AMD Instinct MI350X 288GBBudget pick
C130.3 tok/s decode
~$8,000 MSRP
Can it run?
Runs with offload (needs ~5.5 GB host RAM)
Using Q4_K_M in llama.cpp
Fit status
Runs with offload (needs ~5.5 GB host RAM)
Decode
69.9 tok/s
TTFT
2768 ms
Safe context
15K
Memory
102.6 GB / 96.0 GB
Offload
10%
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Very compromised (needs ~10 GB host RAM) | 68.0 tok/s | 4138 ms | 28K |
| Chat | C | Runs with offload (needs ~3.1 GB host RAM) | 71.0 tok/s | 1488 ms | 8K |
| Coding | C | Runs with offload (needs ~5.5 GB host RAM) | 69.9 tok/s | 2768 ms | 15K |
| RAG | C | Very compromised (needs ~10 GB host RAM) | 68.0 tok/s | 5173 ms | 28K |
| Reasoning | C | Runs with offload (needs ~5.5 GB host RAM) | 69.9 tok/s | 3271 ms | 15K |
How Mixtral 8x22B (141B params) fits at each quantization level on NVIDIA GH200 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 55.0 GB | Low | C42 |
Q3_K_SBest for your GPU | 3 | 69.1 GB | Low | C45 |
NVFP4 | 4 | 79.0 GB | Medium | C45 |
Q4_K_M | 4 | 86.0 GB | Medium | C45 |
Q5_K_M | 5 | 101.5 GB | High | F0 |
Q6_K | 6 | 115.6 GB | High | F0 |
Q8_0 | 8 | 150.9 GB | Very High | F0 |
F16 | 16 | 289.0 GB | Maximum | F0 |
ollama run mixtral-8x22bhuggingface-cli download mixtral-8x22bUpgrade options
~$8,000 MSRP
~$15,000 MSRP
NVIDIA H200 141GBBiggest leap~$30,000 MSRP
NVIDIA H200 PCIe 141GBNVIDIA upgrade~$30,000 MSRP