What GPU do you need to run Phi-3 Mini 3.8B?
3.82B params 128K max context
The smallest GPU memory that runs Phi-3 Mini 3.8B well is 10 GB (Q8_0, e.g. GeForce RTX 3080). 90 of 132 GPUs run it well and 41 more are a tight fit.
MMLU 70.9
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 128K |
|---|---|---|---|---|
| Q4_K_M | 2.4 GB | 8.2 GB | 14.1 GB | 49.2 GB |
| Q8_0 | 4.1 GB | 9.9 GB | 15.8 GB | 50.9 GB |
Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.
Cheapest GPUs that can run Phi-3 Mini 3.8B
Lowest launch MSRP among GPUs that run its smallest quantization (Q4_K_M) well, with at least 16K context.
As an Amazon Associate, I earn from qualifying purchases.
GPUs that can run Phi-3 Mini 3.8B
Best quantization that fits on each GPU, with the context and speed it gets. For chips with several memory sizes, the smallest that works is shown.
NVIDIA
- T1000 4 GB Q4_K_M 4K context · ~47 tok/s tight fit
- T400 4GB 4 GB Q4_K_M 4K context · ~24 tok/s tight fit
- T600 4 GB Q4_K_M 4K context · ~47 tok/s tight fit
- GTX 1060 6GB 6 GB Q8_0 5K context · ~38 tok/s tight fit
- GTX 1660 6 GB Q8_0 5K context · ~38 tok/s tight fit
- GTX 1660 SUPER 6 GB Q8_0 5K context · ~66 tok/s tight fit
- GTX 1660 Ti 6 GB Q8_0 5K context · ~57 tok/s tight fit
- RTX 2060 6 GB Q8_0 5K context · ~66 tok/s tight fit
- RTX 3050 6GB 6 GB Q8_0 5K context · ~33 tok/s tight fit
- RTX A2000 6 GB Q8_0 5K context · ~57 tok/s tight fit
- GTX 1070 8 GB Q8_0 10K context · ~51 tok/s tight fit
- GTX 1070 Ti 8 GB Q8_0 10K context · ~51 tok/s tight fit
- GTX 1080 8 GB Q8_0 10K context · ~63 tok/s tight fit
- RTX 2060 SUPER 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 2070 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 2070 SUPER 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 2080 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 2080 SUPER 8 GB Q8_0 10K context · ~98 tok/s tight fit
- RTX 3050 8 GB Q8_0 10K context · ~44 tok/s tight fit
- RTX 3060 Ti 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 3060 Ti GDDR6X 8 GB Q8_0 10K context · ~120 tok/s tight fit
- RTX 3070 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 3070 Ti 8 GB Q8_0 10K context · ~120 tok/s tight fit
- RTX 4060 8 GB Q8_0 10K context · ~54 tok/s tight fit
- RTX 4060 Ti 8 GB Q8_0 10K context · ~57 tok/s tight fit
- RTX 5050 8 GB Q8_0 10K context · ~63 tok/s tight fit
- RTX 5060 8 GB Q8_0 10K context · ~89 tok/s tight fit
- RTX 5060 Ti 8 GB Q8_0 10K context · ~89 tok/s tight fit
- T1000 8GB 8 GB Q8_0 10K context · ~32 tok/s tight fit
- RTX 3080 10 GB Q8_0 16K context · ~150 tok/s
- GTX 1080 Ti 11 GB Q8_0 18K context · ~96 tok/s
