GPUs for on-prem AI / NVIDIA
NVIDIA A100 40GB PCIe for LLMs
Previous-generation workhorse, now mostly sold used or refurbished; no FP8. NVIDIA lists 312 TFLOPS BF16 dense (624 with sparsity). Price is an indicative 2026 street range for one unit, not a quote.
PCIe | NVLink bridge 600 GB/s (2 GPUs) | CUDA | indicative price $4,000-8,000
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 |
|---|---|---|---|---|---|
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 252 | 2 | 186 |
| Qwen3-32B | 32.8B | 1 | 46 | 2 | 35 |
| Olmo 3.1 32B Think | 32B | 1 | 47 | 2 | 39 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 246 | 2 | 151 |
| Gemma 4 31B IT | 30.7B | 1 | 48 | 2 | 36 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 225 | 2 | 119 |
| Granite 4.2 30B | 30B | 1 | 50 | 2 | 37 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 255 | 1 | 162 |
| Qwen3.8-27B | 27.8B | 1 | 55 | 1 | 29 |
| Gemma 3 27B IT | 27B | 1 | 55 | 2 | 45 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 208 | 1 | 92 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 61 | 2 | 48 |
| Magistral Small 1.2 (2509) | 24B | 1 | 61 | 2 | 48 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 61 | 2 | 48 |
| gpt-oss-20b | 21B / 3.6B | 1 | 224 | 1 | 123 |
| Phi-4 | 14B | 1 | 96 | 1 | 39 |
| Gemma 4 12B IT | 11.95B | 1 | 111 | 1 | 52 |
| Qwen3.5-9B | 9.7B | 1 | 135 | 1 | 75 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 140 | 1 | 57 |
| Granite 4.2 8B | 8B | 1 | 150 | 1 | 57 |
| Llama 3.1 8B Instruct | 8B | 1 | 151 | 1 | 62 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 2 | 116 | 4 | 98 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 2 | 194 | 4 | 164 |
| gpt-oss-120b | 117B / 5.1B | 2 | 220 | 4 | 174 |
| Command A (03-2025) | 111B | 2 | 25 | 4 | 24 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 2 | 97 | 4 | 72 |
| Qwen3-Coder-Next | 79.7B / 3B | 2 | 261 | 4 | 195 |
| Llama 3.3 70B Instruct | 70B | 2 | 37 | 4 | 32 |
| Command A+ (05-2026) | 218B / 25B | 4 | 114 | 8 | 102 |
| Step-3.7-Flash | 198B / 11B | 4 | 168 | 8 | 139 |
| Qwen3.8-Flash-Next | 180B / 6B | 4 | 206 | 8 | 168 |
| Mistral Medium 3.5 128B | 128B | 4 | 36 | 8 | 33 |
| Qwen3.5-122B-A10B | 125B / 10B | 4 | 173 | 4 | 124 |
| Ling-3.0-flash | 124B / 5.1B | 4 | 228 | 4 | 183 |
| MiniMax-M3 | 428B / 23B | 8 | 124 | 16 (multi-node) | 76 |
| Qwen3.5-397B-A17B | 403B / 17B | 8 | 138 | 16 (multi-node) | 85 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 8 | 151 | 16 (multi-node) | 89 |
| Trinity Large Thinking | 398B / 13B | 8 | 146 | 16 (multi-node) | 84 |
| GLM-5.3-Flash | 320B / 18B | 8 | 155 | 16 (multi-node) | 101 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 8 | 159 | 16 (multi-node) | 100 |
| Hy3 | 295B / 21B | 8 | 110 | 16 (multi-node) | 60 |
| Muse Spark Open 816B-A42B | 816B / 42B | 16 (multi-node) | 78 | 24 (multi-node) | 66 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 16 (multi-node) | 126 | 24 (multi-node) | 111 |
| GLM-5.3 | 744B / 40B | 16 (multi-node) | 68 | 24 (multi-node) | 58 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 16 (multi-node) | 79 | 24 (multi-node) | 68 |
| DeepSeek-V3.2 | 671B / 37B | 16 (multi-node) | 81 | 24 (multi-node) | 70 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 24 (multi-node) | 71 | 32 (multi-node) | 66 |
| Kimi-K2.6 | 1000B / 32B | 24 (multi-node) | 81 | 32 (multi-node) | 77 |
| Kimi-K2.7-Code | 1000B / 32B | 24 (multi-node) | 81 | 32 (multi-node) | 77 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 32 (multi-node) | 80 | 48 (multi-node) | 70 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 40 (multi-node) | 50 | - | - |
| Kimi-K3 | 2800B / 104B | 48 (multi-node) | 50 | - | - |
Design a deployment on NVIDIA A100 40GB PCIe
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one NVIDIA A100 40GB PCIe?
Of the models we track, Qwen3.6-35B-A3B (36B total, 3B active) fits on a single 40 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 40GB PCIe for LLM inference?
Generation speed is mostly set by memory bandwidth: 1555 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 194 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is NVIDIA A100 40GB PCIe a good choice for on-prem AI?
Previous-generation workhorse, now mostly sold used or refurbished; no FP8. NVIDIA lists 312 TFLOPS BF16 dense (624 with sparsity). Price is an indicative 2026 street range for one unit, not a quote.