SemiAnalysis posted on September 13 that AMD released its DeepSeek V4.1 Flash image a full two days after vLLM on the CUDA side had already added support for the model. The image itself works out of the box, with no missing features. The gap shows up in cost efficiency: by the firm's own measurements, the same model running on AMD hardware delivers up to 14.8x worse performance per dollar than an H200, and up to 42x worse than a B200 or B300.
SemiAnalysis attributes the gap to the ecosystem. Nvidia's collaboration with its developer community lets the CUDA path get optimized on day one of a model's release, and the firm puts the size of that developer base at six million.
Where the two-day gap comes from
In the first few days after a new model ships, inference-side work is concentrated on kernel adaptation, quantization path selection, and scheduler tuning. The core development effort behind inference frameworks like vLLM and SGLang has long revolved around CUDA, and the reference implementations that model makers build themselves also tend to run on CUDA first. AMD's side has to wait for that mainline work to land before doing its own round of porting and validation, and a two-day lag is the routine result of that pipeline.
DeepSeek V4.1 Flash is a model that entered private testing in early September before its formal release, and its architecture differs from earlier generations. The bigger the architectural change, the higher the porting cost, and the wider the day-one performance gap gets stretched.
The same order of magnitude showed up three weeks earlier
Rewind three weeks: on August 24, SemiAnalysis published a benchmark suite called AgentX 1.0, built specifically to test long-context agentic inference. That suite used 393 anonymized Claude Code call traces with contexts exceeding one million tokens, ran on more than 1,000 GPUs drawing roughly 2 megawatts of power, and carried a budget the firm put at over $3 million.
That earlier test's conclusions varied by model: on SGLang, Qwen3.5 saw Nvidia outperform AMD by more than 20x at the 90 tok/s/user tier; GLM 5.3 saw Nvidia's cost efficiency come in roughly 5x higher at the 150 tok/s/user tier; and B300's FP4 path delivered a 12x gain in performance per dollar over an H100. The same material included an even blunter line: across all of Alibaba, AMD's ATOM stack is used by only one small advertising business unit.
The 5x-to-20x gap from three weeks ago and the 14.8x-to-42x gap from this week land in the same order of magnitude. The difference is that AgentX tested models that had already been running stably for a while, while this latest test covers a model in its first two days after release, where the cold-start gap widens further.
The hardware itself isn't the bottleneck
The MI355X's spec sheet isn't behind on paper. SemiAnalysis made a pointed claim in that August report: at certain tiers, even giving away AMD's chips for free wouldn't win on a per-token cost basis, until vLLM and SGLang catch up on the software side. After August 21, a round of optimizations on vLLM contributed by Inferact and Nvidia flipped the script, with the B200 overtaking the MI355X on performance per dollar.
For buyers, the takeaway from these numbers is that the deciding variable in procurement has shifted from peak compute to software-stack maturity and how quickly a platform adapts to new models. For an inference service provider that wants to go live the same day a new model ships, AMD's two-day lag translates directly into two days of lost revenue.
Where the numbers need caveats
Both the 14.8x and 42x figures come solely from SemiAnalysis; AMD has not responded, and no third party has reproduced the results. Performance-per-dollar is an extremely price-sensitive metric, and the market-price gap between a used H200 and a brand-new B300 alone can swing the multiple in either direction. The firm's earlier benchmark reports have disclosed their testing methodology and cost model in public, but this latest post did not come with a full methodology writeup.
AMD's software stack has been closing the gap for the past couple of years. Whether that pace can outrun Nvidia's per-generation optimization investment will come down to how many days separate the two sides' images the next time a model ships with a major architectural overhaul.
Sources: SemiAnalysis, CocoLoop, public vLLM project records; the performance-per-dollar multiples and the AgentX benchmark's compute scale are recorded as stated by SemiAnalysis and have not been independently verified.