Discrete GPU
AMD Radeon Instinct MI360X
192 GB$18000~172.8 GB usable
- Type
- Discrete GPU
- Provider
- AMD
- VRAM
- 192 GB
- Price
- $18000
Usable for models (est.): 172.8 GB · Price per GB: $94
Popular models that run here
The largest models this device holds fully in memory, each at its best-fitting quantization.
Llama 3.1 405B405B parametersBest fit: INT2
DeepSeek-V4-Flash (MoE)284B parametersBest fit: FP4
Qwen 3 235B-A22B (MoE)235B parametersBest fit: FP4
Mistral Large 2 (MoE)141B parametersBest fit: FP8- Qwen 2.5 72B72B parametersBest fit: FP16
- Llama 3 70B70.6B parametersBest fit: FP16
Llama 3.3 70B70.6B parametersBest fit: FP16- Llama 3.1 70B70B parametersBest fit: FP16
Every model: does it fit?
39 of 41 models run fully in memory, 1 more with partial offload. Estimated at a 4096-token context. How fit is calculated
| Model | Params | Best fit | Memory needed (est.) | Fit |
|---|---|---|---|---|
| DeepSeek-V4-Pro (MoE) | 1600B | — | ~480.0 GB | ✕ Does not fit |
| DeepSeek-V3 (MoE) | 671B | — | ~201.3 GB | ◐ Partial offload |
| Llama 3.1 405B | 405B | INT2 | ~121.5 GB | ✓ Runs |
| DeepSeek-V4-Flash (MoE) | 284B | FP4 | ~170.4 GB | ✓ Runs |
| Qwen 3 235B-A22B (MoE) | 235B | FP4 | ~141.0 GB | ✓ Runs |
| Mistral Large 2 (MoE) | 141B | FP8 | ~169.2 GB | ✓ Runs |
| Qwen 2.5 72B | 72B | FP16 | ~172.8 GB | ✓ Runs |
| Llama 3 70B | 70.6B | FP16 | ~169.4 GB | ✓ Runs |
| Llama 3.3 70B | 70.6B | FP16 | ~169.4 GB | ✓ Runs |
| Llama 3.1 70B | 70B | FP16 | ~168.0 GB | ✓ Runs |
| Llama 2 70B | 70B | FP16 | ~168.0 GB | ✓ Runs |
| Falcon 40B | 40B | FP16 | ~96.0 GB | ✓ Runs |
| Qwen 2.5 32B | 32B | FP32 | ~153.6 GB | ✓ Runs |
| Qwen 3 32B | 32B | FP32 | ~153.6 GB | ✓ Runs |
| OLMo 3 32B | 32B | FP32 | ~153.6 GB | ✓ Runs |
| Gemma 4 31B | 31B | FP32 | ~148.8 GB | ✓ Runs |
| Qwen 3 30B-A3B (MoE) | 30B | FP32 | ~144.0 GB | ✓ Runs |
| Gemma 3 27B | 27B | FP32 | ~129.6 GB | ✓ Runs |
| Phi-4 | 14B | FP32 | ~67.2 GB | ✓ Runs |
| Llama 2 13B | 13B | FP32 | ~62.4 GB | ✓ Runs |
| OLMo 2 13B | 13B | FP32 | ~62.4 GB | ✓ Runs |
| Mistral Nemo 12B | 12B | FP32 | ~57.6 GB | ✓ Runs |
| Gemma 3 12B | 12B | FP32 | ~57.6 GB | ✓ Runs |
| Gemma 2 9B | 9B | FP32 | ~43.2 GB | ✓ Runs |
| Llama 3 8B | 8B | FP32 | ~38.4 GB | ✓ Runs |
| Llama 3.1 8B | 8B | FP32 | ~38.4 GB | ✓ Runs |
| Qwen 3 8B | 8B | FP32 | ~38.4 GB | ✓ Runs |
| Llama 3.3 8B | 8B | FP32 | ~38.4 GB | ✓ Runs |
| Llama 2 7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| Mistral 7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| Qwen 2.5 7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| Falcon 7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| OLMo 2 7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| OLMo-7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| Mistral Small 3 7B | 7B | FP32 | ~33.6 GB | ✓ Runs |
| Gemma 3 4B | 4B | FP32 | ~19.2 GB | ✓ Runs |
| Phi-3 | 3.8B | FP32 | ~18.2 GB | ✓ Runs |
| Phi-3-mini-128K | 3.8B | FP32 | ~18.2 GB | ✓ Runs |
| Llama 3.2 3B | 3B | FP32 | ~14.4 GB | ✓ Runs |
| Llama 3.2 1B | 1B | FP32 | ~4.8 GB | ✓ Runs |
| Gemma 3 1B | 1B | FP32 | ~4.8 GB | ✓ Runs |
Measured throughput
No one has submitted a measured benchmark for this device yet. The fit list above is an estimate.
Similar GPUs
- NVIDIA RTX 4090 (24 GB)
- NVIDIA RTX 6000 Ada (48 GB)
- NVIDIA A100 40GB PCIe (40 GB)
- NVIDIA A100 80GB PCIe (80 GB)
- NVIDIA H100 PCIe (80 GB)
- NVIDIA H100 SXM (80 GB)
- NVIDIA RTX 4080 Super (16 GB)
- NVIDIA RTX 4070 Ti Super (16 GB)
Alternatives with similar memory
- Mac Studio M5 Ultra 128GB (128 GB)
- Mac Studio M5 Ultra (96 GB)
- Mac Studio M4 Max 64GB (64 GB)
- MacBook Pro M4 Max 16-inch (48 GB)