What happened

Anthropic told IPO investors that its Claude Mythos Preview model completed the software work needed to get an Anthropic model running on AMD's MI355X GPUs in about three days. According to the company, that job previously took six to 12 months and a team of dozens of engineers.

Why it matters

The low-level code that ties an AI model to a specific company's chips has long been part of NVIDIA's CUDA moat. Anthropic's claim suggests a model that can write such code in days could erode that advantage.

Details

  • The three days cover one Anthropic model running on MI355X GPUs at a "useful level of performance."
  • Mythos ported or wrote almost 100% of the kernels.
  • Anthropic already runs its models on NVIDIA and AMD GPUs, Google TPUs and AWS Trainium, managed as one fungible pool of compute.
  • The company began working with AMD GPUs this year, with an agreement for up to 2 gigawatts of capacity.

Competitive backdrop

AMD's CUDA equivalent is ROCm, with HIP as its programming layer, and CUDA code has to be converted to run there. AMD executives have said AI is getting better at this conversion and that generated kernels can sometimes outperform hand-tuned versions.