Local LLM Hardware

Local LLM on NVIDIA H100 80GB

80GB VRAM · HBM3 · 3350 GB/s memory bandwidth · 700W TDP · released 2022

At a glance

VRAM80 GB
MemoryHBM3
Memory bandwidth3350 GB/s
TDP700 W
Released2022
2026 price (used / new)~$30000
Tierdata-center flagship
Best forfastest single-node inference in 2024-2025

What models fit on NVIDIA H100 80GB?

Memory bandwidth is the bottleneck for LLM token generation. The NVIDIA H100 80GB's 3350 GB/s bandwidth gives roughly the throughput below at Q4_K_M quantization. Numbers are conservative and assume a single request (batch=1), 2048-token context.

ModelQuantizationFits?Approx tokens/sec
Llama 3.2 1BQ4_K_MYes (0.8 GB)~603 tok/s
Llama 3.1 8BQ4_K_MYes (~5 GB)~201
Llama 3.1 8BQ8_0Yes (~8 GB)~151
Qwen 2.5 14BQ4_K_MYes (~9 GB)~117
Mistral Nemo 12BQ4_K_MYes (~8 GB)~134
Qwen 3 32BQ4_K_MYes (~20 GB)~60
DeepSeek R1 distilled 32BQ4_K_MYes (~20 GB)~60
Llama 3.1 70BQ4_K_MYes (~42 GB)~34
Llama 4 Scout 109B (MoE)Q4_K_MYes (~64 GB)~27

Token estimates use a simple bandwidth heuristic (~tokens/sec ≈ bandwidth_GB/s × model_efficiency). Real numbers vary with model architecture, batch size, and KV-cache size. Treat as ballpark.

Build recommendations for NVIDIA H100 80GB

Best model to download first

Llama 3.1 8B for everyday chat and coding; Qwen 3 32B for top quality on 24GB. Start with ollama pull llama3.1:8b.

Recommended inference backend

Use Ollama for general use or Mullama if you need a drop-in Ollama alternative with native bindings for 6 languages. For production serving on multi-GPU setups, see vLLM.

Other GPUs to consider

Sources

[1] nvidia.com/h100 · [2] vLLM benchmarks