AI Trends

Anthropic says Claude Mythos cuts AMD GPU adaptation cycle to about three days

cnBetaUpdated Nov 17 16:39

AI summary

Anthropic co-founder and chief compute officer Tom Brown said at the IPO roadshow that Claude Mythos can automatically port or write nearly 100% of GPU kernel code, enabling the model to run on AMD Instinct MI355X, shortening adaptation work that originally required dozens of engineers and months to about three days. A footnote in the roadshow materials limits this result to a single model completing the Prefill task on MI355X and reaching a usable performance level; whether the Decode stage is included is not clearly stated. Anthropic did not disclose Claude's actual running speed on MI355X, nor did it provide performance comparison data with similar NVIDIA hardware.

Why read this

SelectedAI score 78

Anthropic disclosed its own engineering practices in its IPO roadshow, which readers can use to understand how AI automatically writing GPU kernels affects the migration costs of the CUDA ecosystem.

Topics

Tags

Body · AI translation

Anthropic co-founder and chief compute officer Tom Brown presented the result during the company's IPO roadshow for investors. He said Claude Mythos can automatically port or write nearly 100% of GPU kernel code, enabling Claude models to run on AMD Instinct MI355X accelerators. Previously, similar hardware adaptation work often required a team of dozens of engineers and took months.

However, the three-day figure Anthropic disclosed has a clear scope of application, and does not mean that the entire model deployment, production-environment validation and performance optimization process can be completed within three days. A footnote in the roadshow materials states that this result specifically refers to an Anthropic model completing prompt processing (Prefill) tasks on AMD MI355X GPUs and reaching a performance level with practical value. Anthropic did not clearly state whether the model's answer-generation phase (Decode) is also included.

GPU kernels are an important link between AI software and the underlying hardware, responsible for translating model computation tasks into instructions that can be executed efficiently on a specific processor. GPU architectures, memory systems and compute units differ among vendors, so the same AI model often requires low-level code to be written or adjusted for different hardware platforms. This kind of work has long been an important source of advantage for Nvidia's CUDA ecosystem.

Nvidia's CUDA platform has been developed for nearly 20 years and has mature compilers, development tools, optimization libraries and a large developer community. AMD has its own ROCm software platform and supports related compute tasks through the HIP programming layer. To help developers migrate CUDA code to the AMD platform, AMD previously launched the HIPify tool, but code conversion does not mean all programs can directly achieve ideal performance; developers usually still need to do a great deal of debugging and optimization.

Advances in AI programming tools are changing this situation. Anush Elangovan, AMD's vice president of AI software, said in April this year that Claude has an advantage over HIPify in code conversion because it has built-in web search capabilities and can obtain more technical materials during the process. AMD has also previously demonstrated research results on using AI agents to automatically write GPU kernels, but early systems still fell short of human experts in some tests.

Anthropic does not rely solely on Nvidia GPUs to run its models. The company already uses different types of accelerators, including Nvidia and AMD GPUs, Google TPU and Amazon AWS Trainium, and manages these hardware resources as a unified, flexibly schedulable compute resource pool. Tom Brown previously called the company's early decision to adopt a multi-chip platform strategy an unusual choice. As software adaptation time is greatly shortened, this strategy may become more attractive.

Anthropic began using AMD GPUs for the first time this year and reached a cooperation agreement with AMD involving up to 2 gigawatts of compute. According to information previously disclosed by both sides, AMD is also using Claude to optimize workloads for the Instinct series GPUs and to accelerate development of the ROCm software platform. AMD believes that AI-generated kernel code is expected to further lower the technical threshold for customers to adopt its accelerators.

In fact, AMD management has previously spoken many times about the impact of AI programming tools on the GPU software ecosystem. AMD CEO Lisa Su said last September that a large amount of AI software was originally developed around CUDA, but a considerable portion of that code can now be converted through software tools into a form suitable for the AMD platform. In July this year, AMD senior vice president Vamsi Boppana also said that AI is expected to eliminate the remaining software obstacles to adopting AMD GPUs.

At AMD's Advancing AI event held in July this year, OpenAI's Philippe Tillet also discussed the role of AI agents in GPU kernel development. He said that after OpenAI received AMD Helios racks, one engineer took only a few days to confirm that the system could run normally, and noted that AI agents are gradually taking on some of the work previously handled by kernel engineers. AMD also revealed that one engineer had Claude help start up an MI355X rack and keep the system running over the weekend.

The data disclosed by Anthropic this time has attracted attention also because the results no longer come only from AMD's own demonstrations or promotions. As an AI company that actually purchases and uses chips, Anthropic disclosed to investors the change in time in its own engineering practice. Its IPO prospectus also mentions that, with the help of Claude, the company has been able to bring new chip platforms into production environments with far less engineering investment than in the past, although the document does not give the specific three-day figure.

However, faster adaptation does not directly prove that AMD GPUs have caught up with Nvidia in performance. Anthropic did not publish the actual running speed of Claude on MI355X, nor did it provide performance comparison data with similar Nvidia hardware. The report also noted that although the chart the company showed marked the adaptation progress of both approaches as close to 100%, it did not explain in detail which work items this percentage corresponds to.

Nvidia CEO Jensen Huang previously said that the company itself also makes extensive use of AI to assist in developing GPU kernels, and believes that no one understands its own chip architecture better than Nvidia, so Nvidia can more fully unleash hardware performance. Regarding the practice of companies such as Anthropic adopting other accelerators, he once called Anthropic a special case rather than an industry trend.

At the same time, AI automatically writing GPU kernels also faces a new resource bottleneck. An analysis published by SemiAnalysis in July this year argued that software development costs are falling, but code generated by AI agents still needs to be tested and validated on actual GPUs. Each agent needs corresponding GPU resources to run tests, and AMD's internal GPU resources for such development are insufficient, which may become a constraint on further expanding applications.

Overall, the significance of Claude Mythos is not only that it shortens a specific software development process, but also that AI may change the cost of software migration between different GPU platforms. CUDA's advantage will not disappear immediately because of this; its mature toolchain, developer ecosystem and long-accumulated performance optimization experience remain valuable. But if AI can increasingly automate code porting, kernel writing and hardware adaptation, customers' dependence on a specific software ecosystem when choosing GPUs may decline, and competitors such as AMD may also have the opportunity to win over customers who previously mainly used Nvidia platforms at a lower migration cost.

Source

cnBeta