GPUs for gaming and local AI
56 cards · specs from the makers · rental prices from 1h ago · How we compute this
The short answer
Buy by VRAM first. 16 GB is the sweet spot for 1440p gaming and for local AI models up to about 14B; 24 GB opens up 24–32B models and 4K; 70B-class models need 48 GB or more, which most people rent by the hour. Right now the cheapest 16 GB+ consumer card to rent is the RTX 4060 Ti from $0.053/hr.
Start with a guide
- Best GPU for local AI
- Best budget GPU for gaming
- RTX 5060 Ti vs RX 9060 XT
- Best GPU for 1440p: RTX 5070 vs RX 9070 XT
- Buying a used RTX 3090
- Is the RTX 5090 worth it for local AI?
- Is 8GB of VRAM enough for gaming?
- Best budget GPU for local AI under $500
VRAM calculatorCloud GPU pricesBest GPU for each AI modelWhat each model costs to runCan my PC run it? (games)AI video on a GPU (Wan 2.2)
8–12 GB
1080p and light 1440p gaming; small local models (7–8B) at good quality.
| GPU | VRAM | Bandwidth | Largest model at Q4+ | Rent from |
|---|---|---|---|---|
| RTX 3080 | 10 GB | 760 GB/s | Gemma 3 12B | — |
| RTX 3080 Ti | 12 GB | 912 GB/s | DeepSeek R1 Distill 14B | $0.095/hr |
| RTX 3080 | 12 GB | 912 GB/s | DeepSeek R1 Distill 14B | — |
| RTX 5070Buy | 12 GB | 672 GB/s | DeepSeek R1 Distill 14B | $0.098/hr |
| RTX 4070 Ti | 12 GB | 504 GB/s | DeepSeek R1 Distill 14B | $0.092/hr |
| RTX 4070 SUPER | 12 GB | 504 GB/s | DeepSeek R1 Distill 14B | $0.093/hr |
| RTX 4070 | 12 GB | 504 GB/s | DeepSeek R1 Distill 14B | $0.093/hr |
| RTX 3060 | 12 GB | 360 GB/s | DeepSeek R1 Distill 14B | $0.033/hr |
16 GB
The 1440p gaming sweet spot; local models up to about 14B at Q4, more with MoE offload.
| GPU | VRAM | Bandwidth | Largest model at Q4+ | Rent from |
|---|---|---|---|---|
| RTX 5080Buy | 16 GB | 960 GB/s | Gemma 4 26B-A4B (MoE) | $0.131/hr |
| RTX 5070 TiBuy | 16 GB | 896 GB/s | Gemma 4 26B-A4B (MoE) | $0.121/hr |
| RTX 4080 SUPER | 16 GB | 736 GB/s | Gemma 4 26B-A4B (MoE) | $0.136/hr |
| RTX 4080 | 16 GB | 716.8 GB/s | Gemma 4 26B-A4B (MoE) | $0.130/hr |
| RTX 4070 Ti SUPER | 16 GB | 672 GB/s | Gemma 4 26B-A4B (MoE) | $0.067/hr |
| AMD Radeon RX 9070 XTBuy | 16 GB | 640 GB/s | Gemma 4 26B-A4B (MoE) | — |
| AMD Radeon RX 9070Buy | 16 GB | 640 GB/s | Gemma 4 26B-A4B (MoE) | — |
| RTX 5060 TiBuy | 16 GB | 448 GB/s | Gemma 4 26B-A4B (MoE) | $0.065/hr |
| RTX A4000 | 16 GB | 448 GB/s | Gemma 4 26B-A4B (MoE) | $0.053/hr |
| AMD Radeon RX 9060 XTBuy | 16 GB | 320 GB/s | Gemma 4 26B-A4B (MoE) | — |
| RTX 4060 Ti | 16 GB | 288 GB/s | Gemma 4 26B-A4B (MoE) | $0.053/hr |
20–24 GB
4K gaming; 24–32B local models at Q4 with room for context.
| GPU | VRAM | Bandwidth | Largest model at Q4+ | Rent from |
|---|---|---|---|---|
| AMD Radeon RX 7900 XT | 20 GB | 800 GB/s | NVIDIA Nemotron 3 Nano 30B-A3B (MoE) | — |
| RTX 4090 | 24 GB | 1008 GB/s | Qwen3.6 35B-A3B (MoE) | $0.170/hr |
| RTX 3090 Ti | 24 GB | 1008 GB/s | Qwen3.6 35B-A3B (MoE) | $0.101/hr |
| AMD Radeon RX 7900 XTX | 24 GB | 960 GB/s | Qwen3.6 35B-A3B (MoE) | — |
| RTX 3090 | 24 GB | 936 GB/s | Qwen3.6 35B-A3B (MoE) | $0.080/hr |
| RTX A5000 | 24 GB | 768 GB/s | Qwen3.6 35B-A3B (MoE) | $0.106/hr |
32–48 GB
Top-end gaming and workstation cards; 70B-class models at low-bit quants.
