Anthropic told IPO investors that its Claude Mythos Preview model completed the low-level software work needed to run one of its models on AMD MI355X GPUs at useful performance in about three days, with basically no human intervention, versus six to 12 months and dozens of engineers previously. Mythos ported or wrote almost 100% of the kernels, a result that directly challenges Nvidia's CUDA moat.
GPU Kernel
GPU 内核
Low-level parallel compute functions on GPUs that determine real model performance on specific chips.
Curated Intelligence
Berkeley researchers introduced K-Search, an AI agent that automatically transfers decades of CUDA kernel optimization expertise to Apple Silicon's MLX framework. By employing architecture-aware strategy translation rather than instruction-level copying, the tool significantly lowers the barrier and time cost for developing high-performance kernels like Flash Attention and Mamba across hardware platforms.
The AI model Fable wrote the first genuine and fastest megakernel on the KernelBench-Mega benchmark, achieving an 18.71x speedup. By utilizing a single cooperative kernel launch per token, it outperformed other models that rely on multi-step decompositions, marking a significant breakthrough in automating core AI research tasks like low-level compute optimization.
An article titled 'From Zero to GPU: A Guide to Building and Scaling Production-Ready CUDA Kernels' has been published. Due to missing body content, it is only confirmed as a tutorial for developing high-performance CUDA code, with specific technical details unavailable.
The source text contains only the title 'Creating custom kernels for the AMD MI300' with no body content, code, or technical details. It is impossible to evaluate implementation methods or performance data due to missing information.