AMD vs NVIDIA AI GPUs: How Claude Mythos is moving beyond CUDA
The time it took for Anthropic's Claude Mythos Preview to create the software needed for its AI to run on AMD GPUs was only three days, down from six to twelve months. Just imagine the situation where it took six to twelve months – the period during which NVIDIA's software empire had made it impossible for engineers to escape from its clutches, because this empire did not depend only on silicon. The code in those kernels told the GPU what to do to calculate the numbers for a particular model, and hence tied the model exclusively to the hardware of that vendor.
What Anthropic said
Here are some quotations taken from Anthropic IPO roadshow, which began on November 10, when Tom Brown, one of the company's co-founders, explains how Claude is trained and operates using AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. The last one is the newcomer for 2023. Setting up a new chip platform used to require efforts of several dozen people during many months. However, this time, Mythos succeeded in porting Claude onto AMD platform in just three days without almost any human intervention and did most of the kernel development.
An attractive title in such case could be "CUDA is dead". Hold your horses. The note under the slide says that three days relate to the prompt processing of one Anthropic’s model on MI355X GPUs in "useful" performance. One model. One step of inference process. One chip. And "useful" performance doesn't have to mean NVIDIA's performance level either.
Why AMD may be better placed than you think
The past of AMD and Anthropic in this field is known. Via an agreement signed in July, AMD will leverage Claude to optimise the load on their Instinct GPUs and speed up ROCm, the company's software stack. SemiAnalysis had previously argued that since the compilers and most of the kernels of AMD were open-source, AMD was perfect for the agentic era. That makes sense. An agent able to read code is an agent that can also write better code.
A practice run had also happened. Tom Brown had revealed how he used Claude to get Anthropic's inference stack up and running on AMD hardware over the course of just a weekend. Over the course of just a weekend, not even a quarter. Even Mythos shaving half the effort will mean a lot to the buyer.
What it means for buyers
Any company intending to acquire AI computing capabilities – whether large hyperscalers, Indian startups, and data center construction firms acquiring their capacity at NVIDIA prices – can benefit from reduced switching costs. The simple fact that there is a competitive threat can influence the negotiation process even before the competition manages to start stealing customers from the market. It should be remembered that NVIDIA has the superior environment, the larger customer base, and more experience among developers, none of which disappears in 72 hours.
The irony of the situation lies in the following: the software required for AI applications can now be produced by AI agents themselves, but the hardware is needed for them to test the code. AMD appears to have faced chronic GPU shortages within its company for this exact reason, which has been holding it back as the new bottleneck. The software becomes cheaper while the hardware becomes more expensive.