DeepSeek-R1’s 671B MoE fits one eight-GPU HGX B300 node at FP8, with room for KV cache. See the fit math and why the node beats a multi-server build.
Tag: Blackwell Ultra
The answer is set by three things: the model’s weight footprint, the KV cache your context and concurrency produce, and how your serving framework spreads both across GPUs. An eight-GPU
The NVIDIA HGX B300 is NVIDIA’s Blackwell Ultra platform in its densest single-node form: eight B300 SXM GPUs wired together on one baseboard by fifth-generation NVLink and NVSwitch into a