ywkang

ywkang pushed to master at ywkang/kernbench2

2026-07-23 00:46:56 +00:00

ae131aa2af report(S5.2): explain why dense mapping wins throughput (per-PE efficiency)

ywkang pushed to master at ywkang/kernbench2

2026-07-23 00:39:15 +00:00

baf22f01ab plot: extend batch-scaling curves with saturated-throughput ceiling

ywkang pushed to master at ywkang/kernbench2

2026-07-23 00:37:12 +00:00

13f5cbc317 gqa decode: fix cube_base output-store; measure batch scaling on one SIP

ywkang pushed to master at ywkang/kernbench2

2026-07-22 23:25:51 +00:00

8c108154a7 plot: fix Fig 16 suptitle to match KV-footprint panel

ywkang pushed to master at ywkang/kernbench2

2026-07-22 23:25:08 +00:00

711a9a257f gqa short-context: exclude KV deploy from wall; correct HBM peak; batch scaffold

ywkang pushed to master at ywkang/kernbench2

2026-07-22 17:33:58 +00:00

cb6d21f145 report(1H): tighten §5 GQA (mapping tie, taxonomy bridge, dedup, Summary)

ywkang pushed to master at ywkang/kernbench2

2026-07-22 17:22:07 +00:00

d88e6cc72e report(1H): rebuild main.pdf

ywkang pushed to master at ywkang/kernbench2

2026-07-22 17:20:34 +00:00

edb30326ce report(1H): add Agentic Workloads (§6) and HW-Spec Search (§7) sections

ywkang pushed to feat/gqa-composite-commands at ywkang/kernbench2

2026-06-24 02:39:48 +00:00

a4a2683aad paper(gqa): single-rank decode-streaming study — composite's masked bandwidth win

ywkang pushed to master at ywkang/kernbench2

2026-06-24 02:38:09 +00:00

faf011dae5 paper(ipcq): regen alternatives figures + comparison plot + allreduce section

1dade267ff experiment(gqa-decode): 16x16x16 MAC-block dispatch model

Compare 2 commits »

ywkang pushed to feat/gqa-composite-commands at ywkang/kernbench2

2026-06-24 02:19:57 +00:00

faf011dae5 paper(ipcq): regen alternatives figures + comparison plot + allreduce section

ywkang pushed to feat/gqa-composite-commands at ywkang/kernbench2

2026-06-24 01:24:29 +00:00

1dade267ff experiment(gqa-decode): 16x16x16 MAC-block dispatch model

ywkang pushed to master at ywkang/kernbench2

2026-06-24 00:06:18 +00:00

24c705419e paper(gqa): two-regime "use of composite commands" — decode + compute-bound prefill

cd6f0ed91d feat(gqa): Case-6 composite-command decode variants + on-chip-operand DMA fix

Compare 2 commits »

ywkang pushed to feat/gqa-composite-commands at ywkang/kernbench2

2026-06-24 00:04:29 +00:00

24c705419e paper(gqa): two-regime "use of composite commands" — decode + compute-bound prefill

ywkang created branch feat/gqa-composite-commands in ywkang/kernbench2

2026-06-23 23:07:50 +00:00

ywkang pushed to feat/gqa-composite-commands at ywkang/kernbench2

2026-06-23 23:07:50 +00:00

cd6f0ed91d feat(gqa): Case-6 composite-command decode variants + on-chip-operand DMA fix

ywkang pushed to master at ywkang/kernbench2

2026-06-22 07:55:55 +00:00

e9a5c438e3 paper(1H): 6-case GQA long-ctx + IPCQ design alternatives + GEMM terminology

ywkang pushed to master at ywkang/kernbench2

2026-06-17 22:12:36 +00:00

b6315c3c90 paper(1H): reflect D8 single-op cost model in §2/§3.4 + figure/diagram regen

ywkang pushed to master at ywkang/kernbench2

2026-06-17 22:06:39 +00:00

821bbf26a2 gemm(bench): composite-vs-async dispatch comparison harness + single-op-model sweep

ywkang pushed to master at ywkang/kernbench2

2026-06-17 22:04:57 +00:00

2d8271c981 perf(cost-model): D8 single-op-cmd fast-path (FIXED=8 single-op / 40 composite)