AMD Instinct MI355X Trailed Nvidia Systems in Efficiency
Benchmark results indicate the MI355X delivered lower tokens per megawatt than competing Blackwell architectures.
Updated on Oct. 14, 2026 in Data Centers
Live Poll
Do you believe performance efficiency matters more than hardware cost when selecting new AI server systems?
The AMD Instinct MI355X configuration recorded a lower energy efficiency rating than Nvidia B200 and B300 systems in recent agentic AI testing. Added to the AA-AgentPerf benchmark on October 7, the setup measured throughput relative to power consumption during model inference.
Why it matters
This efficiency gap provides data center operators with comparative performance metrics across specific hardware and software stacks. These results highlight the ongoing competitive trade-offs in power utilization for large-scale agentic AI workloads.
The AMD MI355X configuration achieved 128 million tokens per second per megawatt, compared to 193 million for the B200 and 220 million for the B300. Testing utilized the 552 billion parameter DeepSeek V4.1 Flash model at a target of 100 tokens per second per user.
The players
AMD
A semiconductor company focused on high-performance computing, data center GPUs, and adaptive processors.
Nvidia
A leading designer of graphics processing units and AI-accelerator hardware used in high-density data centers.
The details
The benchmark uses AA-AgentPerf, an assessment that replays agentic coding sessions by incorporating tool-call delays and prompt caching. The MI355X configuration ran on a node with eight GPUs, though the serving architecture utilized two-GPU tensor parallelism, a technique where a single neural network layer is split across multiple processors. The setup relied on MXFP4, a low-precision data format, alongside a nightly vLLM build on ROCm 10 and DSpark speculative decoding, a method that uses a smaller model to predict tokens before verification.
Timeline
October 7, 2026: The AMD MI355X configuration was added to the benchmark results.
The Tech Race
This development places the AMD Instinct MI355X directly into the competitive landscape defined by the AA-AgentPerf benchmark. It updates the performance hierarchy previously dominated by Nvidia's Blackwell series, reflecting the ongoing race to optimize inference throughput.
Data center operators and enterprise AI architects can use these benchmark figures to inform hardware procurement decisions based on energy-per-token efficiency. As these software configurations and nightly builds improve, the performance delta between these architectures may shift.
The takeaway
The benchmark underscores that architectural efficiency is sensitive to software-stack optimization, including parallelism and decoding strategies. Operators should continue to monitor updated test results as vLLM builds and ROCm driver improvements are integrated into production environments.
Further reading
For more on evolving infrastructure requirements, explore the Data Centers section.
Live Poll
Do you believe performance efficiency matters more than hardware cost when selecting new AI server systems?
