The next GPU kernel may be written by a model trained to write kernels
KernelZero beats frontier models at CUDA and Triton generation, while new hardware work removes the communication and memory costs that slow inference down.
- KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA kernel generation and DeepSeek-V4-Pro on Triton kernel generation. Self-reported m:…9778
- SURMA cuts communication submission latency from 1.2 microseconds to 400 nanoseconds. Self-reported m:…9779
- SURMA saturates attainable bandwidth at 512 B granularity with a single SM, where NVLink requires over 14% of SMs. Self-reported m:…9779
- TileBench covers 45 operators with matched semantics for Triton and cuTile. Self-reported m:…9766
- A coding agent wrote every line of a plasma physics simulation that ran across 98% of a Cerebras wafer. Self-reported m:…0611
A small model beats the frontier at kernels
A 7-billion-parameter model built to write GPU kernels now beats the frontier models at their own game, according to a preprint. KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA kernel generation and surpasses DeepSeek-V4-Pro on Triton kernel generation. The framework co-evolves a proposer and a coder, and it uses Correctness-Aware Group Relative Policy Optimization, known as CA-GRPO, which optimizes performance only after correctness becomes sufficiently reliable. m:…9778
That ordering is the point. Kernel generation has always been a correctness problem first and a speed problem second: a fast kernel that computes the wrong answer is not a kernel. The preprint's bet is that a small model trained specifically for the task, with correctness enforced before performance, can beat a generalist many times its size, and the reported results say the bet pays. m:…9778
The plumbing underneath
Below the kernel, the communication layer is getting the same treatment. SURMA, a preprint on an on-die communication engine for SuperPOD, reduces communication submission latency from 1.2 microseconds to 400 ns, and saturates attainable bandwidth at 512 B granularity with a single SM, whereas NVLink requires over 14% of SMs. It occupies under 0.5% of GPU area and accelerates inference on four models by 1.05 to 1.26 times. m:…9779
Choosing between kernel frameworks is getting a measuring stick. TileBench, a preprint on tile-based programming models, provides 45 operators covering diverse AI-kernel patterns with matched semantics for Triton and cuTile. Its findings split the field: Triton is more token-efficient than cuTile for LLM-generated kernels under iterative refinement, cuTile excels on Tensor-Core/TMA-friendly kernels, and Triton is superior for irregular, streaming, and bandwidth-bound operators. m:…9766
Hardware that meets the model halfway
Hardware is adapting to the models, not just the other way round. A preprint on an integrated GPU core reports that RescuCore achieves 1.4 to 9.4 times speedup and 1.4 to 5.0 times energy reduction during prefill compared to conventional GPU execution, and 1.3 to 2.8 times speedup with 1.2 to 2.6 times energy reduction during autoregressive decoding, at an estimated area overhead of 5.9% of an A100 die for state size 32 and MIMO rank 4. m:…9774
Cerebras offers a live example of the loop closing. The company says a researcher ran a plasma physics simulation across 98% of a Cerebras wafer, and a coding agent wrote every line of the simulation. Its Tungsten language and compiler allow programming the Wafer-Scale Engine at a high level, the company says, and the Wafer-Scale Engine contains 900,000 cores. The company reports SNAP running at 1.6 PFLOP/s on a single CS-4 system and wafer-scale molecular dynamics running 748 times faster than the Frontier supercomputer. m:…0611
The thread running through these is that the stack below the model is becoming a software problem: kernels written by models, communication offloaded to on-die engines, and frameworks measured on matched operators. The prize is not a faster chip but a shorter path from an idea to a kernel that runs correctly at speed. m:…9778 m:…9779
What to watch
What to watch is whether the small specialist stays ahead once the frontier models absorb the same trick. KernelZero-7B surpasses Claude-4.5-Sonnet on CUDA kernel generation, the preprint reports, and surpasses DeepSeek-V4-Pro on Triton kernel generation, but those are point-in-time results, and the framework's edge is a training method, CA-GRPO, that optimizes performance only after correctness becomes sufficiently reliable. The durable question is whether correctness-gated training becomes the standard way to build code-writing models, or a one-off win. m:…9778
What is still unresolved
- KernelZero's reported wins are against named frontier models on kernel generation; the preprint does not show the kernels running in production serving.
- The Cerebras speed figures are the company's own numbers and have not been independently verified.
- TileBench covers 45 operators; it does not settle which framework wins on workloads outside that set.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.