Artificial Analysis has integrated the AMD Instinct MI355X into its AA-AgentPerf benchmark, with data published on October 14, 2026. The accelerator is based on the CDNA 4 architecture, featuring 288 GB of HBM3E memory with 8 TB/s bandwidth and a power consumption of up to 1,400 W.
Testing was conducted on a single 8-GPU node using ROCm nightly builds and developmental AVM technologies. In the DeepSeek V4.1 Flash model, the MI355X showed competitive output speeds but trailed Nvidia’s Blackwell chips in efficiency, delivering 128 million tokens per second per megawatt compared to 193 million for the B200.
However, the MI355X outperformed the Nvidia B200 by 15–40% in throughput per megawatt under mid-range loads when running the larger DeepSeek V4 Pro 0813 model. Analysts attribute this advantage primarily to the higher memory capacity of the AMD hardware.
The AA-AgentPerf benchmark, which launched in June 2026, focuses on multi-step agentic workloads rather than simple prompts. Research indicates that AMD’s performance has improved significantly through software optimization alone, as the hardware remained unchanged during the testing period.
