{ "date": "2026-10-22", "change": "opt-in count-first dispatch with a streaming FFN expansion, one NVLink domain (docs/design.md section 16); off by default", "decision": "not the default: below 2,048 tokens it does not win, and under compute load at 3A:1F it raises FFN GEMM time by up to 1.5 us even with the epilogue capped at 64 SMs (gate: <= 1 us). Recommended opt-in setting: MIN_BUCKET=2048, EPILOGUE_SMS=64, UNGATE=1 (no payload gate measures better).", "native_source_sha256": { "shipped": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "stage_a_before_review": "9a775052", "stage_b_rejected": "00a7c096", "control_tree_final_role_suites": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" }, "setup": "4 x H200 (kermit), BF16, hidden 2048, top-k 8 of 128 experts (96 at 1A:3F), capacity 2048, CUDA graphs, 24 transport SMs; 3A:1F with the single-FFN collect view; benchmark bucket = token count", "design": "three independent designs, three adversarial reviews and a synthesis; stage A implemented and reviewed by ordering/PTX and host/JIT reviewers (4 minor findings, all fixed)", "correctness_shipped": { "native_source_sha256": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "min_bucket": 1, "summary_and_log_hashes": [ "cf-built exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "count-first cache entries: before 43 after 60", "bf16-small exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "bf16-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "fp8-small exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "fp8-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "f1-small exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "f1-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "f1-fp8-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "f3-small exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "f3-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "f3-fp8-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "view-n4 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "view-n3-large exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "h7168 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "h8192-fp8 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "sms7 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "sms23 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "arm-source exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "arm-source-f1 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "knob-UNGATED_TOKENS exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "knob-UNGATE exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "knob-WARP_COUNT exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "knob-UNIT_MIN_ROWS exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "knob-EPILOGUE_SMS exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_PAYLOAD_FIRST exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_SHARED_FFN_RECEIVE exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_SHARED_ROLE_KERNELS exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_KEEP_ENTRY_BARRIERS exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_UNION_NOTIFY exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_ALL_RANK_EXIT exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_GRID_SYNC_EXIT exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "ctl-AFDEP_UNION_BUFFERS exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "mixed-ffn-f2 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "mixed-ffn-f1 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "mixed-ffn-f3 exit=0 mismatch=0 timeout=1 trap=0 \"native_source_sha256\": \"958652a3", "n11-disagree exit=1 agreement_error=3", "invalid-eager-dummy exit=1 assertion=yes", "invalid-eager-duplicate exit=1 assertion=yes", "invalid-eager-negative exit=1 assertion=yes", "invalid-eager-wide exit=1 assertion=yes", "invalid-graph-dummy exit=1 assertion=yes", "invalid-graph-duplicate exit=1 assertion=yes", "invalid-graph-negative exit=1 assertion=yes", "invalid-graph-wide exit=1 assertion=yes", "done", "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "17efcaa10108c057a74d417cf0426226a6f722a27393bfbe8d160dd5d40f992c arm-source-f1.log", "cf48125a53fcb47b3efe84a5e5c4427a0269cf6f232544e1a494109c79822254 arm-source.log", "ca303bde634e7f1e3dd4457fdbb930a294ea63c5d8ae2562683aef6094f01c0b bf16-large.log", "c38af1e73d9ed2f5be81f94237a2ba67dc9799a38379ee852fe73400a7d3296c bf16-small.log", "7616f43f879e0861c143b7f91259a1079bdffc1d0d8842d1a13acb87f03e627c cf-built.log", "36b1a1b75669039d4348a7df1f9a54aca5b7317ccaa899770b3db7c7994f1c1b ctl-AFDEP_ALL_RANK_EXIT.log", "34f2461ecdb8cf4022f557de6a2d6624430e7334abaa70f02d70b1f3946be92a ctl-AFDEP_GRID_SYNC_EXIT.log", "c6b90b09f461ad5d12c22ae495f73b9856668caf00c9f7cb9a1cd3a07a9b29a4 ctl-AFDEP_KEEP_ENTRY_BARRIERS.log", "d91cc8959dcbd0115dc2b8c12b0ec8ce16566f3c716a2747d8a0feab507861f1 ctl-AFDEP_PAYLOAD_FIRST.log", "777f0599d004fe5c7a92838410c10063d0a2affb53347560ff63cbe3323584ec ctl-AFDEP_SHARED_FFN_RECEIVE.log", "5718cd28b9d20205e60d8b95923fb35ff64eb75395d449ca540ddf448fc4df6f ctl-AFDEP_SHARED_ROLE_KERNELS.log", "8826180b058bbb0f07f2bc0e7e89f30bbc12307df2c2fae26830607667c4b385 ctl-AFDEP_UNION_BUFFERS.log", "0c1aa7bccf4258e6ac7999609d17924ab2e073aa8d5724ba584d0dc783f7b35b ctl-AFDEP_UNION_NOTIFY.log", "c12182141bd807374f5fcd0f229d24b7fe6c1edaa0481dad336c79d311ee37cf f1-fp8-large.log", "5cec59f50c41b431e4e913c7d6bda22e9b1bc87f8c5932ef59e4ef3e647a110b f1-large.log", "66022d9a41b2556126f33a69e0c429ea33d1cf88f64e3ebc51f896acc9c1f21b f1-small.log", "7f2faf0bb340490e210a5f773cd8603daecfae3593e38a5215bd4fa8752703a3 f3-fp8-large.log", "9dd18a396f919386de3038dbaf7365e7df4a6266709bcc5dd7da27cc29fc6652 f3-large.log", "5a36847aed546aa021d1e888341db3d8b3f318462ff5d3af37b9d4f899498a62 f3-small.log", "586a5726c245512c2214d1bd199e320b29642f99601feabfc58362894fa6f98e fp8-large.log", "ecf6d771ee92fd754848db63beea1dcb78d00da83b23d895f6d4d8a9f4e9231a fp8-small.log", "4c0b049f0f88b6a0f532776f872efe97ebf6683e6032d619ee1deb25d3540b93 h7168.log", "6b6637106f9877abade2840676bc3e22d011eb003c410f92341409a7a6e70bff h8192-fp8.log", "24c697fb2d35e9d2db248dfe53a25112fc038c09bbf336e99f243e2dbd3b430d invalid-eager-dummy.log", "a48add85f771b2368dfc2022d94af2f8852dc2eb84a0d370e397d58f8e261f34 invalid-eager-duplicate.log", "0139be12493cbe99142da22672ed13f03ffb04cd92ece0823333a7b45f39fd15 invalid-eager-negative.log", "f37aa94f82564ca1a51f9a3e287c19ef2fc45664f8679f1e4fbdf80c58a3c98f invalid-eager-wide.log", "7c3aa6c238938b60b14b5f7e83d7af11013232855f72c686c941e8958e310c66 invalid-graph-dummy.log", "4ecd3ba5d51a5064c7d2b6ddd6a9bd51be30c453f6de7a9fa591b8712712a26b invalid-graph-duplicate.log", "d32ba8f587b0dbb60a1d4f92b17c52e6860d13cf58488c0fd8551e9af3da848c invalid-graph-negative.log", "85b97d6a2d81fb82bd1150df1140146350d361806e43d5e79cdf26be1d5e08ea invalid-graph-wide.log", "90535c512048bf8d2d54834950032bcd5dc44925deba1be426ed407ff066fd7f knob-EPILOGUE_SMS.log", "fdd1d56615da12cebc0c58a1874716b6e3f4f9f4679cb0a22d31fcc1c2ca9981 knob-UNGATE.log", "51a3fdd4c94ab875516582ad83f0101b9d3f2f6eaa84cbe2b05fb5bba64f2a74 knob-UNGATED_TOKENS.log", "351a5eacd73ea8d1c732449b8ced8034431e1b3ed590646441e5be26e2aabed9 knob-UNIT_MIN_ROWS.log", "e9882ecbb6608004ef7cde0bb8d88dc2a7d1be53c31f43898dfdb7480c56dce9 knob-WARP_COUNT.log", "3abe3934c94c77980b963ab867197f4ea9d0d7edb7e84002b2cdce86a39e6e91 mixed-ffn-f1.log", "766107cabaf09b047ce59eb9b371b58ee5beef0bc2638c31bb95d2ceff2fd333 mixed-ffn-f2.log", "63bcba45056773dbee8fb26569c3519b8089ae9f0b54336c8b8c2a52ab5edd28 mixed-ffn-f3.log", "2d71dc866ca6bf3dd2be4904fa1a45b7f88e67b5064ef4367f20b8fbe262b443 n11-disagree.log", "9a77b13c20254a50f3e3fd6e573cf47c3692ec0f6d871e360af09d6727753440 sms23.log", "9179340e18708afe3920b787188aeba071a51fad02298ec87a70ceeca7ae6d5f sms7.log", "e88bd403aa1db1ca6105d66c6bf81f1f0acf85d8464b13aa76e3b864681da627 view-n3-large.log", "66022d9a41b2556126f33a69e0c429ea33d1cf88f64e3ebc51f896acc9c1f21b view-n4.log" ], "runner": "# cf-validate.sh : count-first correctness on a 4-GPU pod (MIN_BUCKET=1: count-first at every size).\nset -u\ntree=$1; out=$2; mkdir -p $out; : > $out/summary.txt; cd $tree\nexport PYTHONPATH=$tree PYTHONSAFEPATH=1 AFDEP_COUNT_FIRST_MIN_BUCKET=1\npython -c \"from afdep._build import source_digest; print(source_digest())\" | tee $out/digest.txt\nL=\"--capacity 2048 --hidden-size 2048 --top-k 8 --num-sms 24\"\nrun() { name=$1; np=$2; shift 2; timeout 900 torchrun --standalone --nproc-per-node=$np -m tests.gpu.deepep \"$@\" > $out/$name.log 2>&1; echo \"$name exit=$? mismatch=$(grep -c 'Mismatched\\|not equal' $out/$name.log) timeout=$(grep -c 'timeout' $out/$name.log) trap=$(grep -ci 'not count-first\\|P not cleared\\|illegal' $out/$name.log) $(grep -o '\"native_source_sha256\": \"[0-9a-f]\\{8\\}' $out/$name.log | sort -u | tr '\\n' ' ')\" | tee -a $out/summary.txt; }\nbefore=$(ls /work/cache/deepep-jit/cache | grep -c \"afdep_source_dispatch\\|afdep_ffn_stream\")\nrun cf-built 4 --payload bf16 --hybrid off\necho \"count-first cache entries: before $before after $(ls /work/cache/deepep-jit/cache | grep -c 'afdep_source_dispatch\\|afdep_ffn_stream')\" | tee -a $out/summary.txt\nfor p in bf16 fp8; do run $p-small 4 --payload $p --hybrid off; run $p-large 4 --payload $p --hybrid off $L; done\nfor f in 1 3; do run f$f-small 4 --payload bf16 --hybrid off --ffn-ranks $f; run f$f-large 4 --payload bf16 --hybrid off --ffn-ranks $f $L --experts-per-ffn 32; run f$f-fp8-large 4 --payload fp8 --hybrid off --ffn-ranks $f $L --experts-per-ffn 32; done\nrun view-n4 4 --payload bf16 --hybrid off --ffn-ranks 1 --collect-view\nrun view-n3-large 3 --payload fp8 --hybrid off --ffn-ranks 1 --collect-view $L --experts-per-ffn 32\nrun h7168 4 --payload bf16 --hybrid off --capacity 256 --hidden-size 7168 --top-k 8 --num-sms 24\nrun h8192-fp8 4 --payload fp8 --hybrid off --capacity 256 --hidden-size 8192 --top-k 8 --num-sms 24\nrun sms7 4 --payload bf16 --hybrid off --num-sms 7\nrun sms23 4 --payload bf16 --hybrid off $L --num-sms 23\nfor arm in source; do export AFDEP_COUNT_FIRST_ARM=$arm; run arm-$arm 4 --payload bf16 --hybrid off $L; run arm-$arm-f1 4 --payload bf16 --hybrid off --ffn-ranks 1 $L --experts-per-ffn 32; unset AFDEP_COUNT_FIRST_ARM; done\nfor k in \"UNGATED_TOKENS=48\" \"UNGATE=1\" \"WARP_COUNT=1\" \"UNIT_MIN_ROWS=32\" \"EPILOGUE_SMS=48\"; do export AFDEP_COUNT_FIRST_$k; run knob-${k%%=*} 4 --payload bf16 --hybrid off --ffn-ranks 1 $L --experts-per-ffn 32; unset AFDEP_COUNT_FIRST_${k%%=*}; done\nfor c in AFDEP_PAYLOAD_FIRST AFDEP_SHARED_FFN_RECEIVE AFDEP_SHARED_ROLE_KERNELS AFDEP_KEEP_ENTRY_BARRIERS AFDEP_UNION_NOTIFY AFDEP_ALL_RANK_EXIT AFDEP_GRID_SYNC_EXIT AFDEP_UNION_BUFFERS; do\n export $c=1; run ctl-$c 4 --payload bf16 --hybrid off $L; unset $c\ndone\nfor f in 2 1 3; do AFDEP_PAYLOAD_FIRST=1 run mixed-ffn-f$f 4 --payload bf16 --hybrid off --ffn-ranks $f $L --experts-per-ffn 32 --role-env ffn:AFDEP_SHARED_FFN_RECEIVE=1; done\ntimeout 300 torchrun --standalone --nproc-per-node=4 -m tests.gpu.deepep --payload bf16 --hybrid off --role-env ffn:AFDEP_SHARED_FFN_RECEIVE=1 > $out/n11-disagree.log 2>&1; echo \"n11-disagree exit=$? agreement_error=$(grep -c 'same count-first configuration' $out/n11-disagree.log)\" | tee -a $out/summary.txt\nfor e in eager graph; do for r in dummy duplicate negative wide; do\n timeout 300 torchrun --standalone --nproc-per-node=4 -m tests.gpu.deepep --payload bf16 --invalid-route $r --invalid-execution $e > $out/invalid-$e-$r.log 2>&1; c=$?