Evaluating Llama 3.3 70B Inference on NVIDIA H100 and A100 GPUs (genai-perf)

Third-party NIM genai-perf study: llama-3.3-70b-instruct bf16, TP=4 PP=1, on 4xH100 80GB SXM5 vs 4xA100 80GB PCIe, container nvcr.io/nim/meta/llama-3.3-70b-instruct:1.8.2, genai-perf concurrency sweep 1..500, workloads 200->200 and 1000->200 synthetic. Throughput figures are read off published chart PNGs and are stated as approximations ('~7,000 TPS', '~2,600 TPS at 250 users'), so confidence is kept low. Raw data and scripts published in a public GitHub repo.