AMD Instinct MI350X 288GBBest value
C71.7 tok/s decode
~$8,000 MSRP
Can it run?
Too heavy
Using Q4_K_M in vLLM
Fit status
Too heavy
Decode
35.3 tok/s
TTFT
5478 ms
Safe context
6K
Memory
258.7 GB / 96.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | F | Too heavy | 35.3 tok/s | 7968 ms | 12K |
| Chat | F | Too heavy | 35.3 tok/s | 2988 ms | 4K |
| Coding | F | Too heavy | 35.3 tok/s | 5478 ms | 6K |
| RAG | F | Too heavy | 35.3 tok/s | 9961 ms | 12K |
| Reasoning | F | Too heavy | 35.3 tok/s | 6474 ms | 6K |
How Llama 4 Maverick 17B 128E (400B params) fits at each quantization level on NVIDIA GH200 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 156.0 GB | Low | F0 |
Q3_K_S | 3 | 196.0 GB | Low | F0 |
NVFP4 | 4 |
Upgrade options
~$8,000 MSRP
224.0 GB |
| Medium |
| F0 |
Q4_K_M | 4 | 244.0 GB | Medium | F0 |
Q5_K_M | 5 | 288.0 GB | High | F0 |
Q6_K | 6 | 328.0 GB | High | F0 |
Q8_0 | 8 | 428.0 GB | Very High | F0 |
F16 | 16 | 820.0 GB | Maximum | F0 |