Local LLM on NVIDIA H100 80GB

Is the NVIDIA H100 80GB the right card for this?

This is an accelerator you rent or requisition rather than one you drop into a desktop. HBM3 stacks put 3350 GB/s behind 80 GB, which is the combination that makes batched serving worthwhile: the same weights stream fast enough to answer many concurrent requests, not just one. At roughly $30000 it costs about $375 per gigabyte of memory, so nobody buys one to chat with a 7B model. Its case is throughput per rack slot, and the figure our catalogue records for it is FP16 70B: 50-70, FP16 405B: 18-25, with vLLM: 100+ tok/s aggregate.

One other card here carries 80 GB, so the two load the same weights and differ only in how fast they read them. At 3350 GB/s this is the quickest of that group, so what you gain over the others is tokens per second, not a model you could not otherwise open.

Released in 2022, this is one of the older entries in the guide, and age here is mostly an advantage: the current generation of local-AI tooling was written while this hardware was already widespread, so backend support is settled and the failure modes are documented. What it cannot have is any of the memory technology that arrived after it, and memory is what sets generation speed.

What the price gap buys

The cheapest 80 GB option in this guide is the NVIDIA A100 80GB at about $15000. This card costs $15000 more and reads memory 64% faster, which is roughly the gain in tokens per second. It buys no additional model: both load exactly the same weights, so the question is only whether you are waiting on the output.

A part at about $30000 is not bought the way a graphics card is bought. Most people meet this hardware by the hour from a cloud provider, and the figure above is useful mainly as the number to compare a rental bill against. If you are specifying one, the 700 W draw and the cooling it implies decide the chassis before the accelerator does.

Where the NVIDIA H100 80GB sits in this guide

Across the 20 cards catalogued here, the NVIDIA H100 80GB ranks 2nd on memory bandwidth at 3350 GB/s, 19th cheapest at about $30000, and 8th on bandwidth per watt at 4.8 GB/s per watt. Those three positions, not the capacity figure, are what separate it from other 80 GB parts.

HBM3 stacks memory on the package beside the processor, which is why 80 GB delivers 3350 GB/s where board-mounted GDDR at the same capacity would deliver a fraction of it. It is also why the part costs what it costs: HBM is expensive to package and cannot be soldered on later.

This is a CUDA card, which in practice means every local runner in our tool directory supports it first. Ollama and llama.cpp pick it up without configuration; vLLM and TensorRT-LLM target it specifically. Driver and CUDA toolkit versions are the usual source of trouble, not the card.

Which catalogue models fit on NVIDIA H100 80GB?

Of the 46 models in our catalogue with per-size memory figures, 41 have at least one size that fits in 80 GB at Q4_K_M with room for context. The largest of each are below, biggest first.

Model	Largest size that fits	Weights at Q4_K_M
Mistral Large	123B	70 GB
GPT-OSS	120B	70 GB
Llama 3.2 Vision	90B	55 GB
Qwen 2.5	72B	44 GB
Llama 3.1	70B	42 GB
Llama 3.3	70B	42 GB
DeepSeek R1	70B	42 GB
Code Llama	70B	42 GB

A further 33 smaller models also fit; see the model catalogue.

The next sizes up are out of reach without offloading: Qwen 3 235B at 130 GB, minimax M2 230B at 130 GB, minimax M2.1 230B at 130 GB.

Expected tokens per second on NVIDIA H100 80GB

Memory bandwidth, not compute, is the bottleneck for single-stream token generation. The NVIDIA H100 80GB's 3350 GB/s puts the following ceiling on a batch of one at 2048 tokens of context.

Model	Quantization	Status	Approx tokens/sec
Llama 3.2 1B	Q4_K_M	Fits in VRAM	~603 tok/s
Llama 3.1 8B	Q4_K_M	Fits in VRAM	~201 tok/s
Llama 3.1 8B	Q8_0	Fits in VRAM	~151 tok/s
Mistral Nemo 12B	Q4_K_M	Fits in VRAM	~134 tok/s
Qwen 2.5 14B	Q4_K_M	Fits in VRAM	~117 tok/s
Qwen 3 32B	Q4_K_M	Fits in VRAM	~60 tok/s
DeepSeek R1 distilled 32B	Q4_K_M	Fits in VRAM	~60 tok/s
Llama 3.1 70B	Q4_K_M	Fits in VRAM	~34 tok/s
Llama 4 Scout 109B (MoE)	Q4_K_M	Fits in VRAM	~27 tok/s

These come from a bandwidth heuristic (tokens/sec ≈ bandwidth in GB/s × a per-model efficiency factor), not from a test rig in our office. Real numbers move with model architecture, batch size and KV-cache size. Treat them as an order of magnitude, and treat the "reported throughput class" row in the table above as the community-measured figure.

Build recommendations for NVIDIA H100 80GB

Best model to download first

Llama 3.1 8B to check the stack works, then the largest model in the table above that you actually need. At 80 GB the constraint stops being capacity. Start with ollama pull llama3.1:8b.

Recommended inference backend

At this capacity the point is serving several people at once, so use vLLM or TensorRT-LLM for continuous batching rather than a single-stream runner.

Sources

[1] nvidia.com/h100 · [2] vLLM benchmarks · VRAM figures for catalogue models from src/data/models.json (2026-06-29); card specifications from src/data/gpus.json (2026-06-29).