Before using Claude to port AMD kernels, stop treating three days as a deadline

Before using Claude to port AMD kernels, stop treating three days as a deadline

Before asking Claude to move a workload onto AMD hardware, check whether the model in the demonstration is available to you. Anthropic's roadshow describes Mythos getting Claude running on AMD in three days, but CellCog's account of the presentation says Mythos is not a public model. The internal result covers prompt processing for one model, not a delivery estimate for migrating an entire service.

The result has a narrower acceptance test than it sounds

Anthropic's investor website dates the start of its roadshow to November 10, Eastern time. In transcript excerpts published on November 11, Simon Willison records co-founder Tom Brown saying that Mythos needed almost no human intervention and handled nearly all the required kernel writing or porting. That is an account of work inside Anthropic, not an independently reproduced experiment.

CellCog's transcription of the slide footnote supplies the boundary: Claude Mythos Preview brought up prompt processing for one Anthropic model on AMD MI355X GPUs in approximately three days, reaching a usable level of performance. The presentation supplies no specific speed, cost or performance comparison with another chip, nor acceptance results for training, answer generation or production operations.

Define a small test before changing hardware

There is no migration-service quote to adopt here, and elapsed days do not tell you what your model usage will cost. Anthropic's registration filing describes years of investment and a proprietary orchestration layer behind its multi-chip capabilities. Our reading is that a model may shorten the coding stage without making that existing infrastructure unnecessary.

An engineering team can separately evaluate a coding tool it can actually access, without assuming it has the capabilities attributed to Mythos. Choose a computation with known reference outputs, fix the inputs and acceptance criteria, and ask the tool to propose a port. Check correctness before measuring the complete task's runtime. Include human repairs, testing and deployment work in the record.

Keep the original implementation as a reference, run both versions against the same test inputs, and record the hardware and software environments. Review output differences, full-task runtime and human effort together before expanding the scope of the port.

Share this article