\n grep -q -i assert $out/invalid-$e-$r.log && a=yes || a=no; echo \"invalid-$e-$r exit=$c assertion=$a\" | tee -a $out/summary.txt\ndone; done\necho done >> $out/summary.txt\n" }, "default_off_shipped": { "summary": [ "off-f1-large exit=0 \"native_source_sha256\": \"958652a3", "off-f3-large exit=0 \"native_source_sha256\": \"958652a3", "off-view exit=0 \"native_source_sha256\": \"958652a3", "off-fp8 exit=0 \"native_source_sha256\": \"958652a3" ], "matrix": [ "bf16-direct-small exit=0", "bf16-hybrid-small exit=0", "bf16-generic-control exit=0", "bf16-off-large exit=0", "bf16-on-large exit=0", "fp8-direct-small exit=0", "fp8-hybrid-small exit=0", "fp8-generic-control exit=0", "fp8-off-large exit=0", "fp8-on-large exit=0", "invalid-eager-dummy exit=1 assertion=yes", "invalid-eager-duplicate exit=1 assertion=yes", "invalid-eager-negative exit=1 assertion=yes", "invalid-eager-wide exit=1 assertion=yes", "invalid-graph-dummy exit=1 assertion=yes", "invalid-graph-duplicate exit=1 assertion=yes", "invalid-graph-negative exit=1 assertion=yes", "invalid-graph-wide exit=1 assertion=yes", "done" ] }, "faults_shipped": { "patches": "\"\"\"cf-faults.py : scratch fault injection for count-first (never commit). Faults: P1 P2 P3 P4 (must pass), N2 N3 (must fail).\"\"\"\nimport sys\nfrom pathlib import Path\ntree, fault = Path(sys.argv[1]), sys.argv[2]\nI = tree / \"afdep/third_party/deepep_v2/deep_ep/include/deep_ep/impls\"\nSPIN = lambda ns: ('{ unsigned long long _t0, _t1; asm volatile(\"mov.u64 %0, %%globaltimer;\" : \"=l\"(_t0)); '\n 'do { asm volatile(\"mov.u64 %0, %%globaltimer;\" : \"=l\"(_t1)); } while (_t1 - _t0 < ' + f'{ns}ull); }}')\ndef patch(name, old, new):\n p = I / name; s = p.read_text(); assert s.count(old) == 1, (name, old[:60]); p.write_text(s.replace(old, new))\ndef P1(): # late rows: delay the NVLink store of every 5th token by 20 us (flag follows the store)\n a = \" // Issue TMA NVLink stores\\n\"\n patch(\"afdep_source_dispatch.cuh\", a, f\" if (token_idx % 5 == 2) {{ {SPIN(20000)} if (token_idx % 97 == 2 and lane_idx == 0) printf(\\\"AFDFAULT P1\\\\n\\\"); }}\\n\" + a)\ndef P2(): # deferred flag: 10 us between store completion and flag release\n a = \" if (slot_idx >= 0) {\\n afdep_cf_st_release_sys(\"\n patch(\"afdep_count_first.cuh\", a, f\" if (slot_idx >= 0 and slot_idx % 7 == 3) {{ {SPIN(10000)} if (slot_idx % 97 == 3) printf(\\\"AFDFAULT P2\\\\n\\\"); }}\\n\" + a)\ndef P3(): # late counts: source SM0 thread 0 waits 30 us before its count puts\n a = \" // TODO: for further optimization, we can fuse rank and expert counters\\n\"\n patch(\"afdep_source_dispatch.cuh\", a, f\" if (thread_idx == 0) {{ {SPIN(30000)} printf(\\\"AFDFAULT P3\\\\n\\\"); }}\\n\" + a)\ndef P4(): # late prefix sums: FFN publishes P 20 us late\n a = \" afdep_cf_st_release_gpu(ptr, 1);\\n}\"\n patch(\"afdep_count_first.cuh\", a, f\" {SPIN(20000)} printf(\\\"AFDFAULT P4\\\\n\\\");\\n\" + a)\ndef N2(): # epilogue skips the row wait (with P1): must read rows before they land\n P1()\n a = \" const int lane_idx, const int rank_idx) {\\n if (lane_idx == 0) {\\n const auto flag\"\n patch(\"afdep_count_first.cuh\", a, \" const int lane_idx, const int rank_idx) {\\n if (true) { if (lane_idx == 0 and slot % 97 == 0) printf(\\\"AFDFAULT N2\\\\n\\\"); __syncwarp(); return; }\\n if (lane_idx == 0) {\\n const auto flag\")\ndef N3(): # epilogue skips the prefix wait (with P4): must read stale prefix sums\n P4()\n a = \" const int thread_idx, const int rank_idx) {\\n if (thread_idx == 0) {\\n const auto ptr = afdep_cf_word(workspace, kAfdCfPrefixReady);\\n comm::timeout_while\"\n patch(\"afdep_count_first.cuh\", a, \" const int thread_idx, const int rank_idx) {\\n if (true) { if (thread_idx == 0) printf(\\\"AFDFAULT N3\\\\n\\\"); __syncthreads(); return; }\\n if (thread_idx == 0) {\\n const auto ptr = afdep_cf_word(workspace, kAfdCfPrefixReady);\\n comm::timeout_while\")\ndef P4b(): # late prefix sums: FFN CTA0 thread 0 spins 20 us between the epilogue trigger and the expert reduction\n a = \" cudaTriggerProgrammaticLaunchCompletion();\\n\\n // Reduce expert count and add stats\\n\"\n patch(\"afdep_ffn_stream_dispatch.cuh\", a, \" cudaTriggerProgrammaticLaunchCompletion();\\n\" + f\" if (thread_idx == 0) {{ {SPIN(20000)} printf(\\\"AFDFAULT P4b\\\\n\\\"); }}\\n\" + \"\\n // Reduce expert count and add stats\\n\")\ndef N3b(): # epilogue skips the prefix wait while the prefix sums are late (P4b): must read stale prefix sums\n P4b()\n a = \" const int thread_idx, const int rank_idx) {\\n if (thread_idx == 0) {\\n const auto ptr = afdep_cf_word(workspace, kAfdCfPrefixReady);\\n comm::timeout_while\"\n patch(\"afdep_count_first.cuh\", a, \" const int thread_idx, const int rank_idx) {\\n if (true) { if (thread_idx == 0) printf(\\\"AFDFAULT N3b\\\\n\\\"); __syncthreads(); return; }\\n if (thread_idx == 0) {\\n const auto ptr = afdep_cf_word(workspace, kAfdCfPrefixReady);\\n comm::timeout_while\")\nglobals()[fault]()\nprint(\"applied\", fault)\n", "summary": [ "applied P1", "P1-f1 tree=00064fc5 exit=0 mismatch=0 timeout_msgs=0 markers=11332 assert=0", "P1-f2 tree=00064fc5 exit=0 mismatch=0 timeout_msgs=0 markers=7551 assert=0", "applied P2", "P2-f1 tree=c75ef6e1 exit=0 mismatch=0 timeout_msgs=0 markers=4837 assert=0", "P2-f2 tree=c75ef6e1 exit=0 mismatch=0 timeout_msgs=0 markers=3088 assert=0", "applied P3", "P3-f1 tree=36e598f5 exit=0 mismatch=0 timeout_msgs=0 markers=2304 assert=0", "P3-f2 tree=36e598f5 exit=0 mismatch=0 timeout_msgs=0 markers=1536 assert=0", "applied P4", "P4-f1 tree=4af0b7a9 exit=0 mismatch=0 timeout_msgs=0 markers=768 assert=0", "P4-f2 tree=4af0b7a9 exit=0 mismatch=0 timeout_msgs=0 markers=1536 assert=0", "applied N2", "N2-f1 tree=f1ba5ca8 exit=1 mismatch=8 timeout_msgs=0 markers=162 assert=12", "N2-f2 tree=f1ba5ca8 exit=1 mismatch=8 timeout_msgs=0 markers=124 assert=12", "applied N3", "N3-f1 tree=366a89ce exit=0 mismatch=0 timeout_msgs=0 markers=102144 assert=0", "N3-f2 tree=366a89ce exit=0 mismatch=0 timeout_msgs=0 markers=204288 assert=0" ], "log_hashes": [ { "log": "N2-f1.log", "log_sha256": "31461173dd595216773f46b636ea0529854c9de11945cf73ffa563d9fbde6370" }, { "log": "N2-f2.log", "log_sha256": "d06b39d42ef7894d65c3ccefe8314f3de4d56448067030e5f7b86cfe9a4670bc" }, { "log": "N3-f1.log", "log_sha256": "e412b51b164f3d8d9b06086d8d20237616ec2a3e32674c35de86a45c038236b4" }, { "log": "N3-f2.log", "log_sha256": "c17784fbca39d642f85f5cef4216182dc3d5cbbd6c82799baaa9c8be0db4ca8d" }, { "log": "P1-f1.log", "log_sha256": "575f0ce5f347ec518bf7476e72ec3188c144791b9c1432e2764e8da0391c9bb8" }, { "log": "P1-f2.log", "log_sha256": "3dde61b0b373ef88566fba980aa4bc2ed051a9ccc2a3301d59b3b48933f350dc" }, { "log": "P2-f1.log", "log_sha256": "5e414ac2f635b2d7fe605ea03b7cda548017c4b771cc857590616048997095e7" }, { "log": "P2-f2.log", "log_sha256": "50b22b25235d9852d5045c60f459176f7759dfe7e178fac2f58daa28f257362d" }, { "log": "P3-f1.log", "log_sha256": "975101912dbfdc3ba753dd9baf3b3ddcb8d1ab849ac6e807d9c4d0febfd8f915" }, { "log": "P3-f2.log", "log_sha256": "aeb94c33cf176831a79e361d4587cb986cc941b4a9a01575b52b42378075bdb7" }, { "log": "P4-f1.log", "log_sha256": "4a274a9e9561b9fdb6fde408186cc8bf9f87afef6aa929fc829f67b4014c6f5d" }, { "log": "P4-f2.log", "log_sha256": "d8f96e0983b284f27f185b597dc9ecd13813240453629dc9e774cb9beb4be344" } ], "corrected_prefix_faults": { "summary": [ "applied P4b", "P4b-f1 tree=8a3529a5 exit=0 mismatch=0 timeout_msgs=0 markers=768 assert=0", "P4b-f2 tree=8a3529a5 exit=0 mismatch=0 timeout_msgs=0 markers=1536 assert=0", "applied N3b", "N3b-f1 tree=d4ccdcd5 exit=1 mismatch=0 timeout_msgs=0 markers=266 assert=1921", "N3b-f2 tree=d4ccdcd5 exit=1 mismatch=0 timeout_msgs=0 markers=532 assert=2774" ], "log_hashes": [ { "log": "N3b-f1.log", "log_sha256": "5e0a43491ec18ff8a4f25a388f59b46e6e8536a27e4acb60a325d82db0ff86ea" }, { "log": "N3b-f2.log", "log_sha256": "e6443346b4a93fbc30f08e39fb4d5a3b40df241f70d856332c7fc965a17048d6" }, { "log": "P4b-f1.log", "log_sha256": "d665d698660305e8c76e44ba3404e3e595e7671a8c418c56dd77008bc9863ca8" }, { "log": "P4b-f2.log", "log_sha256": "2fc05a90055a286bc2f39d84462c1b8d1d8d49a48653993b856c10fa5ef7effb" } ], "note": "N3 (with P4, which delays only the P release after the prefix sums exist) passed: it did not create the race. P4b delays the prefix sums themselves; N3b = P4b + no prefix wait fails." }, "expected": "P1-P4 pass (late rows, deferred flags, late counts, late prefix sums); N2 (epilogue skips the row wait, with P1) and N3 (epilogue skips the prefix wait, with P4) must fail" }, "screens": { "mechanism_3a1f_stage_a": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "CF": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1" } }, "CF_ungate": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_UNGATE": "1" } }, "CF_src": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_ARM": "source" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1", "8", "32", "128", "512", "2048", "--ffn-ranks", "1", "--attention-ranks", "3" ], "processes": [ { "order": [ "C", "CF", "CF_ungate", "CF_src", "ISO" ], "median_us": { "C": { "1": 18.482080101966858, "8": 20.016080141067505, "32": 23.28320026397705, "128": 35.30160069465637, "512": 96.11303806304932, "2048": 320.81119537353516 }, "CF": { "1": 20.903279781341553, "8": 23.903679847717285, "32": 28.062880039215088, "128": 36.65295958518982, "512": 96.81399822235107, "2048": 285.41640281677246 }, "CF_ungate": { "1": 20.854320526123047, "8": 23.661999702453613, "32": 28.371039628982544, "128": 40.10967969894409, "512": 95.5295991897583, "2048": 283.7823963165283 }, "CF_src": { "1": 19.397440552711487, "8": 21.974560022354126, "32": 26.359840631484985, "128": 38.62631916999817, "512": 100.21207809448242, "2048": 330.59295654296875 }, "ISO": { "1": 18.468000292778015, "8": 20.250319242477417, "32": 23.14360022544861, "128": 35.28175950050354, "512": 96.1407995223999, "2048": 320.70240020751953 } }, "native_source_sha256": { "C": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF_ungate": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF_src": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "CF", "CF_ungate", "CF_src", "ISO", "C" ], "median_us": { "CF": { "1": 21.03495955467224, "8": 23.694480657577515, "32": 27.926799058914185, "128": 36.70248031616211, "512": 96.68416023254395, "2048": 285.2403163909912 }, "CF_ungate": { "1": 20.82479953765869, "8": 23.601360321044922, "32": 28.34928035736084, "128": 39.934799671173096, "512": 95.7046365737915, "2048": 283.57672691345215 }, "CF_src": { "1": 19.726000428199768, "8": 21.802879571914673, "32": 