ToolJoltTools

H100 80GB LLM Calculator — What Can It Run?

Check which models fit on a H100 80GB: max parameters per precision, cache headroom and estimated tokens/sec.

Memory needed (GB)
Headroom on this card (GB)
Max params at this quant (B)
Est. decode speed (tok/s)

H100 80GB: 80 GB, 3350 GB/s, ~989 TFLOPS FP16. The H100 added FP8 tensor cores (989 dense BF16 TFLOPS, ~2× that in FP8) and 3.35 TB/s HBM3 — the card the 2023–25 frontier was trained and served on, and the unit in which AI compute is now priced.

Formula

needed = params × bpw ÷ 8 + reserve · fits if ≤ 80 GB · decode tok/s ≈ 0.6 × 3350 GB/s ÷ weight-GB (bandwidth-bound)
References: NVIDIA/Apple official H100 80GB specifications; llama.cpp quantization size tables

About H100 80GB LLM Calculator — What Can It Run?

"Will it run?" is the first question of local AI, and for the H100 80GB this calculator answers it precisely: enter any model's parameter count and quantization and get the memory bill against this card's 80 GB, the largest model it can hold at that quant, and a bandwidth-derived decode-speed estimate (token generation streams the whole model per token, so 3350 GB/s is the speed limit that matters). The H100 added FP8 tensor cores (989 dense BF16 TFLOPS, ~2× that in FP8) and 3.35 TB/s HBM3 — the card the 2023–25 frontier was trained and served on, and the unit in which AI compute is now priced.

How to use H100 80GB LLM Calculator — What Can It Run?

  1. 1Enter your values into H100 80GB LLM Calculator — What Can It Run? — sensible, domain-typical defaults are pre-filled so you see a real result immediately.
  2. 2The result recomputes live using the formula shown on the page; there is no button to press.
  3. 3Adjust any input to compare scenarios, then read the worked example to see the substituted numbers.

Why use H100 80GB LLM Calculator — What Can It Run??

  • Computes H100 80GB LLM instantly in your browser — no sign-up, no upload, no server round-trip.
  • 100% free and unlimited, with the exact formula shown: needed = params × bpw ÷ 8 + reserve.
  • Runs entirely client-side, so every value you enter stays private on your device.
  • Live recompute as you type, with a worked example and authoritative references for trust.

Frequently asked questions

What makes H100 worth 2–3× the A100's price?+

Training: ~3× BF16 compute plus FP8 (≈2× again) with Transformer Engine — large-run cost per token drops despite the premium. Serving: 64% more bandwidth speeds decode, and FP8 weights double effective capacity. For light inference, the premium often isn't justified.

How many H100s for Llama-3-70B at full precision?+

BF16 weights are ~141 GB — two H100s minimum with tensor parallelism, leaving ~19 GB across the pair for cache (tight); four is the comfortable production setup. In FP8 it serves on a single card (~71 GB) for moderate contexts.

How is the tokens/sec estimate for the H100 80GB derived?+

Decode is memory-bound: each token reads every weight once, so speed ≈ effective bandwidth ÷ model size. We assume ~60% of the 3350 GB/s peak is achievable, matching llama.cpp benchmarks within ~20%. Prompt prefill is compute-bound and much faster per token.

Why reserve memory beyond the weights?+

The KV cache grows with context (use our per-model KV-cache calculators), CUDA/Metal runtimes take hundreds of MB, and allocator fragmentation wastes more. The default reserve suits 2–8K contexts; long-context work needs significantly more.

Related tools

Related ML & AI tools

Sponsored