AMD’s MI355X Trails Nvidia B200 and B300 in AI Efficiency Test

AMD’s Instinct MI355X delivered lower throughput per megawatt than the listed Nvidia B200 and B300 systems in an agentic AI benchmark, giving data center operators a comparison across specific hardware and software configurations.

At a target output speed of 100 tokens per second per user, the MI355X result was about 128 million tokens per second per megawatt. Nvidia’s B200 delivered about 193 million, while its B300 delivered about 220 million at the same target speed.

The MI355X configuration, added to the benchmark results on Oct. 7, used MXFP4 precision, a nightly vLLM build on ROCm 10, two-GPU tensor parallelism and DSpark speculative decoding. The node contained eight GPUs, but the serving configuration used two.

AA-AgentPerf replays agentic coding sessions and includes prompt caching and tool-call delays. The benchmark allows optimizations such as key-value cache reuse and speculative decoding, making the tested software configuration part of the comparison. Benchmark results can change as configurations improve.

DeepSeek V4.1 Flash has 552 billion parameters, with 8 billion active per input token and 16 billion per output token. The figures compare performance on this model and workload; they do not describe performance across all models or deployments.