- RTX 2080 Ti 11 GB Q8_0 18K context · ~122 tok/s
- RTX 2060 12GB 12 GB Q8_0 21K context · ~66 tok/s
- RTX 3060 12 GB Q8_0 21K context · ~71 tok/s
- RTX 3080 12GB 12 GB Q8_0 21K context · ~180 tok/s
- RTX 3080 Ti 12 GB Q8_0 21K context · ~180 tok/s
- RTX 4070 12 GB Q8_0 21K context · ~100 tok/s
- RTX 4070 SUPER 12 GB Q8_0 21K context · ~100 tok/s
- RTX 4070 Ti 12 GB Q8_0 21K context · ~100 tok/s
- RTX 5070 12 GB Q8_0 21K context · ~133 tok/s
- RTX A2000 12GB 12 GB Q8_0 21K context · ~57 tok/s
- RTX 4060 Ti 16GB 16 GB Q8_0 32K context · ~57 tok/s
- RTX 4070 Ti SUPER 16 GB Q8_0 32K context · ~133 tok/s
- RTX 4080 16 GB Q8_0 32K context · ~142 tok/s
- RTX 4080 SUPER 16 GB Q8_0 32K context · ~145 tok/s
- RTX 5060 Ti 16GB 16 GB Q8_0 32K context · ~89 tok/s
- RTX 5070 Ti 16 GB Q8_0 32K context · ~177 tok/s
- RTX 5080 16 GB Q8_0 32K context · ~190 tok/s
- RTX 2000 Ada 16 GB Q8_0 32K context · ~44 tok/s
- RTX A4000 16 GB Q8_0 32K context · ~89 tok/s
- RTX PRO 2000 Blackwell 16 GB Q8_0 32K context · ~57 tok/s
- Tesla T4 16 GB Q8_0 32K context · ~63 tok/s
- Tesla V100 16GB 16 GB Q8_0 32K context · ~178 tok/s
- RTX 4000 Ada 20 GB Q8_0 43K context · ~71 tok/s
- RTX 3090 24 GB Q8_0 54K context · ~185 tok/s
- RTX 3090 Ti 24 GB Q8_0 54K context · ~199 tok/s
- RTX 4090 24 GB Q8_0 54K context · ~199 tok/s
- RTX 4500 Ada 24 GB Q8_0 54K context · ~85 tok/s
- RTX A5000 24 GB Q8_0 54K context · ~152 tok/s
- RTX PRO 4000 Blackwell 24 GB Q8_0 54K context · ~133 tok/s
- RTX Spark 24 GB Q8_0 54K context · ~59 tok/s
- Tesla P40 24 GB Q8_0 54K context · ~68 tok/s
- RTX 5090 32 GB Q8_0 76K context · ~354 tok/s
- RTX 5000 Ada 32 GB Q8_0 76K context · ~114 tok/s
- RTX PRO 4500 Blackwell 32 GB Q8_0 76K context · ~177 tok/s
- Tesla V100 32GB 32 GB Q8_0 76K context · ~178 tok/s
- L40S 48 GB Q8_0 120K context · ~171 tok/s
- RTX 6000 Ada 48 GB Q8_0 120K context · ~190 tok/s
- RTX A6000 48 GB Q8_0 120K context · ~152 tok/s
- RTX PRO 5000 Blackwell 48 GB Q8_0 120K context · ~266 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 128K context · ~266 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 128K context · ~276 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 128K context · ~354 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 128K context · ~54 tok/s
AMD
- Radeon RX 5600 XT 6 GB Q8_0 5K context · ~57 tok/s tight fit
- Radeon RX 5700 8 GB Q8_0 10K context · ~89 tok/s tight fit
- Radeon RX 5700 XT 8 GB Q8_0 10K context · ~89 tok/s tight fit
- Radeon RX 6600 8 GB Q8_0 10K context · ~44 tok/s tight fit
- Radeon RX 6600 XT 8 GB Q8_0 10K context · ~51 tok/s tight fit
- Radeon RX 6650 XT 8 GB Q8_0 10K context · ~55 tok/s tight fit
