Nvidia’s Blackwell GPUs Crush MLPerf Benchmarks

Nvidia’s Blackwell architecture has once again demonstrated its superiority in MLCommons’ latest MLPerf benchmarks, particularly in generative AI workloads. According to the official MLCommons results published this week, Blackwell-based systems outperformed all competitors in both the Llama 3.1 405B and interactive Llama 2 70B tests.

The benchmarks, which measure inference performance – how quickly AI models can process requests after training – are increasingly important as enterprises shift focus from model development to deployment. “Inference is where the rubber meets the road for business applications,” said MLCommons executive director David Kanter in the organization’s press release.

The 4-bit Precision Advantage

Nvidia’s standout performance came in the 4-bit precision category, where Blackwell chips demonstrated a 2.3x improvement over AMD’s recently launched MI325X accelerators. This efficiency gain is particularly significant for cost-conscious data center operators, as lower precision requires less memory bandwidth and power.

“What we’re seeing with Blackwell is that Nvidia isn’t just competing on raw performance anymore,” noted Tirias Research analyst Kevin Krewell in a blog post analyzing the results. “They’re winning the efficiency battle that matters most for scaled deployments.”

Notable Absences and Alternatives

While Google participated with its 5th-generation Trillium TPUs, showing respectable results in some categories, Intel was conspicuously absent from the competition. Industry observers suggest Intel may be focusing its efforts on next-generation hardware rather than benchmarking current offerings.

The MLPerf results come as data center operators face increasingly complex hardware decisions. With Nvidia commanding premium prices, some are looking to alternatives like AMD or developing custom silicon. However, as this benchmark shows, Nvidia continues to set the performance bar that others must clear.