Anthropic says its Claude Mythos Preview model cut the software work needed to run an AI workload on AMD accelerators from an estimated six to 12 months to about three days. The task was to get prompt processing for one model running on AMD Instinct MI355X GPUs at useful performance. Anthropic says Mythos wrote or ported almost all the required low-level kernels with virtually no human intervention; previously, the effort would have involved dozens of engineers. Such kernels adapt models to specific chips and have traditionally made moving workloads between hardware platforms time-consuming.
The result could weaken Nvidia’s CUDA advantage by making it faster and cheaper to support rival chips, though it concerns one model and workload rather than proving that all software can be ported so easily. Anthropic already runs models across Nvidia and AMD GPUs, Google TPUs, and AWS Trainium, managing them as a shared pool of compute. Its early choice to support multiple platforms now looks useful as AI tools simplify hardware-specific engineering. Anthropic began working with AMD GPUs this year and has an agreement covering up to two gigawatts of capacity. If the reported speedup generalizes, choosing accelerators could depend more on price and availability and less on the cost of adapting software.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.