The AI model Fable produced CUDA code described by KernelBench-Mega maintainers and its official leaderboard as the first genuine — and fastest — megakernel submitted to the benchmark. This outcome is being read as an example of AI systems increasingly handling tasks that are central to AI research and development, such as kernel design and low-level performance engineering.
Results and comparisons
- Fable's CUDA kernel achieved an 18.71× speedup on an NVIDIA RTX PRO 6000 Blackwell compared with an optimized PyTorch baseline in KernelBench-Mega tests.
- For context, other recent submissions recorded smaller gains: Claude Opus 4.8 (writing Triton) 14.4×, GLM-5.2 (Triton) 11.14×, and GPT 5.5 (Triton) 4.34×.
What sets this submission apart
A notable technical detail is that torch.profiler reported exactly one cooperative kernel launch per decoded token for Fable's solution. By contrast, other high-scoring entries decomposed the workload into anywhere from four to fourteen kernel launches per token. Reducing the number of kernel launches can lower synchronization overheads and make more efficient use of GPU resources, which helps explain the observed performance advantage.
Broader significance for AI R&D automation
Designing and optimizing kernels is a fundamental input task for accelerating model training and inference. If AI systems can autonomously create and refine these low-level components, they can take on more of the engineering work previously performed by human specialists. Observers argue that progress on these fronts is meaningful beyond raw speed improvements: it signals growing capability for AI systems to perform core R&D tasks, which in turn is relevant to discussions about recursive self-improvement and the automation of scientific and engineering workflows. Benchmarks like KernelBench-Mega therefore serve both as performance tests and as indicators of how effectively models can build parts of their own toolchain.
Where to find the results and further analysis
The KernelBench-Mega official leaderboard lists the submissions and rankings. Elliot Arledge, one of the benchmark maintainers, has published commentary and analysis on the results and methodology.
