Claude Mythos adapted AI model for AMD GPUs in three days
In a presentation for investors, Anthropic stated that Claude Mythos Preview adapted another AI model for AMD Instinct MI355X in approximately three days instead of the usual 6–12 months.
In materials for investors, the company provided an example of porting an AI model to AMD accelerators. According to a slide, Claude Mythos Preview managed to launch another model on AMD Instinct MI355X in about three days. Typically, preparing a new platform for inference—processing requests with an already trained model—takes from 6 to 12 months.
This scenario could reduce developers' dependence on CUDA, NVIDIA's software platform. When switching to a different architecture, engineers usually have to adapt low-level operations, verify the correctness of calculations, and manually tune performance. Anthropic views this result as a sign that modern AI models are capable of independently porting neural networks between computing platforms.
The AMD Instinct MI355X is built on the CDNA 4 architecture and is equipped with 288 GB of HBM3E memory. In terms of hardware specifications, the accelerator competes with NVIDIA solutions, though its widespread adoption has long been hindered by software compatibility issues.
However, experts consider conclusions about the collapse of CUDA's advantage premature. NVIDIA's ecosystem includes libraries, compilers, optimized compute kernels, and tools for distributed computing. A key role is played by the NCCL library, which is tied to NVIDIA hardware, including the NVLink interconnect and InfiniBand networking technology. Therefore, switching to another platform may require restructuring the entire data center infrastructure, not just porting a single model.