26.72919988632202, "128": 38.64863991737366, "512": 100.19984245300293, "2048": 330.44206619262695 }, "ISO": { "1": 18.16215991973877, "8": 20.619360208511353, "32": 23.73687982559204, "128": 35.3384006023407, "512": 95.96776008605957, "2048": 321.7595100402832 }, "C": { "1": 17.976239919662476, "8": 20.230560302734375, "32": 24.01319980621338, "128": 35.414639711380005, "512": 96.18759632110596, "2048": 321.7479133605957 } }, "native_source_sha256": { "CF": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF_ungate": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF_src": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3" } }, { "order": [ "CF_ungate", "CF_src", "ISO", "C", "CF" ], "median_us": { "CF_ungate": { "1": 20.741440057754517, "8": 23.59231948852539, "32": 28.343679904937744, "128": 39.99760031700134, "512": 95.73367595672607, "2048": 283.72623443603516 }, "CF_src": { "1": 19.336639642715454, "8": 21.81015968322754, "32": 26.59783959388733, "128": 38.3024001121521, "512": 100.19471645355225, "2048": 330.23326873779297 }, "ISO": { "1": 18.547439575195312, "8": 20.69696068763733, "32": 23.484959602355957, "128": 35.51759958267212, "512": 96.20903968811035, "2048": 321.14110946655273 }, "C": { "1": 18.49943995475769, "8": 20.446479320526123, "32": 24.039440155029297, "128": 35.211039781570435, "512": 96.35871887207031, "2048": 321.0296821594238 }, "CF": { "1": 20.932079553604126, "8": 23.796000480651855, "32": 27.957359552383423, "128": 36.58552050590515, "512": 96.75007820129395, "2048": 285.29343605041504 } }, "native_source_sha256": { "CF_ungate": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF_src": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3", "CF": "9a7750524cdc3103426dcf7e27dd8d8fd55bea9b24fcffe7986b8fdc9fd21aa3" } } ], "table": [ { "tokens": 1, "C": { "median_us": 18.482080101966858, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 20.932079553604126, "paired_change_pct": [ 13.10025530685277, 17.015347195406115, 13.149801317205867 ], "geomean_change_pct": 14.407211949823395, "repeats_faster": 0 }, "CF_ungate": { "median_us": 20.82479953765869, "paired_change_pct": [ 12.835354089303696, 15.84624832961008, 12.119286359370186 ], "geomean_change_pct": 13.588883711181365, "repeats_faster": 0 }, "CF_src": { "median_us": 19.397440552711487, "paired_change_pct": [ 4.952691719192459, 9.73374029472871, 4.525540719098653 ], "geomean_change_pct": 6.378059105173417, "repeats_faster": 0 }, "ISO": { "median_us": 18.468000292778015, "paired_change_pct": [ -0.07618086877214658, 1.034254109353161, 0.2594652624890781 ], "geomean_change_pct": 0.40477086892745096, "repeats_faster": 1 } }, { "tokens": 8, "C": { "median_us": 20.230560302734375, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 23.796000480651855, "paired_change_pct": [ 19.422382800483962, 17.122216602053065, 16.38189689098779 ], "geomean_change_pct": 17.635070090222115, "repeats_faster": 0 }, "CF_ungate": { "median_us": 23.601360321044922, "paired_change_pct": [ 18.214952856357126, 16.66192121161838, 15.38573031906365 ], "geomean_change_pct": 16.74847424458592, "repeats_faster": 0 }, "CF_src": { "median_us": 21.81015968322754, "paired_change_pct": [ 9.784532573230242, 7.772000605281226, 6.669511857390642 ], "geomean_change_pct": 8.067673805211228, "repeats_faster": 0 }, "ISO": { "median_us": 20.619360208511353, "paired_change_pct": [ 1.1702546140855974, 1.9218444766674558, 1.2250586674829034 ], "geomean_change_pct": 1.4384765616859196, "repeats_faster": 0 } }, { "tokens": 32, "C": { "median_us": 24.01319980621338, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 27.957359552383423, "paired_change_pct": [ 20.5284484995518, 16.297699949542622, 16.29788119892823 ], "geomean_change_pct": 17.691247086517127, "repeats_faster": 0 }, "CF_ungate": { "median_us": 28.34928035736084, "paired_change_pct": [ 21.851976134385698, 18.057071053169295, 17.90490844275321 ], "geomean_change_pct": 19.257441744786163, "repeats_faster": 0 }, "CF_src": { "median_us": 26.59783959388733, "paired_change_pct": [ 13.213992632567795, 11.310446346287772, 10.642508404351458 ], "geomean_change_pct": 11.717021399435023, "repeats_faster": 0 }, "ISO": { "median_us": 23.484959602355957, "paired_change_pct": [ -0.5995741004058863, -1.1507003766729973, -2.306545198629917 ], "geomean_change_pct": -1.3548426333275465, "repeats_faster": 3 } }, { "tokens": 128, "C": { "median_us": 35.30160069465637, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 36.65295958518982, "paired_change_pct": [ 3.8280385703246766, 3.6364639462032278, 3.903550513876408 ], "geomean_change_pct": 3.7892901083397046, "repeats_faster": 0 }, "CF_ungate": { "median_us": 39.99760031700134, "paired_change_pct": [ 13.620002803486253, 12.763535070894982, 13.593919876050386 ], "geomean_change_pct": 13.325120128792122, "repeats_faster": 0 }, "CF_src": { "median_us": 38.62631916999817, "paired_change_pct": [ 9.418038870529344, 9.131817328511316, 8.779520144133013 ], "geomean_change_pct": 9.109479550885723, "repeats_faster": 0 }, "ISO": { "median_us": 35.3384006023407, "paired_change_pct": [ -0.05620480024248753, -0.21527568728818558, 0.87063546831736 ], "geomean_change_pct": 0.19857655386603668, "repeats_faster": 2 } }, { "tokens": 512, "C": { "median_us": 96.18759632110596, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 96.75007820129395, "paired_change_pct": [ 0.7293080870484303, 0.5162452649095206, 0.40614833178014287 ], "geomean_change_pct": 0.5504777778371217, "repeats_faster": 0 }, "CF_ungate": { "median_us": 95.7046365737915, "paired_change_pct": [ -0.6070340559917398, -0.502101898567231, -0.6486625420726755 ], "geomean_change_pct": -0.5859519537053326, "repeats_faster": 3 }, "CF_src": { "median_us": 100.19984245300293, "paired_change_pct": [ 4.264811636423538, 4.171271853496328, 3.9809553576306334 ], "geomean_change_pct": 4.138945954607198, "repeats_faster": 0 }, "ISO": { "median_us": 96.1407995223999, "paired_change_pct": [ 0.02888417628872375, -0.2285494631890983, -0.15533538190631857 ], "geomean_change_pct": -0.11839225529759467, "repeats_faster": 2 } }, { "tokens": 2048, "C": { "median_us": 321.0296821594238, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 285.29343605041504, "paired_change_pct": [ -11.03290442079209, -11.346646070922295, -11.131757620861404 ], "geomean_change_pct": -11.170532630933161, "repeats_faster": 3 }, "CF_ungate": { "median_us": 283.72623443603516, "paired_change_pct": [ -11.542240293046046, -11.863693550784149, -11.619937281956293 ], "geomean_change_pct": -11.675396599845023, "repeats_faster": 3 }, "CF_src": { "median_us": 330.44206619262695, "paired_change_pct": [ 3.049071014508775, 2.7021629266286373, 2.8668958323294857 ], "geomean_change_pct": 2.8726123605941734, "repeats_faster": 0 }, "ISO": { "median_us": 321.14110946655273, "paired_change_pct": [ -0.03391252162785463, 0.0036042750258591383, 0.03470934724147945 ], "geomean_change_pct": 0.0014630979667140664, "repeats_faster": 1 } } ] }, "traces_3a1f_stage_a": { "C-t2048": [ "## C T=2048", "anchor device 1 dispatch start = 0; period median 321.57 us", " dev0 0:dispatch_impl start -0.02 end 79.56 dur 79.57", " dev0 1:dispatch_copy_epilogue start 70.55 end 145.20 dur 75.07", " dev0 2:combine_impl start 145.19 end 319.77 dur 174.53", " dev1 0:dispatch_impl start 0.00 end 79.39 dur 79.39", " dev1 1:combine_impl start 79.94 end 319.84 dur 239.87", " dev2 0:dispatch_impl start 0.20 end 79.53 dur 79.33", " dev2 1:combine_impl start 80.08 end 319.90 dur 239.94", " dev3 0:dispatch_impl start 0.06 end 79.46 dur 79.39", " dev3 1:combine_impl start 80.00 end 320.03 dur 240.03" ], "C-t512": [ "## C T=512", "anchor device 1 dispatch start = 0; period median 96.74 us", " dev0 0:dispatch_impl start 0.03 end 23.09 dur 23.04", " dev0 1:dispatch_copy_epilogue start 21.74 end 39.80 dur 18.10", " dev0 2:combine_impl start 39.81 end 94.73 dur 54.82", " dev1 0:dispatch_impl start 0.00 end 22.85 dur 22.85", " dev1 1:combine_impl start 23.39 end 94.96 dur 71.57", " dev2 0:dispatch_impl start -0.07 end 22.84 dur 22.91", " dev2 1:combine_impl start 23.40 end 94.84 dur 71.49", " dev3 0:dispatch_impl start -0.10 end 22.93 dur 23.01", " dev3 1:combine_impl start 23.49 end 94.85 dur 71.30" ], "CF-t2048": [ "## CF T=2048", "anchor device 1 dispatch start = 0; period median 289.38 us", " dev0 0:dispatch_impl start -0.09 end 90.79 dur 90.91", " dev0 1:afdep_ffn_stream_epilogue_impl start 6.28 end 112.74 dur 106.45", " dev0 2:combine_impl start 112.75 end 287.18 dur 174.66", " dev1 0:dispatch_impl start 0.00 end 89.01 dur 89.01", " dev1 1:combine_impl start 89.55 end 287.49 dur 197.97", " dev2 0:dispatch_impl start 0.03 end 89.43 dur 89.44", " dev2 1:combine_impl start 90.00 end 287.42 dur 197.57", " dev3 0:dispatch_impl start -0.12 end 89.70 dur 89.79", " dev3 1:combine_impl start 90.25 end 287.41 dur 197.17" ], "CF-t512": [ "## CF T=512", "anchor device 1 dispatch start = 0; period median 97.09 us", " dev0 0:dispatch_impl start -0.21 end 29.48 dur 29.66", " dev0 1:afdep_ffn_stream_epilogue_impl start 5.67 end 40.25 dur 34.61", " dev0 2:combine_impl start 40.23 end 94.95 dur 54.66", " dev1 0:dispatch_impl start 0.00 end 27.25 dur 27.25", " dev1 1:combine_impl start 27.78 end 95.18 dur 67.42", " dev2 0:dispatch_impl start -0.10 end 27.13 dur 27.23", " dev2 1:combine_impl start 27.70 end 95.05 dur 67.30", " dev3 0:dispatch_impl start -0.23 end 27.29 dur 27.49", " dev3 1:combine_impl start 27.84 end 95.04 dur 67.14" ], "CFungate-t2048": [ "## CFungate T=2048", "anchor device 1 dispatch start = 0; period median 285.68 us", " dev0 0:dispatch_impl start -0.08 end 88.19 dur 88.24", " dev0 1:afdep_ffn_stream_epilogue_impl start 16.73 end 109.10 dur 92.37", " dev0 2:combine_impl start 109.05 end 283.63 dur 174.40", " dev1 0:dispatch_impl start 0.00 end 86.88 dur 86.88", " dev1 1:combine_impl start 87.42 end 283.89 dur 196.35", " dev2 0:dispatch_impl start 0.20 end 86.95 dur 86.72", " dev2 1:combine_impl start 87.50 end 283.98 dur 196.66", " dev3 0:dispatch_impl start 0.17 end 87.28 dur 87.09", " dev3 1:combine_impl start 87.82 end 283.97 dur 196.05" ], "CFungate-t512": [ "## CFungate T=512", "anchor device 1 dispatch start = 0; period median 95.90 us", " dev0 0:dispatch_impl start 0.07 end 28.42 dur 28.32", " dev0 1:afdep_ffn_stream_epilogue_impl start 15.15 end 39.20 dur 23.95", " dev0 2:combine_impl start 39.20 end 93.93 dur 54.75", " dev1 0:dispatch_impl start 0.00 end 25.92 dur 25.92", " dev1 1:combine_impl start 26.46 end 94.14 dur 67.66", " dev2 0:dispatch_impl start -0.06 end 25.83 dur 25.86", " dev2 1:combine_impl start 26.39 end 94.05 dur 67.65", " dev3 0:dispatch_impl start -0.08 end 26.11 dur 26.18", " dev3 1:combine_impl start 26.67 end 94.01 dur 67.33" ], "CFwarp-t2048": [ "## CFwarp T=2048", "anchor device 1 dispatch start = 0; period median 295.12 us", " dev0 0:dispatch_impl start -0.06 end 97.24 dur 97.34", " dev0 1:afdep_ffn_stream_epilogue_impl start 13.83 end 118.66 dur 104.64", " dev0 2:combine_impl start 118.66 end 293.20 dur 174.46", " dev1 0:dispatch_impl