NVIDIA A100 40GBBudget pickB66.9 tok/s decode
~$10,000 MSRP
Can it run?
Tight fit
Using Q4_K_M in Ollama
Fit status
Tight fit
Decode
66.1 tok/s
TTFT
2929 ms
Safe context
18K
Memory
28.9 GB / 32.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs with offload (needs ~1.1 GB host RAM) | 58.7 tok/s | 4795 ms | 30K |
| Chat | C | Tight fit | 61.5 tok/s | 1717 ms | 10K |
| Coding | C | Tight fit | 66.1 tok/s | 2929 ms | 18K |
| RAG | C | Runs with offload (needs ~1.1 GB host RAM) | 58.7 tok/s | 5994 ms | 30K |
| Reasoning | C | Tight fit | 61.5 tok/s | 3720 ms |
How Baichuan M2 32B Q4 K M (32B params) fits at each quantization level on RTX 5090 32GB (32.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 12.5 GB | Low | D37 |
Q3_K_S | 3 | 15.7 GB | Low | D39 |
NVFP4 | 4 |
Upgrade options
NVIDIA A100 40GBBudget pick~$10,000 MSRP
RTX PRO 5000 Blackwell 48GBBiggest leap
RTX 6000 Ada 48GBNVIDIA upgrade
| 18K |
17.9 GB |
| Medium |
| C41 |
Q4_K_M | 4 | 19.5 GB | Medium | C42 |
Q5_K_MBest for your GPU | 5 | 23.0 GB | High | C44 |
Q6_K | 6 | 26.2 GB | High | C44 |
Q8_0 | 8 | 34.2 GB | Very High | F0 |
F16 | 16 | 65.6 GB | Maximum | F0 |