"Combined with the software optimizations available in the NVIDIA NIM microservice, a single server with eight H200 GPUs connected using NVLink and NVLink Switch can run the full, 671-billion-parameter DeepSeek-R1 model at up to 3,872 tokens per second".