AMD MI355X joins AA-AgentPerf with DeepSeek V4.1 Flash results

AMD finally has a seat at the table in one of AI’s more practical benchmarks. Artificial Analysis has added the AMD Instinct MI355X to AA-AgentPerf, with results reported on October 14, 2026.

The headline matchup is DeepSeek V4.1 Flash, where the MI355X still trails Nvidia’s Blackwell chips on efficiency. On the larger DeepSeek V4 Pro 0813, though, AMD’s chip flips the script against the B200.

What the benchmark actually measured

The MI355X is built on AMD’s CDNA 4 architecture. Each GPU carries 288 GB of HBM3E memory running at 8 TB/s, and the chip can draw up to 1,400 W.

Artificial Analysis tested the accelerator on a single 8-GPU node. The setup ran ROCm nightly builds, AMD’s software stack in its freshest and least settled form, alongside developmental versions of AVM technologies.

The model in focus, DeepSeek V4.1 Flash, uses a Mixture of Experts design with 552 billion parameters. Only about 8 billion active input parameters are engaged at a time.

The numbers: close on speed, behind on efficiency

According to Artificial Analysis, the MI355X posted output speeds on DeepSeek V4.1 Flash that rival Nvidia’s Blackwell systems. It slipped slightly behind at higher concurrency, meaning when many users hit the system at once.

Efficiency is where the gap shows. The MI355X delivered approximately 128 million tokens per second per megawatt at a rate of 100 tokens per second per user.

Nvidia’s B200 reached around 193 million on the same measure. The B300 hit around 220 million.

Then came DeepSeek V4 Pro 0813, the bigger model. Here the MI355X beat the B200 in throughput per megawatt by 15-40% under mid-range loads.

Artificial Analysis linked that advantage largely to the MI355X’s higher memory capacity per GPU.

Why AA-AgentPerf matters

AA-AgentPerf launched in June 2026. At the time, its results showed Nvidia’s Blackwell lineup in the lead.

The benchmark aims to reflect real agentic workloads, the kind where AI systems take multi-step actions rather than answering a single prompt. Artificial Analysis designed it around realistic production conditions instead of simplified lab tests.

Since launch, AMD’s results have improved substantially through continued software optimization, according to the research findings. The hardware did not change in those months. The code running on it did.

What this means for the AI hardware race

For data center operators, the split result is useful information. Which chip wins depends heavily on which model you run and how many users you serve at once.

The software story deserves close watching. These results came from ROCm nightly builds and developmental tooling. The research findings say further developing technologies are expected to enhance the MI355X’s capabilities.

Nvidia still holds the efficiency crown on the lighter workload. AMD, for its part, has shown that on the right model at the right load, its memory-heavy design can beat the B200 outright.