Artificial Analysis added the AMD Instinct MI355X accelerator to its AA-AgentPerf benchmark, with results published on October 14, 2026. On the DeepSeek V4.1 Flash model, the MI355X matches Nvidia Blackwell systems on speed but trails on efficiency, posting approximately 128 million tokens per second per megawatt versus 193 million for the B200 and 220 million for the B300. However, on the larger DeepSeek V4 Pro 0813 model, the MI355X outperforms the B200 by 15 to 40 percent in throughput per megawatt under mid-range loads, thanks to its superior memory capacity of 288 GB of HBM3E per GPU. This chip, built on the CDNA 4 architecture with a maximum power draw of 1,400 W, was tested on an 8-GPU node running ROCm. These results demonstrate that the choice between AMD and Nvidia heavily depends on which model is run and how many simultaneous users are served.
Source: Read the original article
