AMD Instinct MI355X Trailed Nvidia in AI Benchmark

Benchmark results indicate the MI355X produced lower token throughput per megawatt than competing Nvidia hardware.

Updated on Oct. 14, 2026 in Data Centers

AMD Instinct MI355X Trailed Nvidia in AI Benchmark

Live Poll

Do you believe performance efficiency matters more than hardware cost when selecting new AI server systems?

The AMD Instinct MI355X reached 128 million tokens per second per megawatt during testing on the agentic AI benchmark AA-AgentPerf. This performance level lagged behind the throughput figures reported for Nvidia B200 and B300 systems.

Why it matters

Data center operators utilize these benchmark results to evaluate the efficiency of specific hardware and software configurations for artificial intelligence workloads. Understanding these power-to-throughput ratios helps facilities optimize their operational costs.

The MI355X configuration utilized MXFP4 precision, a nightly vLLM build on ROCm 10, two-GPU tensor parallelism, and DSpark speculative decoding. The test model, DeepSeek V4.1 Flash, features 552 billion parameters with 8 billion active per input token.

The players

AMD

AMD is a global semiconductor company that designs high-performance computing and graphics processing units for data centers.

Nvidia

Nvidia is a technology corporation specializing in the design and manufacture of graphics processing units used extensively for AI training.

The details

The benchmark test used a node containing eight GPUs, though the specific serving configuration applied only two. AA-AgentPerf recreates real-world agentic coding sessions by incorporating variables such as prompt caching and tool-call delays to simulate realistic usage.

The AMD Instinct MI355X benchmark results were officially added to the database on Oct. 7, 2026.

The Tech Race

The development and refinement of benchmarks like the AA-AgentPerf represent an intensifying arms race to define power-efficiency standards for the next generation of AI infrastructure. This competition pushes hardware manufacturers to prioritize token throughput optimization as large-scale model requirements continue to evolve.

For businesses and data center managers, these efficiency metrics directly influence the operational costs and electricity demand of AI-driven services. Lower throughput-per-megawatt ratios can translate into higher energy expenditures when scaling artificial intelligence applications to meet user demand.

The takeaway

Benchmark performance is dynamic and frequently updated through software optimization and configuration adjustments. Operators should track these figures over time as manufacturers refine their implementations to maximize output efficiency.