<h1>NVIDIA Blackwell Revolutionizes Generative AI in MLPerf Inference Debut</h1>
NVIDIA’s Blackwell platform has set a new benchmark for generative AI with its first appearance in the MLPerf Inference v4.1 benchmarks. The Blackwell series has demonstrated performance improvements of up to 4x compared to its predecessor, the H100 Tensor Core GPU. Notably, the H200 GPU also showcased substantial gains across various benchmarks, excelling in tests including the Mixtral 8x7B MoE (Mixture of Experts) LLM (Large Language Model).
These benchmarks underscore NVIDIA’s relentless innovation in AI technology, supported by advances in their software stack, such as the adoption of FP4 precision and TensorRT-LLM, which have driven significant performance enhancements. From data centers to edge deployments, NVIDIA continues to lead with platforms like the Triton Inference Server and Jetson, the latter of which demonstrated major gains for edge AI tasks with the Jetson AGX Orin.
The latest MLPerf Inference results introduce a new generative AI benchmark, providing businesses with standardized tests to guide AI infrastructure investments. While new entries like AMD’s MI300x and Google’s TPUv6e aim to compete, Nvidia’s dominance remains evident. Notably, the performance of MoE models was evaluated for the first time, showcasing their potential for efficient and specialized deployment.
Despite Nvidia’s strong presence, competitors are making strides. Companies such as AMD Instinct and Untether AI are closing the gap in power efficiency and specific AI tasks. Google’s Trillium has shown moderate performance, whereas Untether AI’s at-memory computing approach shows promise in edge applications. Additionally, companies like Cerebras and Furiosa are developing new, efficient AI inference chips, although they have not yet participated in MLPerf.
In summary, NVIDIA’s Blackwell platform has established itself as a new standard in AI inference, delivering remarkable improvements and setting the stage for the future of AI technology.

Collections

NVIDIA's Blackwell platform has set a new benchmark in generative AI with significant performance improvements over its predecessor. The latest MLPerf Inference v4.1 results highlight NVIDIA's advancements in AI technology, including the adoption of FP4 precision and TensorRT-LLM. Competitors like AMD, Google, and Untether AI are also making strides, but NVIDIA continues to lead, particularly with its diverse applications from data centers to edge deployments.

NVIDIA Blackwell Platform Advances LLM Inference Performance in MLPerf Inference v4.1