start 0.00 end 96.10 dur 96.10", " dev1 1:combine_impl start 96.66 end 293.38 dur 196.94", " dev2 0:dispatch_impl start -0.01 end 96.15 dur 96.21", " dev2 1:combine_impl start 96.70 end 293.43 dur 196.66", " dev3 0:dispatch_impl start 0.00 end 96.41 dur 96.42", " dev3 1:combine_impl start 96.96 end 293.38 dur 196.35" ], "CFwarp-t512": [ "## CFwarp T=512", "anchor device 1 dispatch start = 0; period median 98.51 us", " dev0 0:dispatch_impl start -0.02 end 30.95 dur 30.96", " dev0 1:afdep_ffn_stream_epilogue_impl start 7.15 end 41.76 dur 34.61", " dev0 2:combine_impl start 41.77 end 96.60 dur 54.78", " dev1 0:dispatch_impl start 0.00 end 28.51 dur 28.51", " dev1 1:combine_impl start 29.06 end 96.83 dur 67.65", " dev2 0:dispatch_impl start 0.08 end 28.70 dur 28.61", " dev2 1:combine_impl start 29.26 end 96.70 dur 67.33", " dev3 0:dispatch_impl start 0.11 end 28.78 dur 28.67", " dev3 1:combine_impl start 29.36 end 96.71 dur 67.33" ] }, "ratios_2a2f_1a3f_shipped": { "2a2f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "CF": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "512", "1024", "2048" ], "processes": [ { "order": [ "C", "CF", "ISO" ], "median_us": { "C": { "512": 67.04544067382812, "1024": 105.66088199615479, "2048": 184.22656059265137 }, "CF": { "512": 67.75408029556274, "1024": 102.56103992462158, "2048": 170.72160720825195 }, "ISO": { "512": 67.28208065032959, "1024": 105.87056159973145, "2048": 184.84983444213867 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "CF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "CF", "ISO", "C" ], "median_us": { "CF": { "512": 67.72144079208374, "1024": 102.51839637756348, "2048": 170.97887992858887 }, "ISO": { "512": 67.05568075180054, "1024": 105.77360153198242, "2048": 184.38615798950195 }, "C": { "512": 67.08047866821289, "1024": 105.57576179504395, "2048": 184.50496673583984 } }, "native_source_sha256": { "CF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "ISO", "C", "CF" ], "median_us": { "ISO": { "512": 67.4725604057312, "1024": 106.09367847442627, "2048": 184.7663974761963 }, "C": { "512": 67.30272054672241, "1024": 105.83391666412354, "2048": 184.39743995666504 }, "CF": { "512": 67.90215969085693, "1024": 102.37112045288086, "2048": 171.12104415893555 } }, "native_source_sha256": { "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "CF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 512, "C": { "median_us": 67.08047866821289, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 67.75408029556274, "paired_change_pct": [ 1.0569542307613533, 0.9555121498776353, 0.8906610895742073 ], "geomean_change_pct": 0.9676859674809579, "repeats_faster": 0 }, "ISO": { "median_us": 67.28208065032959, "paired_change_pct": [ 0.3529546142484241, -0.03696741124196157, 0.25235214509773396 ], "geomean_change_pct": 0.18931003799054213, "repeats_faster": 1 } }, { "tokens": 1024, "C": { "median_us": 105.66088199615479, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 102.51839637756348, "paired_change_pct": [ -2.9337650916504865, -2.895897093705835, -3.271915393845126 ], "geomean_change_pct": -3.034006651858512, "repeats_faster": 3 }, "ISO": { "median_us": 105.87056159973145, "paired_change_pct": [ 0.1984458198865724, 0.18739124736086854, 0.2454428773784434 ], "geomean_change_pct": 0.21042348800281996, "repeats_faster": 0 } }, { "tokens": 2048, "C": { "median_us": 184.39743995666504, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 170.97887992858887, "paired_change_pct": [ -7.330622327722114, -7.331015010894859, -7.1998807580243795 ], "geomean_change_pct": -7.2871932395781265, "repeats_faster": 3 }, "ISO": { "median_us": 184.7663974761963, "paired_change_pct": [ 0.33831921275750965, -0.06439325100011661, 0.20008820058345922 ], "geomean_change_pct": 0.15786530937558219, "repeats_faster": 1 } } ] }, "1a3f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "CF": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "512", "1024", "2048", "--ffn-ranks", "3", "--attention-ranks", "1", "--experts", "96" ], "processes": [ { "order": [ "C", "CF", "ISO" ], "median_us": { "C": { "512": 57.65607833862305, "1024": 103.37320327758789, "2048": 189.30071830749512 }, "CF": { "512": 59.95176076889038, "1024": 103.27983856201172, "2048": 186.37264251708984 }, "ISO": { "512": 58.02759885787964, "1024": 103.23087692260742, "2048": 189.58319664001465 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "CF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "CF", "ISO", "C" ], "median_us": { "CF": { "512": 59.720799922943115, "1024": 103.23415756225586, "2048": 186.76112174987793 }, "ISO": { "512": 58.031439781188965, "1024": 103.32119941711426, "2048": 189.8145580291748 }, "C": { "512": 58.13248157501221, "1024": 103.42672348022461, "2048": 189.68799591064453 } }, "native_source_sha256": { "CF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "ISO", "C", "CF" ], "median_us": { "ISO": { "512": 58.132078647613525, "1024": 102.98983573913574, "2048": 189.9343204498291 }, "C": { "512": 57.72200107574463, "1024": 103.1003999710083, "2048": 189.39496040344238 }, "CF": { "512": 59.90544080734253, "1024": 103.34511756896973, "2048": 186.50248527526855 } }, "native_source_sha256": { "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "CF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 512, "C": { "median_us": 57.72200107574463, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 59.90544080734253, "paired_change_pct": [ 3.981683278533854, 2.732239025236516, 3.7826819772459386 ], "geomean_change_pct": 3.497413289288409, "repeats_faster": 0 }, "ISO": { "median_us": 58.031439781188965, "paired_change_pct": [ 0.644373550824251, -0.1738129718286019, 0.7104354738685759 ], "geomean_change_pct": 0.3928583091991511, "repeats_faster": 1 } }, { "tokens": 1024, "C": { "median_us": 103.37320327758789, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 103.27983856201172, "paired_change_pct": [ -0.09031810238622207, -0.18618584393768067, 0.2373585340408324 ], "geomean_change_pct": -0.013212798815132043, "repeats_faster": 2 }, "ISO": { "median_us": 103.23087692260742, "paired_change_pct": [ -0.1376820592453587, -0.10202785079093113, -0.1072393821009876 ], "geomean_change_pct": -0.11565100174987242, "repeats_faster": 3 } }, { "tokens": 2048, "C": { "median_us": 189.39496040344238, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 186.50248527526855, "paired_change_pct": [ -1.5467853564342993, -1.5429938761888473, -1.5272186345467564 ], "geomean_change_pct": -1.5389996535934425, "repeats_faster": 3 }, "ISO": { "median_us": 189.8145580291748, "paired_change_pct": [ 0.1492220077372819, 0.06672120601132203, 0.2847805692600236 ], "geomean_change_pct": 0.16686759503763415, "repeats_faster": 0 } } ] } }, "overlap_shipped": { "3a1f": { "native_source_sha256": [ "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" ], "per_repeat": { "C": { "1024": { "attention.gemm_us_per_gemm": [ 27.06, 27.09, 27.15 ], "attention.exchange_us": [ 243.23, 243.93, 243.12 ], "attention.solo_exchange_us": [ 175.67, 175.33, 175.65 ], "attention.makespan_us": [ 48645.79, 48786.59, 48624.42 ], "ffn.gemm_us_per_gemm": [ 35.04, 35.32, 35.2 ], "ffn.exchange_us": [ 243.19, 243.36, 243.12 ], "ffn.solo_exchange_us": [ 175.4, 174.73, 175.28 ], "ffn.makespan_us": [ 48637.31, 48671.55, 48623.81 ] }, "2048": { "attention.gemm_us_per_gemm": [ 26.79, 26.78, 26.74 ], "attention.exchange_us": [ 480.76, 479.74, 480.18 ], "attention.solo_exchange_us": [ 325.87, 326.27, 326.48 ], "attention.makespan_us": [ 96151.2, 95947.94, 96036.99 ], "ffn.gemm_us_per_gemm": [ 34.4, 34.55, 34.47 ], "ffn.exchange_us": [ 480.63, 479.52, 480.09 ], "ffn.solo_exchange_us": [ 325.71, 326.19, 326.47 ], "ffn.makespan_us": [ 96126.95, 95904.45, 96018.37 ] } }, "CF": { "1024": { "attention.gemm_us_per_gemm": [ 27.31, 27.49, 27.05 ], "attention.exchange_us": [ 234.43, 232.02, 240.8 ], "attention.solo_exchange_us": [ 166.9, 171.68, 166.88 ], "attention.makespan_us": [ 46886.46, 46403.62, 48159.71 ], "ffn.gemm_us_per_gemm": [ 37.64, 38.3, 36.9 ], "ffn.exchange_us": [ 234.31, 232.03, 238.68 ], "ffn.solo_exchange_us": [ 166.36, 171.5, 166.75 ], "ffn.makespan_us": [ 46862.66, 46405.09, 47735.97 ] }, "2048": { "attention.gemm_us_per_gemm": [ 27.42, 27.66, 27.14 ], "attention.exchange_us": [ 446.51, 450.21, 447.0 ], "attention.solo_exchange_us": [ 290.28, 295.78, 290.6 ], "attention.makespan_us": [ 89302.91, 90041.69, 89399.97 ], "ffn.gemm_us_per_gemm": [ 38.17, 38.38, 37.96 ], "ffn.exchange_us": [ 446.4, 449.79, 446.98 ], "ffn.solo_exchange_us": [ 290.16, 295.8, 290.63 ], "ffn.makespan_us": [ 89280.0, 89957.7, 89396.13 ] } } } }, "2a2f": { "native_source_sha256": [ "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" ], "per_repeat": { "C": { "1024": { "attention.gemm_us_per_gemm": [ 28.31, 28.23, 28.02 ], "attention.exchange_us": [ 186.85, 187.73, 187.67 ], "attention.solo_exchange_us": [ 107.07, 107.24, 107.15 ], "attention.makespan_us": [ 37370.82, 37546.4, 37534.4 ], "ffn.gemm_us_per_gemm": [ 31.47, 31.41, 31.24 ], "ffn.exchange_us": [ 186.87, 187.71, 187.75 ], "ffn.solo_exchange_us": [ 106.87, 107.23, 106.93 ], "ffn.makespan_us": [ 37373.18, 37541.6, 37550.82 ] }, "2048": { "attention.gemm_us_per_gemm": [ 27.32, 27.54, 27.55 ], "attention.exchange_us": [ 352.25, 352.11, 353.25 ], "attention.solo_exchange_us": [ 185.87, 185.71, 185.79 ], "attention.makespan_us": [ 70450.59, 70422.21, 70649.89 ], "ffn.gemm_us_per_gemm": [ 32.58, 32.57, 32.61 ], "ffn.exchange_us": [ 352.19, 352.03, 353.16 ], "ffn.solo_exchange_us": [ 185.83, 185.67, 185.86 ], "ffn.makespan_us": [ 70437.63, 70406.24, 70631.97 ] } }, "CF": { "1024": { "attention.gemm_us_per_gemm": [ 28.3, 28.63, 28.42 ], "attention.exchange_us": [ 177.35, 178.8, 179.28 ], "attention.solo_exchange_us": [ 103.96, 104.38, 103.95 ], "attention.makespan_us": [ 35470.72, 35760.96, 35855.93 ], "ffn.gemm_us_per_gemm": [ 33.89, 34.26, 33.75 ], "ffn.exchange_us": [ 177.33, 178.74, 179.27 ], "ffn.solo_exchange_us": [ 103.8, 104.07, 103.64 ], "ffn.makespan_us": [ 35466.88, 35747.68, 35854.88 ] }, "2048": { "attention.gemm_us_per_gemm": [ 28.27, 28.55, 28.61 ], "attention.exchange_us": [ 334.17, 336.24, 337.06 ], "attention.solo_exchange_us": [ 171.2, 171.45, 171.73 ], "attention.makespan_us": [ 66834.21, 67248.35, 67411.26 ], "ffn.gemm_us_per_gemm": [ 34.23, 34.4, 34.48 ], "ffn.exchange_us": [ 334.14, 336.21, 336.99 ], "ffn.solo_exchange_us": [ 171.17, 171.49, 171.72 ], "ffn.makespan_us": [ 68461.82, 68796.77, 68962.4 ] } } } } }, "epilogue_grid_under_load_shipped": { "3a1f": { "native_source_sha256": [ "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" ], "per_repeat": { "C": { "2048": { "attention.gemm_us_per_gemm": [ 26.98, 26.71, 26.84 ], "attention.exchange_us": [ 480.87, 479.46, 479.16 ], "attention.solo_exchange_us": [ 327.55, 326.75, 327.02 ], "attention.makespan_us": [ 96174.85, 95892.99, 95831.68 ], "ffn.gemm_us_per_gemm": [ 34.46, 34.46, 34.31 ], "ffn.exchange_us": [ 480.89, 479.32, 479.12 ], "ffn.solo_exchange_us": [ 327.5, 326.5, 326.9 ], "ffn.makespan_us": [ 96178.18, 95863.81, 95824.26 ] } }, "CF": { "2048": { "attention.gemm_us_per_gemm": [ 27.58, 27.5, 27.65 ], "attention.exchange_us": [ 446.36, 449.35, 450.54 ], "attention.solo_exchange_us": [ 291.53, 294.58, 294.59 ], "attention.makespan_us": [ 89272.58, 89870.98, 90108.41 ], "ffn.gemm_us_per_gemm": [ 37.98, 38.13, 37.86 ], "ffn.exchange_us": [ 446.16, 449.22, 450.06 ], "ffn.solo_exchange_us": [ 290.99, 294.29, 294.16 ], "ffn.makespan_us": [ 89232.25, 89844.32, 90012.16 ] } }, "CF_e32": { "2048": { "attention.gemm_us_per_gemm": [ 26.97, 27.12, 27.04 ], "attention.exchange_us": [ 502.34, 501.72, 501.18 ], "attention.solo_exchange_us": [ 336.84, 336.77, 336.47 ], "attention.makespan_us": [ 100468.32, 100344.76, 100235.36 ], "ffn.gemm_us_per_gemm": [ 34.28, 33.79, 34.27 ], "ffn.exchange_us": [ 502.48, 501.7, 500.86 ], "ffn.solo_exchange_us": [ 336.75, 336.58, 336.05 ], "ffn.makespan_us": [ 100495.58, 100339.42, 100172.13 ] } }, "CF_e64": { "2048": { "attention.gemm_us_per_gemm": [ 27.57, 27.38, 27.39 ], "attention.exchange_us": [ 457.39, 457.83, 457.22 ], "attention.solo_exchange_us": [ 293.67, 293.54, 293.63 ], "attention.makespan_us": [ 91477.5, 91565.73, 91443.58 ], "ffn.gemm_us_per_gemm": [ 35.64, 34.78, 35.96 ], "ffn.exchange_us": [ 457.36, 457.75, 457.02 ], "ffn.solo_exchange_us": [ 293.62, 292.47, 293.48 ], "ffn.makespan_us": [ 91471.55, 91550.37, 91404.38 ] } }, "CF_u32": { "2048": { "attention.gemm_us_per_gemm": [ 27.64, 27.5, 27.43 ], "attention.exchange_us": [ 447.52, 450.16, 448.16 ], "attention.solo_exchange_us": [ 291.61, 296.96, 291.79 ], "attention.makespan_us": [ 89504.67, 90032.03, 89632.06 ], "ffn.gemm_us_per_gemm": [ 37.88, 38.31, 38.45 ], "ffn.exchange_us": [ 446.08, 450.06, 448.16 ], "ffn.solo_exchange_us": [ 291.08, 296.34, 291.83 ], "ffn.makespan_us": [ 89216.06, 90012.61, 89631.52 ] } } } }, "2a2f": { "native_source_sha256": [ "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" ], "per_repeat": { "C": { "2048": { "attention.gemm_us_per_gemm": [ 27.22, 27.28, 27.32 ], "attention.exchange_us": [ 351.98, 352.49, 351.81 ], "attention.solo_exchange_us": [ 185.86, 186.3, 186.44 ], "attention.makespan_us": [ 70395.26, 70498.15, 70362.85 ], "ffn.gemm_us_per_gemm": [ 32.79, 32.7, 32.71 ], "ffn.exchange_us": [ 351.88, 352.45, 351.71 ], "ffn.solo_exchange_us": [ 185.49, 185.97, 186.1 ], "ffn.makespan_us": [ 70375.3, 70490.11, 70341.19 ] } }, "CF": { "2048": { "attention.gemm_us_per_gemm": [ 28.42, 28.44, 28.32 ], "attention.exchange_us": [ 335.27, 335.36, 334.94 ], "attention.solo_exchange_us": [ 171.72, 172.26, 172.08 ], "attention.makespan_us": [ 67053.38, 67071.65, 66988.29 ], "ffn.gemm_us_per_gemm": [ 34.48, 34.3, 34.43 ], "ffn.exchange_us": [ 335.32, 335.39, 334.83 ], "ffn.solo_exchange_us": [ 171.46, 171.76, 171.7 ], "ffn.makespan_us": [ 68967.13, 68593.03, 68851.07 ] } }, "CF_e32": { "2048": { "attention.gemm_us_per_gemm": [ 28.08, 28.28, 28.01 ], "attention.exchange_us": [ 348.2, 346.61, 346.41 ], "attention.solo_exchange_us": [ 174.21, 173.89, 173.77 ], "attention.makespan_us": [ 69639.33, 69321.76, 69281.73 ], "ffn.gemm_us_per_gemm": [ 32.92, 32.49, 32.02 ], "ffn.exchange_us": [ 348.06, 346.61, 346.36 ], "ffn.solo_exchange_us": [ 173.97, 173.62, 173.34 ], "ffn.makespan_us": [ 69611.9, 69322.62, 69271.04 ] } }, "CF_e64": { "2048": { "attention.gemm_us_per_gemm": [ 28.82, 28.34, 28.21 ], "attention.exchange_us": [ 342.15, 339.96, 336.91 ], "attention.solo_exchange_us": [ 171.83, 172.89, 172.24 ], "attention.makespan_us": [ 68429.37, 67992.93, 67382.85 ], "ffn.gemm_us_per_gemm": [ 32.53, 32.32, 31.99 ], "ffn.exchange_us": [ 342.0, 339.86, 336.93 ], "ffn.solo_exchange_us": [ 171.42, 172.46, 172.1 ], "ffn.makespan_us": [ 68400.99, 67972.86, 67386.88 ] } }, "CF_u32": { "2048": { "attention.gemm_us_per_gemm": [ 28.32, 28.57, 28.21 ], "attention.exchange_us": [ 335.75, 336.49, 333.49 ], "attention.solo_exchange_us": [ 172.52, 172.67, 172.89 ], "attention.makespan_us": [ 67149.83, 67297.06, 66697.79 ], "ffn.gemm_us_per_gemm": [ 34.59, 34.34, 34.34 ], "ffn.exchange_us": [ 335.79, 336.58, 333.46 ], "ffn.solo_exchange_us": [ 172.23, 172.65, 172.82 ], "ffn.makespan_us": [ 69171.68, 68670.91, 68682.98 ] } } } } }, "recommended_setting_shipped": { "3a1f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "OFF": {}, "REC": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "2048", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1", "128", "512", "1024", "2048", "--ffn-ranks", "1", "--attention-ranks", "3" ], "processes": [ { "order": [ "C", "OFF", "REC", "ISO" ], "median_us": { "C": { "1": 17.971680164337158, "128": 35.54383993148804, "512": 96.49800300598145, "1024": 171.33743286132812, "2048": 321.32856369018555 }, "OFF": { "1": 18.49503993988037, "128": 35.42879939079285, "512": 96.37263774871826, "1024": 171.3965606689453, "2048": 320.765438079834 }, "REC": { "1": 18.56711983680725, "128": 35.64751982688904, "512": 96.49432182312012, "1024": 171.64008140563965, "2048": 287.68280029296875 }, "ISO": { "1": 18.475199937820435, "128": 35.50175905227661, "512": 96.53703689575195, "1024": 171.41192436218262, "2048": 320.7727241516113 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "OFF", "REC", "ISO", "C" ], "median_us": { "OFF": { "1": 18.00216019153595, "128": 35.60416102409363, "512": 96.31056308746338, "1024": 171.51408195495605, "2048": 321.45408630371094 }, "REC": { "1": 18.473039865493774, "128": 35.31615972518921, "512": 96.2616777420044, "1024": 171.2841510772705, "2048": 287.7348804473877 }, "ISO": { "1": 18.502879738807678, "128": 35.350879430770874, "512": 96.34160041809082, "1024": 171.28376007080078, "2048": 320.59398651123047 }, "C": { "1": 18.465920090675354, "128": 35.35455942153931, "512": 96.27511978149414, "1024": 171.4315128326416, "2048": 320.7480812072754 } }, "native_source_sha256": { "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC", "ISO", "C", "OFF" ], "median_us": { "REC": { "1": 18.41815948486328, "128": 35.55343985557556, "512": 96.46280288696289, "1024": 171.4955234527588, "2048": 287.5588035583496 }, "ISO": { "1": 18.055440187454224, "128": 35.5133593082428, "512": 96.50919914245605, "1024": 171.4855194091797, "2048": 321.91944122314453 }, "C": { "1": 18.033599853515625, "128": 35.74695944786072, "512": 96.65480136871338, "1024": 171.69872283935547, "2048": 321.9063949584961 }, "OFF": { "1": 18.03712010383606, "128": 35.63920021057129, "512": 96.30824089050293, "1024": 171.53528213500977, "2048": 321.4032745361328 } }, "native_source_sha256": { "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 1, "C": { "median_us": 18.033599853515625, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 18.03712010383606, "paired_change_pct": [ 2.9121360426932252, -2.5114367270201, 0.01952050810170025 ], "geomean_change_pct": 0.11558212678099977, "repeats_faster": 1 }, "REC": { "median_us": 18.473039865493774, "paired_change_pct": [ 3.3132109353452455, 0.038556296049474525, 2.1324618183356625 ], "geomean_change_pct": 1.8190463662119072, "repeats_faster": 0 }, "ISO": { "median_us": 18.475199937820435, "paired_change_pct": [ 2.801740120450491, 0.20015059066018637, 0.12110911917755374 ], "geomean_change_pct": 1.033368017528291, "repeats_faster": 0 } }, { "tokens": 128, "C": { "median_us": 35.54383993148804, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 35.60416102409363, "paired_change_pct": [ -0.32365816669480685, 0.7059955112953764, -0.30145007842303784 ], "geomean_change_pct": 0.025812198728858426, "repeats_faster": 2 }, "REC": { "median_us": 35.55343985557556, "paired_change_pct": [ 0.2916958201501174, -0.10861313781979076, -0.5413595877082011 ], "geomean_change_pct": -0.12000500302694617, "repeats_faster": 2 }, "ISO": { "median_us": 35.50175905227661, "paired_change_pct": [ -0.11839148300391011, -0.01040881523810544, -0.6534825429240776 ], "geomean_change_pct": -0.26115774363628663, "repeats_faster": 3 } }, { "tokens": 512, "C": { "median_us": 96.49800300598145, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 96.31056308746338, "paired_change_pct": [ -0.1299148721817689, 0.03681460594355812, -0.3585548501500835 ], "geomean_change_pct": -0.15068325941168803, "repeats_faster": 2 }, "REC": { "median_us": 96.46280288696289, "paired_change_pct": [ -0.0038147762095097626, -0.013962111415966039, -0.19864350144186416 ], "geomean_change_pct": -0.0721802690003881, "repeats_faster": 3 }, "ISO": { "median_us": 96.50919914245605, "paired_change_pct": [ 0.04045046379672623, 0.06905276954998207, -0.15064148308772118 ], "geomean_change_pct": -0.013760332363166317, "repeats_faster": 1 } }, { "tokens": 1024, "C": { "median_us": 171.4315128326416, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 171.51408195495605, "paired_change_pct": [ 0.034509567833351795, 0.048164494934521684, -0.0951904019103389 ], "geomean_change_pct": -0.0041929862682033026, "repeats_faster": 1 }, "REC": { "median_us": 171.4955234527588, "paired_change_pct": [ 0.17663889277275135, -0.08595954905615821, -0.11834647528904396 ], "geomean_change_pct": -0.0093095685472111, "repeats_faster": 2 }, "ISO": { "median_us": 171.41192436218262, "paired_change_pct": [ 0.0434764894106765, -0.08618763225001214, -0.1241729854771556 ], "geomean_change_pct": -0.055653806863820154, "repeats_faster": 2 } }, { "tokens": 2048, "C": { "median_us": 321.32856369018555, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 321.4032745361328, "paired_change_pct": [ -0.17524916051177009, 0.22011202491942772, -0.15629401286921807 ], "geomean_change_pct": -0.037309387510475744, "repeats_faster": 2 }, "REC": { "median_us": 287.68280029296875, "paired_change_pct": [ -10.470828677918886, -10.292563757709194, -10.670055624267727 ], "geomean_change_pct": -10.477948815642247, "repeats_faster": 3 }, "ISO": { "median_us": 320.7727241516113, "paired_change_pct": [ -0.17298167713161172, -0.048042281489235616, 0.0040528131322492555 ], "geomean_change_pct": -0.07235133301737307, "repeats_faster": 2 } } ] }, "2a2f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "OFF": {}, "REC": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "2048", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1", "128", "512", "1024", "2048" ], "processes": [ { "order": [ "C", "OFF", "REC", "ISO" ], "median_us": { "C": { "1": 19.792479872703552, "128": 32.364879846572876, "512": 66.83056116104126, "1024": 105.8798360824585, "2048": 184.443359375 }, "OFF": { "1": 20.149919986724854, "128": 32.42552042007446, "512": 66.9373607635498, "1024": 105.72991847991943, "2048": 184.2695140838623 }, "REC": { "1": 19.75152015686035, "128": 32.403600215911865, "512": 66.95800065994263, "1024": 105.863037109375, "2048": 171.0867977142334 }, "ISO": { "1": 19.847440123558044, "128": 31.875439882278442, "512": 66.77592039108276, "1024": 105.97040176391602, "2048": 184.25447463989258 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "OFF", "REC", "ISO", "C" ], "median_us": { "OFF": { "1": 20.196800231933594, "128": 31.814240217208862, "512": 67.01864004135132, "1024": 106.01280212402344, "2048": 