Skip to content

Topic

GPU kernel optimization

Engineering practice of optimizing GPU kernels for throughput, covering memory access, parallelism, and hardware-specific tuning.

Current stories

leadershipOne report1 publisher

CUDA engineers now supervise the AI that writes their GPU kernels

CUDA engineers, among tech's most sought-after specialists, increasingly supervise AI that writes and tests hundreds of GPU kernels. At base salaries Nvidia advertises as high as $431,250, employers are now paying for judgement about code the engineers did not write.

Reality

Evidence50
Adoption30
Hype gap+35
Incentives70
Confidence45
buildOne report1 publisher

vLLM measured its portability layer at 3.4 percent below native throughput on an H100

PyTorch says vLLM's frontier models now ship as hardware-specific flat definitions that torch.compile cannot trace, and the new HW agnostic layers are what users on other accelerators get instead. The overhead figure came from an H100.

Publishers:pytorch.org

Reality

Evidence55
Adoption45
Hype gap+10
Incentives60
Confidence55
buildOne report1 publisher

Meta's MXFP8 FlashAttention-4 kernel, tackling TMEM scale-factor constraints, hits up to 1.6x over BF16

Meta has open-sourced an MXFP8 forward and backward pass for FlashAttention-4 that it already runs in production ads training. The reported 1.6x forward gain over BF16 sits well under the 2-4x the block-scaled MMA instruction advertises.

Publishers:pytorch.org

Reality

Evidence58
Adoption47
Hype gap+12
Incentives55
Confidence60
leadershipOne report1 publisher

A DeepSeek engineer says AI will probably match or surpass the kernels he writes by hand

Liu Shengyu, who helped write a component of DeepSeek's latest model, wrote that AI-written operators will probably be as good as his own or better. He expects to keep his job and to lose the hands-on part of it.

Reality

Evidence48
Adoption20
Hype gap+24
Incentives66
Confidence57
buildOne report1 publisher

NVIDIA's unoptimized DeepSeek-V3 baseline spends 84% of kernel time moving tokens between GPUs

A JAX and Transformer Engine version of the same GB200 run reports 1,068 TFLOPS per GPU, a 10.4x gain. The 84% share bounds what fixing communication alone could have bought, so the expert matmuls got faster too.

Reality

Evidence45
Adoption
Not scored
Hype gap+35
Incentives85
Confidence55
buildOne report1 publisher

Base Compute hands kernel tuning to agents; the carryover claim is the unmeasured part

Its August 20 bake-off priced three agents on one Nemotron port. The claim that findings transfer to the next model-chip pair is the one with no number attached.

Publishers:runtimewire.com

Reality

Evidence42
Adoption15
Hype gap+28
Incentives80
Confidence45
buildConfirmed4 publishers

Nvidia says the harness, not the model, took Claude Opus 5 from 30.2% to 100% on ARC-AGI-3

A five-person Nvidia team published the architecture behind its AVO agent alongside a perfect public-set score. The model did not change. The system around it did.

Perspective Coverage

4 publishers
Builder
Builder 50%
Operator
Operator 29%
Investor
Investor 21%

Reality

Evidence58
Adoption20
Hype gap+34
Incentives82
Confidence72
buildOne report1 publisher

1,500 submissions in 14 days: what a 12th-place GPU kernel says about agent loops

Sankalp placed 12th of 183 on GPU Mode's B200 QR benchmark using OpenAI Codex. The transferable part is the checker, not the ranking.

Publishers:runtimewire.com

Reality

Evidence45
Adoption22
Hype gap+12
Incentives58
Confidence52