Article
Claude 3 days to break through NVIDIA's 20-year CUDA moat? Anthropic has Mythos port its own GPU low-level code
Beating News,AI— Anthropic disclosed during its IPO roadshow that Claude Mythos Preview took just 3 days to get a key computational stage of a Claude model running on AMD GPU. Claude ported or wrote almost all of the required GPU kernels itself, with little human intervention. Anthropic said that adapting to a new chip platform previously required dozens of engineers working for 6 to 12 months.
One of NVIDIA's hardest-to-replace advantages is CUDA。
CUDA, NVIDIA's 2006 computing platform introduced in GPU. By 2018, mainstream deep learning frameworks such as PyTorch and TensorFlow were already widely using CUDA and its associated computing libraries. NVIDIA also progressively expanded its tools for multi-card communication, debugging, and performance optimization. Developers can efficiently use GPU。
without writing extensive low-level code from scratch. As more AI training and inference programs were developed around CUDA, NVIDIA accumulated a large body of ready-made code and tools. Developers often choose NVIDIA because the software has already been adapted. Even if AMD produces a highly capable GPU, customers looking to switch still face code migration, compatibility testing, and performance optimization.
AMD's MI355X is equipped with 288GB of high-bandwidth memory, making it a capable hardware configuration. The difficulty is that GPU kernels written for NVIDIA may not run directly when switched to AMD, and even if they do, performance may not be optimal. Engineers must also adjust the code for AMD's chip architecture and repeatedly test computational results and execution speed.
AMD has long offered tools such as ROCm and HIPIFY to help developers migrate CUDA programs. Cross-platform programming tools such as Triton are also making migration easier, but complex low-level computations still require extensive manual optimization.
Anthropic had Mythos take over some of that work, porting or rewriting GPU kernels itself. According to the roadshow chart, it achieved usable performance for the prompt-processing stage of a Claude model on 3 within AMD MI355X days.
Anthropic has in fact been reducing its dependence on any single chip vendor for years. Its prospectus disclosed that the company spent years building a compute scheduling system spanning NVIDIA GPU、AMD GPU、Google TPU and Amazon Trainium, allowing it to select different chips based on price and performance.
In July this year, Anthropic also reached an agreement with AMD to begin deploying 2027's next-generation AMD in the first half of GPU, at a scale of up to 2GW. The two parties will also use Claude to accelerate ROCm development and optimize GPU performance.AMD committed to invest up to 5000000000 USD。