H100 vs H200 vs B200: choose by bottleneck and availability.

H100: established Hopper performance

Choose H100 when 80 GB is enough and the software stack benefits from Hopper transformer acceleration. It has the broadest GPURento request geography of the three and the lower public reference rate than H200.

H200: solve a memory-bound workload

Choose H200 when 141 GB and higher memory bandwidth reduce sharding, increase context or raise concurrency enough to offset the higher hourly reference. If the workload remains compute-bound and fits 80 GB, validate the gain.

• Measure GPUs required for the same model
• Keep context and concurrency fixed
• Compare cost per accepted token or checkpoint

B200: use Blackwell where it pays

B200 offers the largest memory of the three and a newer architecture at $5.98 per GPU-hour. Use it when Blackwell throughput, 180 GB memory or FP4 support offsets the higher hourly rate.

Benchmark the same system boundary

Pin the model and dataset revision, container digest, precision, context or batch, quality target and checkpoint policy. Record warm-up separately, then compare p50 and p95 end-to-end runtime, peak memory and total compute cost. Manufacturer peaks are specifications, not results from GPURento infrastructure.

• Use the exact GPU form factor shown in the request
• Include data loading, preprocessing and checkpoint I/O
• Report cost per accepted token, checkpoint or completed job