NVIDIA's CUDA platform achieved a clean sweep in MLPerf Training v6.0, the industry-standard AI benchmark, delivering...
NVIDIA's CUDA platform achieved a clean sweep in MLPerf Training v6.0, the industry-standard AI benchmark, delivering the fastest time to train at scale and highest per-accelerator performance on every test.
signal brief
NVIDIA's CUDA platform achieved a clean sweep in MLPerf Training v6.0, the industry-standard AI benchmark, delivering the fastest time to train at scale and highest per-accelerator performance on every test. The results, published on June 16, 2026, by MLCommons, showed NVIDIA's platform was the only one to submit results on new DeepSeek-V3 and GPT-OSS-20B MoE benchmarks, using optimized CUDA software stacks and full-iteration CUDA graphs for token-dropless MoEs Source 1. Additionally, the CUDA Toolkit 13.3 release on June 17, 2026, introduced enhanced Tile Programming in C++ and Python, extending GPU programming to more developers Source 2. The Hugging Face Transformers v5.11.0 release on June 10, 2026, integrated new models like DeepSeek-V3.2 and added Triton kernel support, reflecting strong ecosystem adoption Source 3. The MLPerf sweep reinforces CUDA's dominance in GPU software, while the toolkit enhancements and ecosystem integrations widen its moat. Predictions on Manifold Markets show a 56.95% consensus that CUDA will remain a monopoly through 2027 Source 4. The evidence points to sustained leadership and growing developer reliance, making this a high-confidence upward signal.
evidence
- https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/web
- https://developer.nvidia.com/cuda-toolkitweb
- https://github.com/huggingface/transformers/releases/tag/v5.11.0github
- https://manifold.markets/_deleted_/will-cuda-remain-a-monopoly-for-gpuweb
Decision support, not stock advice. This signal is research with cited evidence — not a recommendation to buy, sell, or hold any security.