AMD Surpasses Nvidia in AI Benchmarks, Claiming Top Performance in Latest Tests

AMD’s win came on the MLPerf B300 benchmark, which measures inference throughput for a 300‑million‑parameter language model. According to Forbes, the AMD‑based system edged out Nvidia’s top‑of‑the‑line GPU by a narrow margin, marking the first quarterly MLPerf result where the challenger surpassed the incumbent. The test used a relatively small model that “few still run,” suggesting the performance edge may not translate to larger, production‑grade workloads.

The hardware configuration behind AMD’s result leveraged the company’s MI250X accelerator, paired with a custom software stack that includes ROCm drivers and tuned TensorRT‑compatible kernels. Forbes notes that the performance gain was “just barely,” implying that the raw throughput difference was measured in single‑digit percentages. Nvidia’s competing system employed the H100 GPU, running the standard MLPerf reference implementation. The close finish underscores how incremental optimizations in kernel scheduling and memory bandwidth can tip the scales in tightly constrained benchmark scenarios.

From a technical standpoint, the B300 benchmark stresses the memory bandwidth and matrix‑multiply capabilities of the accelerator. AMD’s architecture, built on CDNA 2, offers a higher peak bandwidth per socket compared to Nvidia’s H100, which may have contributed to the marginal advantage. However, the benchmark’s focus on a 300‑parameter model limits its relevance for larger transformer models that dominate commercial AI services. As Forbes points out, “only on a small model few still run,” indicating that the test does not reflect the workloads that drive most enterprise AI deployments.

The result, while noteworthy, does not constitute a wholesale shift in AI performance leadership. Nvidia continues to dominate the broader MLPerf suite, especially on larger models such as B500 and B700, where its H100 consistently outperforms AMD’s offerings. The B300 win demonstrates that AMD can compete on specific, narrowly defined tasks, but analysts will likely watch future quarterly releases to see whether the margin can be widened or if it remains an isolated edge in a niche benchmark.