According to Tom's Hardware, the three days cover prompt processing for one Anthropic model on AMD Instinct MI355X GPUs "at a useful level of performance," with Mythos porting or writing "almost 100% of the kernels" with basically no human intervention. Tom Brown, Anthropic co-founder and chief compute officer, said the work used to require a team of "dozens of engineers." The comparison is based on the company's past experience bringing up new accelerators with traditional engineering methods.
Kernels are low-level code that ties an AI model to one maker's chips, and they are a core part of Nvidia's advantage with its CUDA platform. AMD has an equivalent, ROCm, but moving to it requires adapting the code. Anthropic already runs its models on several kinds of chips: Nvidia and AMD GPUs, Google TPUs and AWS Trainium, and manages them as one pool of compute. The company began working with AMD GPUs this year, under an agreement for up to 2 gigawatts.
The figures come from the company's statements to investors, not from independent testing. It has not been published whether stages of running the model beyond prompt processing were adapted the same way.