NVIDIA Blackwell Tops MLPerf Training 6.0 with Industry-Leading Scale and Performance | NVIDIA Technical Blog - NVIDIA Developer

NVIDIA achieved a dominant performance in MLPerf Training v6.0, the latest industry-standard AI training benchmark developed by MLCommons. The company's GB300 NVL72 platform delivered the fastest time... Read original →

NVIDIA’s MLPerf v6.0 dominance isn’t just a chip win—it ignites a full-stack cascade: EDA and 3nm EUV foundry demand surges upstream, while cloud providers scramble to deploy Spectrum-X and Quantum InfiniBand to handle its communication density. U.S. export controls on advanced AI chips raise NVIDIA’s global delivery costs by 15–20%, straining co-optimization with foundries in Taiwan, China, and South Korea. Competitors like AMD and Intel will likely pivot from raw performance races toward open ecosystems (e.g., ROCm + MI300X clusters) or IP licensing. Chinese GPU firms are fast-tracking MoE-native architectures to bypass CUDA lock-in. Within 18 months, Blackwell’s '10k-GPU-scale' training will become the de facto benchmark for foundation models—but soaring power and interconnect bottlenecks will accelerate adoption of optical I/O and near-memory computing.