185.08112907409668 }, "REC": { "1": 20.238879919052124, "128": 31.82055950164795, "512": 67.02864170074463, "1024": 106.06544017791748, "2048": 170.74447631835938 }, "ISO": { "1": 20.228079557418823, "128": 31.86168074607849, "512": 66.90847873687744, "1024": 105.97559928894043, "2048": 184.6776008605957 }, "C": { "1": 20.208399295806885, "128": 32.356159687042236, "512": 66.91920042037964, "1024": 105.80327987670898, "2048": 184.25511360168457 } }, "native_source_sha256": { "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC", "ISO", "C", "OFF" ], "median_us": { "REC": { "1": 19.83079969882965, "128": 31.749680042266846, "512": 66.92784070968628, "1024": 105.65063953399658, "2048": 171.1568832397461 }, "ISO": { "1": 20.217039585113525, "128": 32.45551943778992, "512": 66.90416097640991, "1024": 105.63536167144775, "2048": 184.1605567932129 }, "C": { "1": 20.16271948814392, "128": 32.49679923057556, "512": 66.79744005203247, "1024": 105.68448066711426, "2048": 184.3407154083252 }, "OFF": { "1": 20.185920000076294, "128": 31.90384030342102, "512": 66.94936037063599, "1024": 106.0529613494873, "2048": 185.16839981079102 } }, "native_source_sha256": { "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 1, "C": { "median_us": 20.16271948814392, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 20.185920000076294, "paired_change_pct": [ 1.805939004713908, -0.05739724212445729, 0.11506638251856938 ], "geomean_change_pct": 0.6177040200823658, "repeats_faster": 1 }, "REC": { "median_us": 19.83079969882965, "paired_change_pct": [ -0.2069458506798294, 0.15083145774719142, -1.646205461071093 ], "geomean_change_pct": -0.5704826817021447, "repeats_faster": 2 }, "ISO": { "median_us": 20.217039585113525, "paired_change_pct": [ 0.27768248955144514, 0.0973865437032595, 0.2694085835075244 ], "geomean_change_pct": 0.21479139586682017, "repeats_faster": 0 } }, { "tokens": 128, "C": { "median_us": 32.364879846572876, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 31.90384030342102, "paired_change_pct": [ 0.18736535957821765, -1.6748571989845806, -1.8246687095159753 ], "geomean_change_pct": -1.1082704567911938, "repeats_faster": 2 }, "REC": { "median_us": 31.82055950164795, "paired_change_pct": [ 0.11963699393462424, -1.655326808171187, -2.2990546946105583 ], "geomean_change_pct": -1.2835283020215305, "repeats_faster": 2 }, "ISO": { "median_us": 31.875439882278442, "paired_change_pct": [ -1.5122563921591725, -1.5282374229404372, -0.1270272573392517 ], "geomean_change_pct": -1.0580135508240551, "repeats_faster": 3 } }, { "tokens": 512, "C": { "median_us": 66.83056116104126, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 66.94936037063599, "paired_change_pct": [ 0.1598065325999487, 0.14859654680123846, 0.22743434252148198 ], "geomean_change_pct": 0.1786064220434147, "repeats_faster": 0 }, "REC": { "median_us": 66.95800065994263, "paired_change_pct": [ 0.19069045162478648, 0.163542420826146, 0.19521804660811437 ], "geomean_change_pct": 0.18314932982661514, "repeats_faster": 0 }, "ISO": { "median_us": 66.90416097640991, "paired_change_pct": [ -0.08176015435038542, -0.01602183444339156, 0.15976798556098792 ], "geomean_change_pct": 0.020610049586333723, "repeats_faster": 2 } }, { "tokens": 1024, "C": { "median_us": 105.80327987670898, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 106.01280212402344, "paired_change_pct": [ -0.14159221253640109, 0.1980300115068223, 0.34866110903613023 ], "geomean_change_pct": 0.13482291677393565, "repeats_faster": 1 }, "REC": { "median_us": 105.863037109375, "paired_change_pct": [ -0.015866073942927272, 0.24778088308226032, -0.03202091064274937 ], "geomean_change_pct": 0.06654914786912158, "repeats_faster": 2 }, "ISO": { "median_us": 105.97040176391602, "paired_change_pct": [ 0.08553628793586121, 0.16286774137082016, -0.04647701853333075 ], "geomean_change_pct": 0.06727167049533911, "repeats_faster": 1 } }, { "tokens": 2048, "C": { "median_us": 184.3407154083252, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 185.08112907409668, "paired_change_pct": [ -0.09425402558638218, 0.4482998904430824, 0.4489970653702047 ], "geomean_change_pct": 0.26735396420654034, "repeats_faster": 1 }, "REC": { "median_us": 171.0867977142334, "paired_change_pct": [ -7.241551935524438, -7.332571139671029, -7.151882935561016 ], "geomean_change_pct": -7.242031335291143, "repeats_faster": 3 }, "ISO": { "median_us": 184.25447463989258, "paired_change_pct": [ -0.10240798896066217, 0.22929472656290084, -0.0977313203506025 ], "geomean_change_pct": 0.009598019596812613, "repeats_faster": 2 } } ] }, "1a3f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "OFF": {}, "REC": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "2048", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1", "512", "2048", "--ffn-ranks", "3", "--attention-ranks", "1", "--experts", "96" ], "processes": [ { "order": [ "C", "OFF", "REC", "ISO" ], "median_us": { "C": { "1": 19.066799879074097, "512": 57.934160232543945, "2048": 189.53527450561523 }, "OFF": { "1": 19.07200038433075, "512": 58.175599575042725, "2048": 189.62632179260254 }, "REC": { "1": 19.08680021762848, "512": 57.50216007232666, "2048": 187.82184600830078 }, "ISO": { "1": 19.044399857521057, "512": 57.35088109970093, "2048": 189.61527824401855 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "OFF", "REC", "ISO", "C" ], "median_us": { "OFF": { "1": 18.856160044670105, "512": 58.34455966949463, "2048": 189.90328788757324 }, "REC": { "1": 19.093680381774902, "512": 57.85936117172241, "2048": 187.5812816619873 }, "ISO": { "1": 18.852400183677673, "512": 58.00119876861572, "2048": 189.45616722106934 }, "C": { "1": 19.16535973548889, "512": 57.432241439819336, "2048": 189.46791648864746 } }, "native_source_sha256": { "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC", "ISO", "C", "OFF" ], "median_us": { "REC": { "1": 19.15399968624115, "512": 58.347599506378174, "2048": 187.2648048400879 }, "ISO": { "1": 18.92575979232788, "512": 58.348400592803955, "2048": 189.94808197021484 }, "C": { "1": 18.810960054397583, "512": 57.877280712127686, "2048": 189.2743968963623 }, "OFF": { "1": 18.89079988002777, "512": 57.780399322509766, "2048": 189.35040473937988 } }, "native_source_sha256": { "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "OFF": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 1, "C": { "median_us": 19.066799879074097, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 18.89079988002777, "paired_change_pct": [ 0.02727518665761597, -1.6133257871817297, 0.42443248722716653 ], "geomean_change_pct": -0.39112503768712337, "repeats_faster": 1 }, "REC": { "median_us": 19.093680381774902, "paired_change_pct": [ 0.10489614765576416, -0.3740047392966961, 1.8236157583215507 ], "geomean_change_pct": 0.5137560897143789, "repeats_faster": 1 }, "ISO": { "median_us": 18.92575979232788, "paired_change_pct": [ -0.1174818097169239, -1.6329437909360256, 0.6102811212097548 ], "geomean_change_pct": -0.3844416857407329, "repeats_faster": 2 } }, { "tokens": 512, "C": { "median_us": 57.877280712127686, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 58.175599575042725, "paired_change_pct": [ 0.41674780738973904, 1.5885123178263294, -0.1673910529760292 ], "geomean_change_pct": 0.6099788445186993, "repeats_faster": 1 }, "REC": { "median_us": 57.85936117172241, "paired_change_pct": [ -0.7456743283811518, 0.7436933004793866, 0.8126138416726603 ], "geomean_change_pct": 0.26762516937977043, "repeats_faster": 1 }, "ISO": { "median_us": 58.00119876861572, "paired_change_pct": [ -1.0067965609612206, 0.9906584081218028, 0.8139979537385988 ], "geomean_change_pct": 0.26187141316074136, "repeats_faster": 1 } }, { "tokens": 2048, "C": { "median_us": 189.46791648864746, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "OFF": { "median_us": 189.62632179260254, "paired_change_pct": [ 0.04803711985792436, 0.22978634430270706, 0.040157487892678034 ], "geomean_change_pct": 0.1059553436830285, "repeats_faster": 0 }, "REC": { "median_us": 187.5812816619873, "paired_change_pct": [ -0.9040156254732867, -0.9957542478033266, -1.0617347561143076 ], "geomean_change_pct": -0.9871893302404122, "repeats_faster": 3 }, "ISO": { "median_us": 189.61527824401855, "paired_change_pct": [ 0.04221047433623504, -0.006201191101828929, 0.3559303766908384 ], "geomean_change_pct": 0.13051797680971866, "repeats_faster": 1 } } ] } }, "gate_vs_no_gate_shipped": { "3a1f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "REC": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64" } }, "REC_nogate": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64", "AFDEP_COUNT_FIRST_UNGATE": "1" } } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1024", "2048", "--ffn-ranks", "1", "--attention-ranks", "3" ], "processes": [ { "order": [ "C", "REC", "REC_nogate" ], "median_us": { "C": { "1024": 171.02272033691406, "2048": 322.0693588256836 }, "REC": { "1024": 164.20031547546387, "2048": 287.64472007751465 }, "REC_nogate": { "1024": 162.72111892700195, "2048": 285.92967987060547 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC_nogate": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC", "REC_nogate", "C" ], "median_us": { "REC": { "1024": 164.18392181396484, "2048": 287.6095962524414 }, "REC_nogate": { "1024": 162.7725601196289, "2048": 285.77839851379395 }, "C": { "1024": 171.29159927368164, "2048": 321.7705535888672 } }, "native_source_sha256": { "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC_nogate": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC_nogate", "C", "REC" ], "median_us": { "REC_nogate": { "1024": 163.0851936340332, "2048": 286.0207939147949 }, "C": { "1024": 171.1695957183838, "2048": 322.0859146118164 }, "REC": { "1024": 164.27311897277832, "2048": 287.7286434173584 } }, "native_source_sha256": { "REC_nogate": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 1024, "C": { "median_us": 171.1695957183838, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "REC": { "median_us": 164.20031547546387, "paired_change_pct": [ -3.9891804129942976, -4.149460621452006, -4.029031392322658 ], "geomean_change_pct": -4.055915007930711, "repeats_faster": 3 }, "REC_nogate": { "median_us": 162.7725601196289, "paired_change_pct": [ -4.854092715609948, -4.973413284816974, -4.723036267288627 ], "geomean_change_pct": -4.850235697348393, "repeats_faster": 3 } }, { "tokens": 2048, "C": { "median_us": 322.0693588256836, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "REC": { "median_us": 287.64472007751465, "paired_change_pct": [ -10.688579277981203, -10.616557965112595, -10.667113846274834 ], "geomean_change_pct": -10.65742213056905, "repeats_faster": 3 }, "REC_nogate": { "median_us": 285.92967987060547, "paired_change_pct": [ -11.2210857583128, -11.185658436931167, -11.197360412512236 ], "geomean_change_pct": -11.201369425706897, "repeats_faster": 3 } } ] }, "2a2f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "REC": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64" } }, "REC_nogate": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_EPILOGUE_SMS": "64", "AFDEP_COUNT_FIRST_UNGATE": "1" } } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1024", "2048" ], "processes": [ { "order": [ "C", "REC", "REC_nogate" ], "median_us": { "C": { "1024": 106.33567810058594, "2048": 184.40255165100098 }, "REC": { "1024": 106.23647689819336, "2048": 171.03591918945312 }, "REC_nogate": { "1024": 103.9086389541626, "2048": 168.64943504333496 } }, "native_source_sha256": { "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC_nogate": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC", "REC_nogate", "C" ], "median_us": { "REC": { "1024": 106.13071918487549, "2048": 170.95239639282227 }, "REC_nogate": { "1024": 103.92399787902832, "2048": 168.55095863342285 }, "C": { "1024": 105.75432300567627, "2048": 184.51751708984375 } }, "native_source_sha256": { "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC_nogate": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } }, { "order": [ "REC_nogate", "C", "REC" ], "median_us": { "REC_nogate": { "1024": 103.68864059448242, "2048": 168.5066318511963 }, "C": { "1024": 106.03591918945312, "2048": 185.2505588531494 }, "REC": { "1024": 106.0805606842041, "2048": 170.85935592651367 } }, "native_source_sha256": { "REC_nogate": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "C": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663", "REC": "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" } } ], "table": [ { "tokens": 1024, "C": { "median_us": 106.03591918945312, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "REC": { "median_us": 106.13071918487549, "paired_change_pct": [ -0.0932906096660635, 0.35591564344752413, 0.04210035155276515 ], "geomean_change_pct": 0.10139840745653217, "repeats_faster": 1 }, "REC_nogate": { "median_us": 103.9086389541626, "paired_change_pct": [ -2.282431625749859, -1.730733150785435, -2.2136636461620562 ], "geomean_change_pct": -2.0759168091123437, "repeats_faster": 3 } }, { "tokens": 2048, "C": { "median_us": 184.51751708984375, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "REC": { "median_us": 170.95239639282227, "paired_change_pct": [ -7.24861578208823, -7.351670947542866, -7.768507159022309 ], "geomean_change_pct": -7.456537834223454, "repeats_faster": 3 }, "REC_nogate": { "median_us": 168.55095863342285, "paired_change_pct": [ -8.54278667329954, -8.653139662965758, -9.038529819079388 ], "geomean_change_pct": -8.745066389872868, "repeats_faster": 3 } } ] }, "overlap_3a1f": { "native_source_sha256": [ "958652a3538af7478544b8259a6f108b1838e1e14d77ea95e0940f8805218663" ], "per_repeat": { "C": { "2048": { "attention.gemm_us_per_gemm": [ 26.97, 26.88, 26.71 ], "attention.exchange_us": [ 479.33, 480.88, 479.85 ], "attention.solo_exchange_us": [ 327.33, 327.15, 327.49 ], "attention.makespan_us": [ 95866.97, 96175.49, 95970.18 ], "ffn.gemm_us_per_gemm": [ 34.25, 33.4, 34.5 ], "ffn.exchange_us": [ 479.25, 480.77, 479.42 ], "ffn.solo_exchange_us": [ 326.75, 326.54, 327.09 ], "ffn.makespan_us": [ 95849.6, 96153.09, 95883.93 ] } }, "REC": { "2048": { "attention.gemm_us_per_gemm": [ 27.28, 27.09, 27.34 ], "attention.exchange_us": [ 456.98, 460.5, 457.21 ], "attention.solo_exchange_us": [ 293.66, 293.87, 293.56 ], "attention.makespan_us": [ 91395.68, 92099.87, 91441.86 ], "ffn.gemm_us_per_gemm": [ 35.68, 34.56, 35.78 ], "ffn.exchange_us": [ 456.89, 459.67, 456.25 ], "ffn.solo_exchange_us": [ 293.13, 293.11, 293.12 ], "ffn.makespan_us": [ 91378.46, 91934.68, 91249.18 ] } }, "REC_nogate": { "2048": { "attention.gemm_us_per_gemm": [ 27.32, 27.22, 27.32 ], "attention.exchange_us": [ 458.48, 459.7, 459.21 ], "attention.solo_exchange_us": [ 292.07, 291.88, 291.73 ], "attention.makespan_us": [ 91695.52, 91939.84, 91842.05 ], "ffn.gemm_us_per_gemm": [ 35.11, 35.2, 35.57 ], "ffn.exchange_us": [ 458.38, 459.52, 459.02 ], "ffn.solo_exchange_us": [ 291.68, 291.46, 291.23 ], "ffn.makespan_us": [ 91675.71, 91904.96, 91803.36 ] } } } } } }, "rejected_stage_b": { "native_source_sha256": "00a7c096 (an unpublished scratch branch)", "mechanism_3a1f": { "arms": { "C": { "_env": { "AFDEP_PAYLOAD_FIRST": "1" } }, "CF": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1" } }, "CF_laneA": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_LANE_RELEASE": "1" } }, "CF_slotA": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_SLOT_BEFORE_GATE": "1" } }, "CF_ungate": { "_env": { "AFDEP_COUNT_FIRST_MIN_BUCKET": "1", "AFDEP_COUNT_FIRST_UNGATE": "1" } }, "ISO": { "_tree": "/work/src-fin4", "_digest": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, "passthrough": [ "--payload", "bf16", "--routing", "random", "--hidden", "2048", "--capacity", "2048", "--tokens", "1", "8", "32", "128", "512", "1024", "2048", "--ffn-ranks", "1", "--attention-ranks", "3" ], "processes": [ { "order": [ "C", "CF", "CF_laneA", "CF_slotA", "CF_ungate", "ISO" ], "median_us": { "C": { "1": 18.453279733657837, "8": 20.073920488357544, "32": 23.42519998550415, "128": 35.30695915222168, "512": 96.10712051391602, "1024": 171.2020778656006, "2048": 321.3301467895508 }, "CF": { "1": 21.594479084014893, "8": 24.077199697494507, "32": 28.678640127182007, "128": 37.706559896469116, "512": 97.5540018081665, "1024": 172.39368438720703, "2048": 293.3555221557617 }, "CF_laneA": { "1": 20.940959453582764, "8": 23.707040548324585, "32": 27.925280332565308, "128": 36.749199628829956, "512": 97.12096214294434, "1024": 166.08951568603516, "2048": 285.86584091186523 }, "CF_slotA": { "1": 21.467599868774414, "8": 24.185199737548828, "32": 28.902640342712402, "128": 37.67624020576477, "512": 97.07695960998535, "1024": 171.97088241577148, "2048": 293.40975761413574 }, "CF_ungate": { "1": 21.44416093826294, "8": 23.95616054534912, "32": 29.367599487304688, "128": 41.45472049713135, "512": 96.33376121520996, "1024": 171.00191116333008, "2048": 291.81488037109375 }, "ISO": { "1": 18.025919795036316, "8": 20.41983962059021, "32": 23.800159692764282, "128": 35.44480085372925, "512": 96.28151893615723, "1024": 171.38360023498535, "2048": 321.3693618774414 } }, "native_source_sha256": { "C": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_laneA": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_slotA": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_ungate": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a" } }, { "order": [ "CF", "CF_laneA", "CF_slotA", "CF_ungate", "ISO", "C" ], "median_us": { "CF": { "1": 21.361520290374756, "8": 24.117599725723267, "32": 29.04536008834839, "128": 37.82536029815674, "512": 97.80967712402344, "1024": 172.5554370880127, "2048": 293.78623962402344 }, "CF_laneA": { "1": 20.645840167999268, "8": 23.596080541610718, "32": 27.758638858795166, "128": 36.93368077278137, "512": 97.5267219543457, "1024": 166.0309600830078, "2048": 285.79400062561035 }, "CF_slotA": { "1": 21.293680667877197, "8": 23.953440189361572, "32": 29.19975996017456, "128": 37.49248027801514, "512": 96.88591957092285, "1024": 171.81056022644043, "2048": 293.389196395874 }, "CF_ungate": { "1": 21.46615982055664, "8": 23.89472007751465, "32": 29.238959550857544, "128": 41.190080642700195, "512": 95.83552360534668, "1024": 171.0028839111328, "2048": 291.806001663208 }, "ISO": { "1": 18.08080017566681, "8": 20.293279886245728, "32": 23.96791934967041, "128": 35.48200011253357, "512": 95.99839687347412, "1024": 171.3212776184082, "2048": 321.6377639770508 }, "C": { "1": 18.103840351104736, "8": 20.412399768829346, "32": 23.74840021133423, "128": 35.58503985404968, "512": 96.1732006072998, "1024": 171.2716007232666, "2048": 321.3287353515625 } }, "native_source_sha256": { "CF": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_laneA": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_slotA": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_ungate": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138" } }, { "order": [ "CF_laneA", "CF_slotA", "CF_ungate", "ISO", "C", "CF" ], "median_us": { "CF_laneA": { "1": 20.95687985420227, "8": 23.963040113449097, "32": 27.860161066055298, "128": 36.8832004070282, "512": 97.24823951721191, "1024": 166.12319946289062, "2048": 286.09463691711426 }, "CF_slotA": { "1": 20.987119674682617, "8": 23.7609601020813, "32": 28.222479820251465, "128": 36.928160190582275, "512": 97.03239917755127, "1024": 170.2074432373047, "2048": 290.5427932739258 }, "CF_ungate": { "1": 21.003679037094116, "8": 23.860880136489868, "32": 28.49120020866394, "128": 40.21911859512329, "512": 95.76344013214111, "1024": 167.85239219665527, "2048": 288.6939239501953 }, "ISO": { "1": 17.7047997713089, "8": 20.51416039466858, "32": 24.00167942047119, "128": 35.514880418777466, "512": 96.51679992675781, "1024": 171.37831687927246, "2048": 321.3226318359375 }, "C": { "1": 18.07647943496704, "8": 20.472400188446045, "32": 23.871759176254272, "128": 35.34536004066467, "512": 96.28991603851318, "1024": 171.34495735168457, "2048": 321.5105438232422 }, "CF": { "1": 21.558640003204346, "8": 24.059759378433228, "32": 28.749040365219116, "128": 37.60968089103699, "512": 97.56344318389893, "1024": 172.53168106079102, "2048": 293.66488456726074 } }, "native_source_sha256": { "CF_laneA": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_slotA": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF_ungate": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "ISO": "9675783bee5e9a41c3239cb0e8f2342dc5cbac8feb459f72d75f6d6f2700425a", "C": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138", "CF": "00a7c096333ef1e89db73dab2a2ac46428d05e790cfbb72459324ed040a1c138" } } ], "table": [ { "tokens": 1, "C": { "median_us": 18.103840351104736, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 21.558640003204346, "paired_change_pct": [ 17.0224447669737, 17.994413760234185, 19.263488671922644 ], "geomean_change_pct": 18.089887162297337, "repeats_faster": 0 }, "CF_laneA": { "median_us": 20.940959453582764, "paired_change_pct": [ 13.48096249463735, 14.041218700536163, 15.934521042097384 ], "geomean_change_pct": 14.480770618338079, "repeats_faster": 0 }, "CF_slotA": { "median_us": 21.293680667877197, "paired_change_pct": [ 16.334874768188822, 17.61968872299413, 16.101809260962895 ], "geomean_change_pct": 16.683553481082058, "repeats_faster": 0 }, "CF_ungate": { "median_us": 21.44416093826294, "paired_change_pct": [ 16.207857073503785, 18.572410075670653, 16.19341649273096 ], "geomean_change_pct": 16.985908983342913, "repeats_faster": 0 }, "ISO": { "median_us": 18.025919795036316, "paired_change_pct": [ -2.3159023479280894, -0.12726678423521376, -2.056150728881234 ], "geomean_change_pct": -1.50459033447089, "repeats_faster": 3 } }, { "tokens": 8, "C": { "median_us": 20.412399768829346, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 24.077199697494507, "paired_change_pct": [ 19.94268738614753, 18.151711699042504, 17.52290477406635 ], "geomean_change_pct": 18.53468140211385, "repeats_faster": 0 }, "CF_laneA": { "median_us": 23.707040548324585, "paired_change_pct": [ 18.098707036695583, 15.59679806802039, 17.0504674238102 ], "geomean_change_pct": 16.91081821027589, "repeats_faster": 