Independent Analysis Of Leading Gpus Amd Nvidia

This June 2025 benchmark from Artificial Analysis compares eight-GPU servers built on AMD's MI300X and NVIDIA's H100 and H200, serving DeepSeek R1 and Meta's Llama 4 Maverick (artificialanalysis.ai). The key variable is concurrency, the number of requests a server handles at once. More simultaneous requests raise total output, called system throughput, but each individual user gets tokens more slowly. The test steps concurrency up in three-minute phases, each request using 1,000 input and 1,000 output tokens, until throughput stops growing.

Below roughly 250 concurrent queries, the NVIDIA systems were marginally faster per query, with lower latency and slightly higher throughput. At 16 concurrent queries on DeepSeek R1, for example, the H200 delivered about 45 output tokens per second per query against about 35 for the MI300X. Under heavy load the order flipped: the MI300X reached 25-35% higher peak system throughput and kept answering nearly every query up to its peak, while the H200's response rate degraded. Because cost per token is a system's cost divided by its throughput, the authors conclude NVIDIA is slightly cheaper per token at low concurrency and the MI300X cheaper at high concurrency. They found no statistically significant accuracy difference between platforms.

The practical reading is that NVIDIA suits latency-sensitive uses such as chatbots and real-time services, while the MI300X suits large batch workloads. Two details matter when weighing this. AMD engaged Artificial Analysis for the study and provided the test system, though the firm says it followed its own published methodology. And each model was compared against only one NVIDIA chip, the H200 for DeepSeek R1 and the H100 for Llama 4 Maverick, using different software versions and settings on each side.