MLPerf Version Comparison

LLM inference · Offline · llama2-70b · on-demand floor pricing

MLPerf is the industry standard benchmark for AI compute, measuring how many tokens per second a GPU actually delivers under controlled conditions. When MLCommons publishes a new version, it resets the performance baseline across the market. Prices don't move on release day, but value does: a GPU that looked expensive at $X/GPU-hr in v4.1 can look cheap if its benchmark score jumps 20% in v5.0. Inference and training software regimes are also getting more efficient every cycle: better kernels, smarter quantization, improved memory management. That means the GPU that was optimal for your workload six months ago may no longer be the right answer at today's prices. This page isolates that signal. The throughput table shows what changed on the benchmark side; the efficiency table translates that into procurement math by holding current market prices constant, so you can see which GPUs repriced themselves through software alone. When the gap is large enough, that's the indicator that it's time to re-evaluate your hardware selection entirely.