NVIDIA GPUs drive AI scaling laws and physical AI

NVIDIA highlights that its accelerated GPU computing platform now dominates leading supercomputing benchmarks and underpins three AI scaling laws — pretraining, post‑training and test‑time compute — enabling generative, agentic and physical AI.

• GPU leadership & efficiency: Over 85% of TOP100 supercomputers now use GPUs; NVIDIA GPUs top the Green500 with 70.1 gigaflops per watt vs 15.5 for CPU-only systems, and an 8,192 H100 GPU system achieved 410 trillion traversed edges per second on Graph500, more than doubling the next best (≈150,000 CPUs), while Snowflake integrated NVIDIA A10 GPUs and CUDA‑X (cuML, cuDF) to deliver up to 200x speedups on some ML workloads.
• AI scaling laws & physical AI stack: NVIDIA positions GPUs as essential across pretraining, post‑training and test‑time scaling, supporting 1.4 million open‑source models, powering hyperscaler recommender and search systems, and enabling physical AI via a three‑system stack (DGX GB300 for training, RTX PRO with Omniverse for simulation, Jetson Thor for real‑time inference) alongside initiatives like Project GR00T for humanoid robots showcased at GTC DC 2025.