B200
NVIDIA's Blackwell-generation datacentre GPU.
Resumen
The B200 is NVIDIA's Blackwell-generation accelerator, built with a higher-bandwidth interconnect and lower-precision formats aimed at very large training runs and inference at scale. It solves the problem that training a frontier model is now limited by how fast many GPUs can exchange gradients, not only by each chip's own throughput. It exists because the unit of scale has moved from one GPU to a cluster, and the interconnect became the thing worth optimising.
Grafo de tecnologías
Cómo se conecta B200 con el resto del stack.