GPU
La VRAM décide quels modèles tiennent ; la bande passante mémoire décide à quelle vitesse ils génèrent.
| GPU | VRAM | Bande passante | Puissance de la carte | Plus grand modèle en Q4_K_M, 8K |
|---|---|---|---|---|
| RTX PRO 6000 BlackwellNVIDIA · Blackwell | 96 Go | 1 792 Go/s | 600 W | gpt-oss-120b |
| H100 PCIeNVIDIA · Hopper | 80 Go | 2 000 Go/s | 350 W | gpt-oss-120b |
| A100 80GBNVIDIA · Ampere | 80 Go | 1 935 Go/s | 300 W | gpt-oss-120b |
| RTX 6000 AdaNVIDIA · Ada Lovelace | 48 Go | 960 Go/s | 300 W | Llama 3.3 70B Instruct |
| Radeon PRO W7900AMD · RDNA 3 | 48 Go | 864 Go/s | 295 W | Llama 3.3 70B Instruct |
| RTX A6000NVIDIA · Ampere | 48 Go | 768 Go/s | 300 W | Llama 3.3 70B Instruct |
| RTX 5090NVIDIA · Blackwell | 32 Go | 1 792 Go/s | 575 W | Mixtral 8x7B Instruct |
| RTX 4090NVIDIA · Ada Lovelace | 24 Go | 1 008 Go/s | 450 W | Qwen3 32B |
| RTX 3090 TiNVIDIA · Ampere | 24 Go | 1 008 Go/s | 450 W | Qwen3 32B |
| RX 7900 XTXAMD · RDNA 3 | 24 Go | 960 Go/s | 355 W | Qwen3 32B |
| RTX 3090NVIDIA · Ampere | 24 Go | 936 Go/s | 350 W | Qwen3 32B |
| Tesla P40NVIDIA · Pascal | 24 Go | 346 Go/s | 250 W | Qwen3 32B |
| RX 7900 XTAMD · RDNA 3 | 20 Go | 800 Go/s | 315 W | Gemma 3 27B IT |
| RTX 5080NVIDIA · Blackwell | 16 Go | 960 Go/s | 360 W | gpt-oss-20b |
| RTX 5070 TiNVIDIA · Blackwell | 16 Go | 896 Go/s | 300 W | gpt-oss-20b |
| RTX 4080 SUPERNVIDIA · Ada Lovelace | 16 Go | 736 Go/s | 320 W | gpt-oss-20b |
| RTX 4080NVIDIA · Ada Lovelace | 16 Go | 716,8 Go/s | 320 W | gpt-oss-20b |
| RTX 4070 Ti SUPERNVIDIA · Ada Lovelace | 16 Go | 672 Go/s | 285 W | gpt-oss-20b |
| RX 9070 XTAMD · RDNA 4 | 16 Go | 640 Go/s | 304 W | gpt-oss-20b |
| RX 9070AMD · RDNA 4 | 16 Go | 640 Go/s | 220 W | gpt-oss-20b |
| RX 7800 XTAMD · RDNA 3 | 16 Go | 624 Go/s | 263 W | gpt-oss-20b |
| RX 7900 GREAMD · RDNA 3 | 16 Go | 576 Go/s | 260 W | gpt-oss-20b |
| Arc A770 16 GBIntel · Xe-HPG (Alchemist) | 16 Go | 560 Go/s | 225 W | gpt-oss-20b |
| RX 6800 XTAMD · RDNA 2 | 16 Go | 512 Go/s | 300 W | gpt-oss-20b |
| RTX 5060 Ti 16 GBNVIDIA · Blackwell | 16 Go | 448 Go/s | 180 W | gpt-oss-20b |
| RTX 4060 Ti 16 GBNVIDIA · Ada Lovelace | 16 Go | 288 Go/s | 165 W | gpt-oss-20b |
| RX 7600 XTAMD · RDNA 3 | 16 Go | 288 Go/s | 190 W | gpt-oss-20b |
| RTX 3080 TiNVIDIA · Ampere | 12 Go | 912 Go/s | 350 W | Qwen3 14B |
| RTX 5070NVIDIA · Blackwell | 12 Go | 672 Go/s | 250 W | Qwen3 14B |
| RTX 4070 TiNVIDIA · Ada Lovelace | 12 Go | 504 Go/s | 285 W | Qwen3 14B |
| RTX 4070 SUPERNVIDIA · Ada Lovelace | 12 Go | 504 Go/s | 220 W | Qwen3 14B |
| RTX 4070NVIDIA · Ada Lovelace | 12 Go | 504 Go/s | 200 W | Qwen3 14B |
| Arc B580Intel · Xe2 (Battlemage) | 12 Go | 456 Go/s | 190 W | Qwen3 14B |
| RTX 3060 12 GBNVIDIA · Ampere | 12 Go | 360 Go/s | 170 W | Qwen3 14B |
| RTX 2080 TiNVIDIA · Turing | 11 Go | 616 Go/s | 250 W | Mistral Nemo 12B Instruct |
| RTX 3080NVIDIA · Ampere | 10 Go | 760 Go/s | 320 W | Mistral Nemo 12B Instruct |
| RTX 5060 Ti 8 GBNVIDIA · Blackwell | 8 Go | 448 Go/s | 180 W | Qwen3 8B |
| RTX 5060NVIDIA · Blackwell | 8 Go | 448 Go/s | 145 W | Qwen3 8B |
| RTX 3070NVIDIA · Ampere | 8 Go | 448 Go/s | 220 W | Qwen3 8B |
| RTX 4060 Ti 8 GBNVIDIA · Ada Lovelace | 8 Go | 288 Go/s | 160 W | Qwen3 8B |
| RTX 4060NVIDIA · Ada Lovelace | 8 Go | 272 Go/s | 115 W | Qwen3 8B |
Spécification officielleDonnées mises à jour il y a 24 joursSource: NVIDIA, AMD, Intel