GPUs for on-prem AI / NVIDIA
NVIDIA H100 SXM 80GB for LLMs
Reference GPU for production LLM serving in HGX/DGX 8-GPU systems, with the broadest engine support (vLLM, TensorRT-LLM, SGLang). NVIDIA lists 1,979/3,958 TFLOPS with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.
SXM5 | NVLink 900 GB/s | CUDA | indicative price $22K-30K
Which models fit, and how many cards
Planning estimates. 4-bit column: one user at 8K context. FP8 column: 10 users at once at 16K context.
| Model | Size | 4-bit: cards | 4-bit: tok/s | FP8, 10 users: cards | FP8: tok/s per user |
|---|---|---|---|---|---|
| Qwen3.5-122B-A10B | 125B / 10B | 1 | 190 | 2 | 153 |
| Ling-3.0-flash | 124B / 5.1B | 1 | 294 | 2 | 230 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 1 | 168 | 2 | 128 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 1 | 253 | 2 | 201 |
| gpt-oss-120b | 117B / 5.1B | 1 | 290 | 2 | 214 |
| Command A (03-2025) | 111B | 1 | 31 | 2 | 28 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 1 | 123 | 2 | 85 |
| Qwen3-Coder-Next | 79.7B / 3B | 1 | 377 | 2 | 253 |
| Llama 3.3 70B Instruct | 70B | 1 | 47 | 2 | 38 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 378 | 1 | 246 |
| Qwen3-32B | 32.8B | 1 | 92 | 1 | 43 |
| Olmo 3.1 32B Think | 32B | 1 | 94 | 1 | 48 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 371 | 1 | 200 |
| Gemma 4 31B IT | 30.7B | 1 | 96 | 1 | 45 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 349 | 1 | 155 |
| Granite 4.2 30B | 30B | 1 | 99 | 1 | 46 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 381 | 1 | 273 |
| Qwen3.8-27B | 27.8B | 1 | 107 | 1 | 60 |
| Gemma 3 27B IT | 27B | 1 | 108 | 1 | 56 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 329 | 1 | 171 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 120 | 1 | 59 |
| Magistral Small 1.2 (2509) | 24B | 1 | 120 | 1 | 59 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 120 | 1 | 59 |
| gpt-oss-20b | 21B / 3.6B | 1 | 348 | 1 | 218 |
| Phi-4 | 14B | 1 | 178 | 1 | 78 |
| Gemma 4 12B IT | 11.95B | 1 | 201 | 1 | 104 |
| Qwen3.5-9B | 9.7B | 1 | 236 | 1 | 143 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 243 | 1 | 111 |
| Granite 4.2 8B | 8B | 1 | 256 | 1 | 112 |
| Llama 3.1 8B Instruct | 8B | 1 | 258 | 1 | 121 |
| Command A+ (05-2026) | 218B / 25B | 2 | 135 | 4 | 120 |
| Step-3.7-Flash | 198B / 11B | 2 | 212 | 4 | 169 |
| Qwen3.8-Flash-Next | 180B / 6B | 2 | 269 | 4 | 209 |
| Mistral Medium 3.5 128B | 128B | 2 | 43 | 4 | 38 |
| MiniMax-M3 | 428B / 23B | 4 | 155 | 8 | 125 |
| Qwen3.5-397B-A17B | 403B / 17B | 4 | 173 | 8 | 143 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 4 | 187 | 8 | 143 |
| Trinity Large Thinking | 398B / 13B | 4 | 185 | 8 | 141 |
| GLM-5.3-Flash | 320B / 18B | 4 | 190 | 8 | 164 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 4 | 197 | 8 | 165 |
| Hy3 | 295B / 21B | 4 | 140 | 8 | 101 |
| Kimi-K2.6 | 1000B / 32B | 8 | 142 | 16 (multi-node) | 86 |
| Kimi-K2.7-Code | 1000B / 32B | 8 | 142 | 16 (multi-node) | 86 |
| Muse Spark Open 816B-A42B | 816B / 42B | 8 | 130 | 16 (multi-node) | 80 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 8 | 207 | 16 (multi-node) | 129 |
| GLM-5.3 | 744B / 40B | 8 | 117 | 16 (multi-node) | 69 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 8 | 132 | 16 (multi-node) | 82 |
| DeepSeek-V3.2 | 671B / 37B | 8 | 137 | 16 (multi-node) | 83 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 16 (multi-node) | 89 | 24 (multi-node) | 77 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 16 (multi-node) | 83 | 16 (multi-node) | 74 |
| Kimi-K3 | 2800B / 104B | 24 (multi-node) | 56 | 40 (multi-node) | 50 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 24 (multi-node) | 57 | 40 (multi-node) | 51 |
Design a deployment on NVIDIA H100 SXM 80GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA H100 SXM 80GB?
Of the models we track, Qwen3.5-122B-A10B (125B total, 10B active) fits on a single 80 GB card in 4-bit with one user at 8K context. More users or longer context need memory for the KV cache, so leave headroom.
How fast is NVIDIA H100 SXM 80GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 3350 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 419 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA H100 SXM 80GB a good choice for on-prem AI?
Reference GPU for production LLM serving in HGX/DGX 8-GPU systems, with the broadest engine support (vLLM, TensorRT-LLM, SGLang). NVIDIA lists 1,979/3,958 TFLOPS with sparsity, halved here. Price is an indicative 2026 street range for one unit, not a quote.