Anthropic demonstrated that it could in just a few days (days, not months) optimize AI models for AMD – the implications of this are significant as it makes it easier for companies to adopt other GPUs. It means you don't have to wait half a year to try new chips and could transform the way data centers approach building out infrastructure.
Claude Mythos Preview generates nearly all GPU kernel code for AMD Instinct MI355X
They revealed this as part of their investor roadshow ahead of their IPO. What they did was use their Claude Mythos Preview model to do the optimization, and specifically focus on adapting the way that prompts get processed for the AMD Instinct MI355X GPU. This is specifically their latest chip for their server offerings.
They claimed that the process to get it to an "usable" state only took ~3 days. Contrast this with the traditional approach, which can take 6-12 months and involve dozens of engineers. One specific thing they did is that for this process, Claude managed to port/write ~100% of the code for the GPU kernel. This was all done with minimal human involvement in the process.
Just to be clear – this is just their prompt processing for one model, it's not the full training process or even the entire inference stack. Another thing that they revealed is that they have a partnership with AMD – they're going to deploy up to 2GW of AMD Instinct MI450 GPUs starting in the 1H2027.
Anthropic has demonstrated that AI models can be automated to do complex tasks like adapting hardware for them. You can see that it only takes a few days (3 days specifically) – vs the traditional approach of 6 months. So it really sets them up for success for their large scale deployments that they're going to do with AMD.
Comments (0)