Manual · GPU

B200: Blackwell

Two dies, one GPU. FP4, NVLink 5, and the architecture behind the 4000-GPU clusters that changed what "serving" means.

B200 is the GPU I've spent the most time with, and the one that broke the mold. It's not a bigger Hopper. It's a different kind of object: two dies, connected by a high-speed bridge, acting as one GPU. That one fact explains everything about it.

What changed from H100

The numbers

Why the 4000-GPU cluster works

At 4000 B200s, the architecture stops being about one GPU and becomes about the fabric. NVLink 5 connects 72 GPUs in a rack (NVL72) into what is effectively one giant GPU, and NVSwitch 4 extends that to 576-GPU domains. At 4000 GPUs, that's ~7 domains, with only inter-domain traffic going over InfiniBand or Ethernet with RDMA. The KV cache pool, the disaggregation, the routing, all of it is built on these two facts: the GPU is fast, and the fabric is faster.

For inference specifically, B200's FP4 is the unlock. A 405B model that needs 8 H100s in FP8 can run on 4 B200s in FP4 with higher throughput. A 671B MoE model like DeepSeek V4 fits in FP4 at ~336GB, which is under the 384GB of a dual-B200 node. That means you can serve a frontier model on two GPUs, not a rack.

Sources

Back to the manual