gpt-neox-20b
by EleutherAI
20.6B params · text-generation · 579 likes · 179.2k downloads
gpt-neox-20b is a 20.6B parameter model. At Q4 quantization it requires 10GB of VRAM. It runs comfortably on GeForce RTX 4090 (63 tok/s), GeForce RTX 5090 (95 tok/s), A100 PCIe 80 GB (116 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 63 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 95 tok/s | comfortable |
| M4 Max 128GB | 128GB | 23 tok/s | tight |
| M4 Pro 48GB | 48GB | 11 tok/s | tight |
| M4 Pro 24GB | 24GB | 11 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 116 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 234 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 56 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 46 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 45 tok/s | comfortable |