Abstract:
Anthropic disclosed during its IPO roadshow that it uses Claude Mythos Preview to help adapt AMD AI accelerators. A company said that Claude AI took only 3 days for the internal model to complete prompt word processing on AMD MI355X and reach usable performance levels.
Claude ported or wrote nearly all of the GPU kernel code in this effort with minimal human intervention. This case has triggered discussions about NVIDIA's CUDA software ecosystem and competition with AI chips. CUDA is NVIDIA's moat, but it remains to be seen how long this moat can last in the AI era.
When AI models run on different chip architectures, they need to adapt to the underlying computing code, software stack, and hardware features. To achieve practical performance, engineers often also optimize for matrix operations, memory access, and multi-chip communications.
A company said that based on the company’s past experience, using traditional engineering methods to integrate new chip platforms usually takes dozens of engineers several months (actually 6 to 12 months), but with the participation of AI, this process can be shortened to just 3 days.
Nvidia’s CUDA, which brings together a large number of development tools, computing libraries and GPU-optimized software, has long been an important foundation for AI development, because adapting code originally optimized for specific hardware to other chips requires the investment of a large amount of engineering resources.
After AI participates in the transplantation of the underlying code, it can help engineers write and adjust GPU kernels faster, and the cost and time for enterprises to access different chip platforms will have the opportunity to decrease. At this stage, the chips actually adapted by Company A include but are not limited to NVIDIA, AMD, Google TPU and Amazon Trainium.
AMD and Anthropic have previously announced that they will deploy up to 2GW of GPU computing power based on the AMD MI450 series. The first 1GW deployment is planned to begin in the first half of 2027. Both parties will also use Claude to optimize AMD GPU workloads and accelerate ROCm software development.
Comments