Achieving Top Inference Performance with the NVIDIA H100 Tensor Core GPU and NVIDIA TensorRT-LLM

NVIDIA developer blog comparing DGX H100 (8x H100 80GB HBM3) with TensorRT-LLM v0.5.0 (batch-1 point) and v0.6.1 (latency-threshold points) against AMD's MI300X launch claims for Llama 2 70B. Figure 1 is an SVG chart; the only figures stated in prose are 1.7 s per inference at batch 1 and >5 inferences/s at a 2.5 s fixed response-time budget. Full build.py and gptSessionBenchmark command lines are published. | AUDIT 2026-10-03 RE-VERIFIED: the URL is live and returns HTTP 200; an earlier probe recorded 000, which was a transient connection failure from the concurrency burst, not a dead link. No change made to the URL.