The tested configuration combines prefill/decode disaggregation with Wide-EP expert parallelism.
On the SemiAnalysis AgentX workload, vLLM reports 954,000 tokens per second per GPU and a P90 rate of 196 tokens per second per user. That is 7.9 times the team's earlier single-node result. It estimates the configuration at $38.96 per GPU-hour using DeepSeek off-peak pricing. A specialized small-batch decode path reached 826 tokens per second for one user.
Against DeepInfer, DeepSeek's offline engine, vLLM says its GB200 decode throughput was within 6% across batch sizes from 8 to 256 under DeepSeek's 131K-context configuration. Its prefill curve was compared with DeepInfer's published chart, but the hardware and setup differ, so the post does not establish an engine-to-engine prefill speed advantage.
The source reports vendor-run benchmarks and implementation details including DEP2 prefill, EPLB balancing, CPU-overhead reductions, and kernels released by DeepSeek. It does not establish equivalent performance on other models, workloads, hardware, or production deployments.