AMD's MI355X Beats Nvidia's B200 on Tokens Per Megawatt for DeepSeek V4 Pro, but Trails on the Lighter Model

Two independent benchmark analyses published this week point to the same conclusion: AMD's software stack is closing the inference gap with Nvidia faster than most buyers expected, though the results are split by model and workload. Artificial Analysis added the AMD Instinct MI355X to its AA-AgentPerf benchmark on October 14, and SemiAnalysis researchers used the same day's podcast episode to argue that AMD has caught up to, and briefly passed, some Nvidia configurations on inference efficiency.

The hardware under test is substantial. The MI355X is built on AMD's CDNA 4 architecture, carries 288GB of HBM3E memory running at 8TB/s per GPU, and can draw up to 1,400 watts. Artificial Analysis ran its tests on a single eight-GPU node using ROCm nightly builds, the freshest and least settled version of AMD's software stack, alongside developmental releases of its serving technologies.

On DeepSeek V4.1 Flash, a mixture-of-experts model with 552 billion total parameters and roughly 8 billion active, the MI355X produced output speeds that rival Nvidia's Blackwell systems, slipping slightly behind at higher concurrency. Efficiency is where the gap shows: at 100 tokens per second per user, AMD's chip delivered approximately 128 million tokens per second per megawatt, against roughly 193 million for the B200 and 220 million for the B300.

The picture flips on the larger DeepSeek V4 Pro 0813 checkpoint. There the MI355X beat the B200 in throughput per megawatt by 15-40% under mid-range loads, an advantage Artificial Analysis attributes largely to the AMD card's higher memory capacity per GPU. Which chip wins, in other words, depends on which model is being served and how many users are hitting it at once.

SemiAnalysis reached a similar verdict from a different benchmark. Its InferenceX platform tracks how competing GPUs run GLM 5.3, a Chinese frontier model used for coding and complex tasks. Researcher Wenyao Gao noted that hardware specs are fixed at launch but the software controlling the chip keeps improving, and AMD's ATOM inference engine improved so quickly over recent weeks that the MI355X overtook some Nvidia configurations before Nvidia clawed the lead back. Podcast host Jordan Nanos conceded that long-standing criticism of AMD's software ecosystem, including from SemiAnalysis itself, now looks outdated.

The supporting numbers back that up. In a September study, SemiAnalysis estimated that when both systems must begin responding within two seconds, the MI355X's inference cost was about 9% lower than the B200's, though Nvidia's GB200 and GB300 configurations remained cheaper in other setups. AMD separately reported in September that its MI355X handled up to 38% more AI workload than in the previous MLPerf round, and the company has shortened its ROCm release cadence to roughly six weeks.

The caveats matter as much as the headlines. The AA-AgentPerf numbers came from nightly software builds on developmental serving tools, and none of these figures are audited by a neutral party; both outlets publish their own methodology and invite operators to reproduce them. Still, for data-center buyers the practical takeaway is that inference economics are now decided as much by software release cadence as by silicon, and that the same accelerator can be the cheaper or more expensive choice depending entirely on the model and concurrency it is asked to serve.