Source profile
developer.nvidia.com
A field appears only when the current profile revision carries supporting evidence. Missing context stays visibly unknown.
NVIDIA's case for next-item generation in recommenders reads as an architecture story. The number that decides whether it ships is terabytes to petabytes of daily user history against GPU HBM.
Reality
NVIDIA says Alibaba's largest open-weight model serves over 4K tokens/sec/GPU and 350 tokens/sec/user in FP8 on a GB300 NVL72. That figure is the self-hosting floor, not a benchmark.
Reality
A vendor walkthrough has a coding agent build an endoscopic segmentation app from HoloHub examples, with the CLI as the shared execution surface and named skills as the spec.
Reality
SkillEvaluator scores each skill with and without installation and calls the gap Skill Lift. The more interesting figure is the control: 39 to 46 out of 100.
Reality
NVIDIA's federated learning SDK treats large vision-language updates as a transport problem: externalize big objects, stream tensors, and aggregate against disk instead of server memory.
Reality
NVIDIA's tutorial post-trains a 4B model into a Franka manipulation policy that runs on Jetson Thor with about 0.6 seconds of latency slack per cycle. Closed-loop success is 22.9%.
Reality
NVIDIA's cuML and cuVS 25.06 distribute all-neighbors kNN construction across GPUs. The vendor claims hundreds of gigabytes in minutes, but the material supplied carries no benchmark table.
Reality
A 66 GB checkpoint becomes 22 GB and, NVIDIA says, up to 4x faster throughput, but only after distillation teaches the quantized student to live with its own noise.
Reality