Grafy News
Try Grafy free

Anthropic’s Claude Mythos Cuts AMD GPU Onboarding to Three Days, Challenging Nvidia’s CUDA Dominance

Anthropic says its Claude Mythos model automated almost all kernel code, shrinking AMD GPU onboarding from months to three days.

Illustration for the news story: Anthropic’s Claude Mythos Cuts AMD GPU Onboarding to Three Days, Challenging Nvidia’s CUDA DominanceSource: Hacker News — front page

What did Anthropic announce about Claude Mythos and AMD GPU onboarding?

Anthropic told investors that its Claude Mythos Preview model was able to generate the low‑level code required to run the model on AMD MI355X GPUs in roughly three days, with "basically no human intervention," according to a report on Tom's Hardware. The company claims the system wrote or ported almost the entire set of kernels—close to 100%—that traditionally tie an AI model to a specific accelerator. This rapid turnaround covers the prompt‑processing stage for a single Anthropic model at a performance level deemed useful for production workloads.

How does the three‑day timeline compare to previous onboarding efforts?

Historically, bringing a new accelerator into service has been a labor‑intensive process. Anthropic’s co‑founder and chief compute officer, Tom Brown, explained to investors that earlier efforts required a team of "dozens of engineers" and could take anywhere from six to twelve months. The three‑day figure therefore represents a reduction of up to 99% in the engineering effort and calendar time. Brown emphasized that this speed-up stems from the model’s ability to write the bulk of the required kernels autonomously, rather than relying on manual hand‑tuning.

Why does autonomous kernel generation matter for Nvidia’s CUDA moat?

Nvidia’s CUDA platform has been a strategic barrier for competitors because most AI software is written against its APIs. Translating CUDA code to run on AMD’s ROCm stack typically involves tools like HIPify, which still require significant manual adjustment and testing. According to the article, AMD’s vice‑president of AI software, Anush Elangovan, even remarked that Claude’s capabilities surpass HIPify because the model can incorporate web‑search results directly into the generated code. If an AI system can reliably produce high‑quality kernels for a new architecture, the cost and risk of moving away from CUDA diminish, potentially eroding Nvidia’s long‑standing advantage.

What are the practical implications for developers and researchers choosing hardware?

The reported reduction in onboarding time could turn hardware selection into a more straightforward purchasing decision rather than a multi‑month engineering project. Anthropic already runs workloads on a heterogeneous fleet that includes Nvidia GPUs, AMD GPUs, Google TPUs, and AWS Trainium, treating the resources as a "fungible pool of compute." For developers, this means that a model like Claude Mythos can be deployed on AMD hardware with far less custom engineering, opening the door to cost‑effective alternatives to Nvidia‑only solutions. Moreover, AMD’s senior vice president of AI, Vamsi Boppana, noted that in some cases the AI‑generated kernels performed better than manually tuned versions, suggesting potential performance gains as well as cost savings.

How can Grafy users explore or apply this development?

Grafy’s Code studio lets users experiment with AI‑assisted code generation across different programming models. By loading a sample CUDA kernel into the Grafy Code environment, creators can prompt the built‑in multi‑model chat (see the Chat app) to translate it to ROCm‑compatible code, effectively mimicking the autonomous kernel generation that Claude Mythos claims to perform. Users interested in the research side can upload relevant PDFs—such as Anthropic’s IPO filing or AMD’s ROCm documentation—into the Reader app to ask targeted questions about kernel optimization or cross‑platform deployment. These workflows give developers a hands‑on way to assess how far AI‑generated code has progressed and whether it could replace parts of their own onboarding pipelines.


Source: Tom's Hardware


Source: Hacker News — front page — Claude Mythos slashed software work to onboard AMD GPUs to around three days