0 }, "CF_slotA": { "median_us": 23.953440189361572, "paired_change_pct": [ 20.480699082053945, 17.34749691674937, 16.0633823262756 ], "geomean_change_pct": 17.949338782021094, "repeats_faster": 0 }, "CF_ungate": { "median_us": 23.89472007751465, "paired_change_pct": [ 19.339720206838496, 17.05982808549029, 16.551454235230167 ], "geomean_change_pct": 17.64411255234979, "repeats_faster": 0 }, "ISO": { "median_us": 20.41983962059021, "paired_change_pct": [ 1.723226573669523, -0.5835662829096644, 0.2039829518675651 ], "geomean_change_pct": 0.44332878257036246, "repeats_faster": 1 } }, { "tokens": 32, "C": { "median_us": 23.74840021133423, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 28.749040365219116, "paired_change_pct": [ 22.426447351266, 22.304491375743773, 20.431176240318205 ], "geomean_change_pct": 21.71726738746933, "repeats_faster": 0 }, "CF_laneA": { "median_us": 27.860161066055298, "paired_change_pct": [ 19.210424456763953, 16.886352814397164, 16.70761614321399 ], "geomean_change_pct": 17.59596313277061, "repeats_faster": 0 }, "CF_slotA": { "median_us": 28.902640342712402, "paired_change_pct": [ 23.382683437485152, 22.95463989291624, 18.225387630103707 ], "geomean_change_pct": 21.498232505616976, "repeats_faster": 0 }, "CF_ungate": { "median_us": 29.238959550857544, "paired_change_pct": [ 25.36755078068822, 23.11970191955446, 19.35107085448784 ], "geomean_change_pct": 22.587542515702165, "repeats_faster": 0 }, "ISO": { "median_us": 23.96791934967041, "paired_change_pct": [ 1.6006681159271396, 0.9243533727859798, 0.5442424383459432 ], "geomean_change_pct": 1.0221441235886308, "repeats_faster": 0 } }, { "tokens": 128, "C": { "median_us": 35.34536004066467, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 37.706559896469116, "paired_change_pct": [ 6.796395956677692, 6.295680581771501, 6.406274678676982 ], "geomean_change_pct": 6.499234011913879, "repeats_faster": 0 }, "CF_laneA": { "median_us": 36.8832004070282, "paired_change_pct": [ 4.084861770140646, 3.7899098167743395, 4.35089744338224 ], "geomean_change_pct": 4.074970776084963, "repeats_faster": 0 }, "CF_slotA": { "median_us": 37.49248027801514, "paired_change_pct": [ 6.710521411170589, 5.360231242647839, 4.47809881720429 ], "geomean_change_pct": 5.512295794591093, "repeats_faster": 0 }, "CF_ungate": { "median_us": 41.190080642700195, "paired_change_pct": [ 17.412321798669605, 15.751115670066174, 13.78896281959323 ], "geomean_change_pct": 15.641309327397646, "repeats_faster": 0 }, "ISO": { "median_us": 35.48200011253357, "paired_change_pct": [ 0.39040944000099387, -0.2895591572715017, 0.47961140562087223 ], "geomean_change_pct": 0.1928974961990848, "repeats_faster": 1 } }, { "tokens": 512, "C": { "median_us": 96.1732006072998, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 97.56344318389893, "paired_change_pct": [ 1.5054881329432668, 1.7015930699923354, 1.3225965893213232 ], "geomean_change_pct": 1.5097746364481912, "repeats_faster": 0 }, "CF_laneA": { "median_us": 97.24823951721191, "paired_change_pct": [ 1.0549079231663372, 1.407378914810864, 0.9952480157064736 ], "geomean_change_pct": 1.152348269716863, "repeats_faster": 0 }, "CF_slotA": { "median_us": 97.03239917755127, "paired_change_pct": [ 1.0091230398781015, 0.7410785531961839, 0.7710912726739849 ], "geomean_change_pct": 0.8403596992860196, "repeats_faster": 0 }, "CF_ungate": { "median_us": 95.83552360534668, "paired_change_pct": [ 0.23582092573579683, -0.3511134077069378, -0.5467612061905824 ], "geomean_change_pct": -0.22123803013672783, "repeats_faster": 2 }, "ISO": { "median_us": 96.28151893615723, "paired_change_pct": [ 0.18146254024535846, -0.18175929751933495, 0.23562580338514216 ], "geomean_change_pct": 0.07827130161046636, "repeats_faster": 1 } }, { "tokens": 1024, "C": { "median_us": 171.2716007232666, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 172.53168106079102, "paired_change_pct": [ 0.6960233990512155, 0.7495909183568905, 0.6925933085212943 ], "geomean_change_pct": 0.7127324942780344, "repeats_faster": 0 }, "CF_laneA": { "median_us": 166.08951568603516, "paired_change_pct": [ -2.986273439729492, -3.059842156042203, -3.0475118553248803 ], "geomean_change_pct": -3.0312144861249446, "repeats_faster": 3 }, "CF_slotA": { "median_us": 171.81056022644043, "paired_change_pct": [ 0.44906262806836406, 0.31468118526238875, -0.6638737036451792 ], "geomean_change_pct": 0.03205757190378744, "repeats_faster": 1 }, "CF_ungate": { "median_us": 171.00191116333008, "paired_change_pct": [ -0.11691838368202978, -0.15689513672962896, -2.0383238637486256 ], "geomean_change_pct": -0.7747794888786519, "repeats_faster": 3 }, "ISO": { "median_us": 171.37831687927246, "paired_change_pct": [ 0.10602813449920312, 0.029004747390581898, 0.01946922051485611 ], "geomean_change_pct": 0.05149319710002764, "repeats_faster": 0 } }, { "tokens": 2048, "C": { "median_us": 321.3301467895508, "paired_change_pct": [ 0.0, 0.0, 0.0 ], "geomean_change_pct": 0.0, "repeats_faster": 0 }, "CF": { "median_us": 293.66488456726074, "paired_change_pct": [ -8.705882380874307, -8.57143874711519, -8.660885246516282 ], "geomean_change_pct": -8.64608587777651, "repeats_faster": 3 }, "CF_laneA": { "median_us": 285.86584091186523, "paired_change_pct": [ -11.036719160033325, -11.058685643869959, -11.015472925080383 ], "geomean_change_pct": -11.036960992332489, "repeats_faster": 3 }, "CF_slotA": { "median_us": 293.389196395874, "paired_change_pct": [ -8.689003958816532, -8.695001685772086, -9.631954890518813 ], "geomean_change_pct": -9.006401530116992, "repeats_faster": 3 }, "CF_ungate": { "median_us": 291.806001663208, "paired_change_pct": [ -9.185339973030137, -9.187704192111534, -10.207012026046824 ], "geomean_change_pct": -9.527967578692609, "repeats_faster": 3 }, "ISO": { "median_us": 321.3693618774414, "paired_change_pct": [ 0.01220398654855348, 0.09617211020676031, -0.058446601803519016 ], "geomean_change_pct": 0.016623197465848172, "repeats_faster": 1 } } ] }, "traces": { "C-t2048": [ "## C T=2048", "anchor device 1 dispatch start = 0; period median 321.45 us", " dev0 0:dispatch_impl start -0.22 end 79.39 dur 79.62", " dev0 1:dispatch_copy_epilogue start 69.85 end 144.92 dur 75.21", " dev0 2:combine_impl start 144.95 end 319.33 dur 174.30", " dev1 0:dispatch_impl start 0.00 end 79.26 dur 79.26", " dev1 1:combine_impl start 79.81 end 319.58 dur 239.78", " dev2 0:dispatch_impl start -0.12 end 79.34 dur 79.49", " dev2 1:combine_impl start 79.89 end 319.64 dur 239.73", " dev3 0:dispatch_impl start -0.04 end 79.30 dur 79.36", " dev3 1:combine_impl start 79.84 end 319.60 dur 239.73" ], "C-t512": [ "## C T=512", "anchor device 1 dispatch start = 0; period median 96.80 us", " dev0 0:dispatch_impl start 0.08 end 23.15 dur 23.07", " dev0 1:dispatch_copy_epilogue start 21.78 end 39.92 dur 17.98", " dev0 2:combine_impl start 39.97 end 94.82 dur 54.85", " dev1 0:dispatch_impl start 0.00 end 22.88 dur 22.88", " dev1 1:combine_impl start 23.42 end 95.07 dur 71.62", " dev2 0:dispatch_impl start 0.11 end 22.96 dur 22.85", " dev2 1:combine_impl start 23.51 end 94.93 dur 71.49", " dev3 0:dispatch_impl start -0.06 end 22.96 dur 23.01", " dev3 1:combine_impl start 23.52 end 94.93 dur 71.49" ], "CF-t2048": [ "## CF T=2048", "anchor device 1 dispatch start = 0; period median 291.62 us", " dev0 0:dispatch_impl start -0.08 end 90.43 dur 90.46", " dev0 1:afdep_ffn_stream_epilogue_impl start 6.33 end 114.93 dur 108.62", " dev0 2:combine_impl start 114.94 end 289.54 dur 174.50", " dev1 0:dispatch_impl start 0.00 end 87.81 dur 87.81", " dev1 1:combine_impl start 88.35 end 289.82 dur 201.31", " dev2 0:dispatch_impl start 0.01 end 88.45 dur 88.50", " dev2 1:combine_impl start 89.00 end 289.83 dur 200.46", " dev3 0:dispatch_impl start -0.10 end 88.45 dur 88.62", " dev3 1:combine_impl start 89.03 end 289.80 dur 200.70" ], "CF-t512": [ "## CF T=512", "anchor device 1 dispatch start = 0; period median 97.82 us", " dev0 0:dispatch_impl start 0.04 end 30.46 dur 30.40", " dev0 1:afdep_ffn_stream_epilogue_impl start 5.94 end 41.15 dur 35.26", " dev0 2:combine_impl start 41.17 end 95.81 dur 54.75", " dev1 0:dispatch_impl start 0.00 end 27.94 dur 27.94", " dev1 1:combine_impl start 28.48 end 96.10 dur 67.62", " dev2 0:dispatch_impl start 0.02 end 28.24 dur 28.19", " dev2 1:combine_impl start 28.81 end 95.97 dur 67.31", " dev3 0:dispatch_impl start -0.10 end 28.13 dur 28.26", " dev3 1:combine_impl start 28.71 end 95.99 dur 67.30" ], "CFungate-t2048": [ "## CFungate T=2048", "anchor device 1 dispatch start = 0; period median 291.07 us", " dev0 0:dispatch_impl start -0.25 end 89.55 dur 89.76", " dev0 1:afdep_ffn_stream_epilogue_impl start 16.30 end 114.10 dur 97.98", " dev0 2:combine_impl start 114.11 end 288.94 dur 174.62", " dev1 0:dispatch_impl start 0.00 end 86.27 dur 86.27", " dev1 1:combine_impl start 86.82 end 289.17 dur 202.11", " dev2 0:dispatch_impl start -0.15 end 87.25 dur 87.39", " dev2 1:combine_impl start 87.79 end 289.15 dur 201.26", " dev3 0:dispatch_impl start -0.02 end 87.12 dur 87.10", " dev3 1:combine_impl start 87.68 end 289.23 dur 201.47" ], "CFungate-t512": [ "## CFungate T=512", "anchor device 1 dispatch start = 0; period median 96.42 us", " dev0 0:dispatch_impl start 0.03 end 28.63 dur 28.58", " dev0 1:afdep_ffn_stream_epilogue_impl start 15.77 end 39.63 dur 23.74", " dev0 2:combine_impl start 39.68 end 94.40 dur 54.59", " dev1 0:dispatch_impl start 0.00 end 26.21 dur 26.21", " dev1 1:combine_impl start 26.75 end 94.67 dur 67.87", " dev2 0:dispatch_impl start -0.15 end 26.22 dur 26.30", " dev2 1:combine_impl start 26.79 end 94.56 dur 67.74", " dev3 0:dispatch_impl start 0.01 end 26.43 dur 26.40", " dev3 1:combine_impl start 27.01 end 94.52 dur 67.55" ], "CFwarp-t2048": [ "## CFwarp T=2048", "anchor device 1 dispatch start = 0; period median 300.27 us", " dev0 0:dispatch_impl start 0.07 end 98.71 dur 98.61", " dev0 1:afdep_ffn_stream_epilogue_impl start 13.93 end 123.78 dur 109.89", " dev0 2:combine_impl start 123.79 end 298.26 dur 174.46", " dev1 0:dispatch_impl start 0.00 end 95.81 dur 95.81", " dev1 1:combine_impl start 96.35 end 298.53 dur 202.08", " dev2 0:dispatch_impl start 0.02 end 96.56 dur 96.61", " dev2 1:combine_impl start 97.11 end 298.56 dur 201.42", " dev3 0:dispatch_impl start 0.03 end 96.27 dur 96.32", " dev3 1:combine_impl start 96.83 end 298.55 dur 201.57" ], "CFwarp-t512": [ "## CFwarp T=512", "anchor device 1 dispatch start = 0; period median 99.97 us", " dev0 0:dispatch_impl start -0.10 end 32.06 dur 32.13", " dev0 1:afdep_ffn_stream_epilogue_impl start 7.58 end 42.89 dur 35.39", " dev0 2:combine_impl start 42.89 end 97.97 dur 54.98", " dev1 0:dispatch_impl start 0.00 end 29.60 dur 29.60", " dev1 1:combine_impl start 30.14 end 98.24 dur 68.06", " dev2 0:dispatch_impl start 0.05 end 29.88 dur 29.79", " dev2 1:combine_impl start 30.45 end 98.11 dur 67.71", " dev3 0:dispatch_impl start -0.13 end 29.79 dur 29.92", " dev3 1:combine_impl start 30.35 end 98.08 dur 67.76" ] } }, "scope": "communication only; synthetic GEMM load; no serving claim" }