AMD Instinct MI300X 192GBBudget pick
C92.2 tok/s decode
~$15,000 MSRP
Can it run?
Tight fit
Using Q4_K_M in llama.cpp
Fit status
Tight fit
Decode
59.4 tok/s
TTFT
3258 ms
Safe context
19K
Memory
105.8 GB / 128.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Tight fit | 57.8 tok/s | 4875 ms | 37K |
| Chat | B | Runs well | 57.8 tok/s | 1828 ms | 10K |
| Coding | C | Tight fit | 59.4 tok/s | 3258 ms | 19K |
| RAG | C | Tight fit | 57.8 tok/s | 6093 ms | 37K |
| Reasoning | C | Tight fit | 57.8 tok/s | 3961 ms | 19K |
How Mixtral 8x22B (141B params) fits at each quantization level on Gaudi 3 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 55.0 GB | Low | D39 |
Q3_K_S | 3 | 69.1 GB | Low | C41 |
NVFP4 | 4 |
ollama run mixtral-8x22bhuggingface-cli download mixtral-8x22bUpgrade options
~$15,000 MSRP
NVIDIA H200 141GBBest value~$30,000 MSRP
NVIDIA H200 PCIe 141GBBiggest leap~$30,000 MSRP
79.0 GB |
| Medium |
| C43 |
Q4_K_M | 4 | 86.0 GB | Medium | C44 |
Q5_K_MBest for your GPU | 5 | 101.5 GB | High | C45 |
Q6_K | 6 | 115.6 GB | High | C45 |
Q8_0 | 8 | 150.9 GB | Very High | F0 |
F16 | 16 | 289.0 GB | Maximum | F0 |