GPUs for on-prem AI / NVIDIA
NVIDIA A100 80GB SXM for LLMs
HGX A100 8-GPU systems remain common on-prem for 70B-class serving with tensor parallel over NVLink; no FP8. Price is an indicative 2026 street range for one unit, not a quote.
SXM4 | NVLink 600 GB/s | CUDA | indicative price $8,000-15K
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 | 131 | 2 | 109 |
| Ling-3.0-flash | 124B / 5.1B | 1 | 216 | 2 | 177 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 1 | 113 | 2 | 93 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 1 | 181 | 2 | 148 |
| gpt-oss-120b | 117B / 5.1B | 1 | 213 | 2 | 160 |
| Command A (03-2025) | 111B | 1 | 19 | 2 | 18 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 1 | 81 | 2 | 57 |
| Qwen3-Coder-Next | 79.7B / 3B | 1 | 295 | 2 | 204 |
| Llama 3.3 70B Instruct | 70B | 1 | 29 | 2 | 24 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 296 | 1 | 175 |
| Qwen3-32B | 32.8B | 1 | 59 | 1 | 27 |
| Olmo 3.1 32B Think | 32B | 1 | 60 | 1 | 30 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 289 | 1 | 138 |
| Gemma 4 31B IT | 30.7B | 1 | 62 | 1 | 28 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 267 | 1 | 104 |
| Granite 4.2 30B | 30B | 1 | 64 | 1 | 29 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 299 | 1 | 198 |
| Qwen3.8-27B | 27.8B | 1 | 70 | 1 | 38 |
| Gemma 3 27B IT | 27B | 1 | 71 | 1 | 35 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 248 | 1 | 115 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 78 | 1 | 37 |
| Magistral Small 1.2 (2509) | 24B | 1 | 78 | 1 | 37 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 78 | 1 | 37 |
| gpt-oss-20b | 21B / 3.6B | 1 | 266 | 1 | 152 |
| Phi-4 | 14B | 1 | 121 | 1 | 50 |
| Gemma 4 12B IT | 11.95B | 1 | 138 | 1 | 67 |
| Qwen3.5-9B | 9.7B | 1 | 167 | 1 | 95 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 173 | 1 | 73 |
| Granite 4.2 8B | 8B | 1 | 183 | 1 | 73 |
| Llama 3.1 8B Instruct | 8B | 1 | 185 | 1 | 80 |
| Command A+ (05-2026) | 218B / 25B | 2 | 93 | 4 | 84 |
| Step-3.7-Flash | 198B / 11B | 2 | 161 | 4 | 132 |
| Qwen3.8-Flash-Next | 180B / 6B | 2 | 221 | 4 | 176 |
| Mistral Medium 3.5 128B | 128B | 2 | 28 | 4 | 25 |
| MiniMax-M3 | 428B / 23B | 4 | 123 | 8 | 101 |
| Qwen3.5-397B-A17B | 403B / 17B | 4 | 143 | 8 | 122 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 4 | 151 | 8 | 116 |
| Trinity Large Thinking | 398B / 13B | 4 | 156 | 8 | 119 |
| GLM-5.3-Flash | 320B / 18B | 4 | 153 | 8 | 136 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 4 | 163 | 8 | 140 |
| Hy3 | 295B / 21B | 4 | 115 | 8 | 82 |
| Kimi-K2.6 | 1000B / 32B | 8 | 121 | 16 (multi-node) | 76 |
| Kimi-K2.7-Code | 1000B / 32B | 8 | 121 | 16 (multi-node) | 76 |
| Muse Spark Open 816B-A42B | 816B / 42B | 8 | 108 | 16 (multi-node) | 69 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 8 | 184 | 16 (multi-node) | 119 |
| GLM-5.3 | 744B / 40B | 8 | 100 | 16 (multi-node) | 62 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 8 | 110 | 16 (multi-node) | 71 |
| DeepSeek-V3.2 | 671B / 37B | 8 | 115 | 16 (multi-node) | 73 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 16 (multi-node) | 81 | 24 (multi-node) | 69 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 16 (multi-node) | 76 | 16 (multi-node) | 65 |
| Kimi-K3 | 2800B / 104B | 24 (multi-node) | 51 | 40 (multi-node) | 46 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 24 (multi-node) | 53 | 40 (multi-node) | 47 |
Design a deployment on NVIDIA A100 80GB SXM
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA A100 80GB SXM?
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 A100 80GB SXM for LLM inference?
Generation speed is mostly set by memory bandwidth: 2039 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 255 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA A100 80GB SXM a good choice for on-prem AI?
HGX A100 8-GPU systems remain common on-prem for 70B-class serving with tensor parallel over NVLink; no FP8. Price is an indicative 2026 street range for one unit, not a quote.