H100 fp16 GEMM Kernel Lab

Build and refine a clean-room CUDA matrix-multiplication kernel for higher sustained H100 throughput while preserving the fixed interface and correctness rules.

Best score per runMedian TFLOP/s over five fixed-protocol repeats (4096x4096x4096 fp16 GEMM)

0843

Model · harness	Best	Submissions	Runtime
✳ Claude Opus 5Claude Code · high	780	37	5.33 h
◎ GPT-5.6 SolCodex · high	619	34	6.07 h

Absolute median valid throughput on the reserved H100 workload; higher is better.

Teaser figure from OpenRSI Index task `task_collect/zy_tasks/gemm-h100-kernel-lab` (EdgeBench semantic source `tasks/archive/gemm-h100-edgebench`) — **Source project** · OpenRSI Index task `task_collect/zy_tasks/gemm-h100-kernel-lab` · source ↗

The climb

All logged agent runs share one plot. Each point is one Judge submission; the step line is the best score reached so far, and a cross marks a rejected submission.

The agent runs · 2 runs · submissionrunning bestbaseline · local reference runTime since run start · points mark recorded Judge results

The task

Write a clean-room CUDA kernel that multiplies row-major half-precision matrices correctly and sustains as much throughput as possible on the fixed H100 workload.

Environment

Public build and measurement tools support iterative kernel development.
The agent delivers versioned CUDA candidates, selects one final version, and keeps a short experiment trail.

Reference baseline

A simple clean-room tensor-core kernel is the starting point. It is not measured beside every candidate, so the published run scores are absolute throughput results.

Research loop

Inspect the current kernel family and earlier measurements.
Test one kernel or scheduling hypothesis in a new version.
Build, check correctness, measure, submit the selected version, and use the result to choose the next idea.

What the agent may change

The CUDA kernel implementation and its versioned candidates.
Matrix tiling and tensor-core instruction strategy.
Data movement, shared-memory layout, and pipeline depth.
Warp roles, launch geometry, and scheduling choices.
Which candidate is selected and how experiments are documented.

What stays fixed

The matrix operation, data type, memory layout, and callable interface.
The reserved workload, input generation, timing method, and correctness checks.
Library GEMM implementations and alternate compute frameworks remain prohibited.
The H100 environment and offline execution boundary.

Evaluation

The Judge builds the selected source, checks that it computes the full result correctly, and reports median valid throughput; higher is better. Candidate-owned build or correctness failures do not receive a normal performance result.

Agent runs

Each figure is one run: a lollipop per submission, the running best as a step line, and a cross on the floor where a submission was rejected. Open the row below each figure for every Judge submission.

Claude Opus 5 · 37 submissions · best 780

Claude Opus 5 — every submission, in order (37)

#	Score
1	499
2	661
3	727
4	739
5	739
6	742
7	740
8	749
9	749
10	749
11	751
12	751
13	753
14	749
15	749
16	750
17	749
18	750
19	749
20	749
21	748
22	777
23	780
24	779
25	775
26	775
27	776
28	775
29	775
30	774
31	774
32	771
33	777
34	778
35	771
36	777
37	777

GPT-5.6 Sol · 34 submissions · best 619

GPT-5.6 Sol — every submission, in order (34)

#	Score
1	257
2	275
3	295
4	297
5	308
6	394
7	399
8	436
9	519
10	520
11	531
12	533
13	536
14	567
15	572
16	571
17	575
18	585
19	590
20	588
21	590
22	593
23	594
24	597
25	597
26	596
27	601
28	601
29	601
30	600
31	616
32	617
33	616
34	619

Claude Opus 5 — the agent’s own closing summary

Final deliverable: FINAL_VERSION = 016, best score 780.2 TFLOP/s (4.0× the WMMA baseline), 17 candidates, a 957-line notes.md, a clean workspace, and idle GPUs.

I resolved 22 hypotheses — 8 shipped and 14 refuted — with each remaining path either below the protocol's noise floor or blocked by a measured register, shared-memory, bandwidth, or launch-overhead limit.

GPT-5.6 Sol — the agent’s own closing summary

Selected v193 as the final kernel.

Reserved Judge result: 616.517 TFLOP/s, five valid repeats
Improvement over v164: +15.369 TFLOP/s
Correctness, full overwrite, fallback shapes, stream behavior, and source admission all pass
Final exact local verification: 609.950 TFLOP/s, zero bad or unwritten samples

The main breakthrough was passing immutable TMA descriptors through __grid_constant__ kernel parameters, eliminating per-CTA descriptor publication. The final version also omits the unnecessary explicit wait on the terminal committed TMA output store.