- Radeon RX 7600 8 GB Q8_0 10K context · ~57 tok/s tight fit
- Radeon RX 9060 8 GB Q8_0 10K context · ~57 tok/s tight fit
- Radeon RX 9060 XT 8 GB Q8_0 10K context · ~63 tok/s tight fit
- Radeon RX 6700 XT 12 GB Q8_0 21K context · ~76 tok/s
- Radeon RX 6750 XT 12 GB Q8_0 21K context · ~85 tok/s
- Radeon RX 7700 XT 12 GB Q8_0 21K context · ~85 tok/s
- Radeon RX 9070 GRE 12 GB Q8_0 21K context · ~85 tok/s
- AMD BC-250 16 GB Q8_0 32K context · ~89 tok/s
- Radeon RX 6800 16 GB Q8_0 32K context · ~101 tok/s
- Radeon RX 6800 XT 16 GB Q8_0 32K context · ~101 tok/s
- Radeon RX 6900 XT 16 GB Q8_0 32K context · ~101 tok/s
- Radeon RX 6950 XT 16 GB Q8_0 32K context · ~114 tok/s
- Radeon RX 7600 XT 16 GB Q8_0 32K context · ~57 tok/s
- Radeon RX 7800 XT 16 GB Q8_0 32K context · ~123 tok/s
- Radeon RX 7900 GRE 16 GB Q8_0 32K context · ~114 tok/s
- Radeon RX 9060 XT 16GB 16 GB Q8_0 32K context · ~63 tok/s
- Radeon RX 9070 16 GB Q8_0 32K context · ~126 tok/s
- Radeon RX 9070 XT 16 GB Q8_0 32K context · ~126 tok/s
- Radeon RX 7900 XT 20 GB Q8_0 43K context · ~158 tok/s
- Radeon RX 7900 XTX 24 GB Q8_0 54K context · ~190 tok/s
- Radeon AI PRO R9700 32 GB Q8_0 76K context · ~126 tok/s
- Ryzen AI Max (Strix Halo) 32 GB Q8_0 76K context · ~51 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 128K context · ~54 tok/s
Intel
- Arc A580 8 GB Q8_0 10K context · ~101 tok/s tight fit
- Arc A750 8 GB Q8_0 10K context · ~101 tok/s tight fit
- Arc A770 8GB 8 GB Q8_0 10K context · ~101 tok/s tight fit
- Arc B570 10 GB Q8_0 16K context · ~75 tok/s
- Arc B580 12 GB Q8_0 21K context · ~90 tok/s
- Arc A770 16GB 16 GB Q8_0 32K context · ~111 tok/s
- Arc Pro B65 32 GB Q8_0 76K context · ~120 tok/s
- Arc Pro B70 32 GB Q8_0 76K context · ~120 tok/s
Apple
- M1 16 GB Q8_0 32K context · ~13 tok/s
- M1 Pro 16 GB Q8_0 32K context · ~40 tok/s
- M2 16 GB Q8_0 32K context · ~20 tok/s
- M2 Pro 16 GB Q8_0 32K context · ~40 tok/s
- M3 16 GB Q8_0 32K context · ~20 tok/s
- M4 16 GB Q8_0 32K context · ~24 tok/s
- M5 16 GB Q8_0 32K context · ~30 tok/s
- M6 16 GB Q8_0 32K context · ~30 tok/s
- M3 Pro 18 GB Q8_0 38K context · ~30 tok/s
- M4 Pro 24 GB Q8_0 54K context · ~54 tok/s
- M5 Pro 24 GB Q8_0 54K context · ~61 tok/s
- M1 Max 32 GB Q8_0 76K context · ~79 tok/s
- M2 Max 32 GB Q8_0 76K context · ~79 tok/s
- M3 Max 36 GB Q8_0 87K context · ~59 tok/s
- M4 Max 36 GB Q8_0 87K context · ~81 tok/s
- M5 Max 36 GB Q8_0 87K context · ~91 tok/s
- M1 Ultra 64 GB Q8_0 128K context · ~158 tok/s
- M2 Ultra 64 GB Q8_0 128K context · ~158 tok/s
- M3 Ultra 96 GB Q8_0 128K context · ~162 tok/s
- M5 Ultra 96 GB Q8_0 128K context · ~237 tok/s