Use-case guide · Vision-language

Should you pick H100 NVL 94GB (per GPU pair) for vision-language?

H100 NVL 94GB (per GPU pair) has 188 GB VRAM. Whether it's the right fit for vision-language depends on your model size, expected QPS, and budget. Below is what we're seeing in production.

VRAM + model fit

H100 NVL 94GB (per GPU pair) fits models up to ~132B parameters in BF16 comfortably with room for KV-cache. For vision-language specifically, you'll want to leave headroom for context length growth.

Pricing

Live pricing across all providers for H100 NVL 94GB (per GPU pair) is on the GPU detail page — click through for the sortable list.

Throughput

On vision-language workloads, H100 NVL 94GB (per GPU pair) typically delivers the throughput published in its FP16 spec, minus the framework overhead (vLLM ≈ 85% MFU, TGI ≈ 70%).

Try the calculator to size the hardware for your specific model, or see all GPUs on the InferenceScore leaderboard.