| GPU | VRAM | Bandwidth | Largest model at Q4+ | Rent from |
|---|---|---|---|---|
| RTX 5090Buy | 32 GB | 1792 GB/s | Mixtral 8x7B (MoE) | $0.200/hr |
| RTX 5000 Ada Generation | 32 GB | 576 GB/s | Mixtral 8x7B (MoE) | $0.267/hr |
| Apple M4 Max (14-core CPU / 32-core GPU) | 36 GB | 410 GB/s | Mixtral 8x7B (MoE) | — |
| RTX 6000 Ada Generation | 48 GB | 960 GB/s | Qwen3-Next 80B-A3B (MoE) | $0.347/hr |
| RTX A6000 | 48 GB | 768 GB/s | Qwen3-Next 80B-A3B (MoE) | $0.107/hr |
| AMD Radeon PRO W7900 | 48 GB | 864 GB/s | Qwen3-Next 80B-A3B (MoE) | — |
64 GB and up
Datacenter GPUs and unified-memory machines; mostly rented by the hour.
| GPU | VRAM | Bandwidth | Largest model at Q4+ | Rent from |
|---|---|---|---|---|
| RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96 GB | 1792 GB/s | Qwen3.5 122B-A10B (MoE) | $0.443/hr |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | 1792 GB/s | Qwen3.5 122B-A10B (MoE) | $0.520/hr |
| Apple M4 Max (16-core CPU / 40-core GPU) | 128 GB | 546 GB/s | Qwen3.8-Flash-Next (MoE) | — |
| DGX Spark (GB10 Grace Blackwell) | 128 GB | 273 GB/s | Qwen3.8-Flash-Next (MoE) | $0.200/hr |
| AMD Ryzen AI Max+ 395 (Strix Halo,) | 128 GB | 256 GB/s | Qwen3.8-Flash-Next (MoE) | — |
| Apple M4 Pro | 64 GB | 273 GB/s | Llama 4 Scout 17B-16E (MoE) | — |
| Apple M2 Max | 96 GB | 409.6 GB/s | Qwen3.5 122B-A10B (MoE) | — |
| Apple M2 Ultra | 192 GB | 819.2 GB/s | Qwen3 235B-A22B (MoE) | — |
| Apple M3 Ultra | 512 GB | 819.2 GB/s | DeepSeek V4.1 Flash (MoE) | — |
Datacenter GPUs (rent by the hour)
Cards nobody buys for a desk, but everyone compares when renting for big models.
| GPU | VRAM | Bandwidth | Largest model at Q4+ | Rent from |
|---|---|---|---|---|
| T4 | 16 GB | 320 GB/s | Gemma 4 26B-A4B (MoE) | $0.053/hr |
| L4 | 24 GB | 300 GB/s | Qwen3.6 35B-A3B (MoE) | $0.045/hr |
| V100 SXM2 | 32 GB | 900 GB/s | Mixtral 8x7B (MoE) | $0.230/hr |
| V100 PCIe | 32 GB | 900 GB/s | Mixtral 8x7B (MoE) | — |
| A100 PCIe | 40 GB | 1555 GB/s | Mixtral 8x7B (MoE) | — |
| A100 SXM | 40 GB | 1555 GB/s | Mixtral 8x7B (MoE) | $1.000/hr |
| A40 | 48 GB | 696 GB/s | Qwen3-Next 80B-A3B (MoE) | $0.173/hr |
| L40S | 48 GB | 864 GB/s | Qwen3-Next 80B-A3B (MoE) | $0.200/hr |
| A100 PCIe | 80 GB | 1935 GB/s | Qwen3.5 122B-A10B (MoE) | $1.790/hr |
| H100 PCIe | 80 GB | 2000 GB/s | Qwen3.5 122B-A10B (MoE) | $0.667/hr |
| H100 SXM | 80 GB | 3350 GB/s | Qwen3.5 122B-A10B (MoE) | $0.667/hr |
| A100 SXM | 80 GB | 2039 GB/s | Qwen3.5 122B-A10B (MoE) | $0.999/hr |
| RTX PRO 6000 Blackwell Server Edition | 96 GB | 1597 GB/s | Qwen3.5 122B-A10B (MoE) | $1.750/hr |
| H200 SXM | 141 GB | 4800 GB/s | Qwen3 235B-A22B (MoE) | $1.316/hr |
| AMD Instinct MI300X | 192 GB | 5300 GB/s | Qwen3 235B-A22B (MoE) | $2.390/hr |
| B200 | 192 GB | 8000 GB/s | Qwen3 235B-A22B (MoE) | $3.125/hr |
Choosing a GPU: common questions
How much VRAM do I need for gaming?
8 GB still runs most games at 1080p with settings turned down; 12–16 GB is the comfortable range for 1440p in new releases; 16 GB or more for 4K and heavy mods. Our guide on whether 8 GB is enough has the game-by-game detail.
How much VRAM do I need for local AI?
About 8 GB runs 7–8B models, 16 GB runs up to about 14B at Q4 quality, 24 GB runs 24–32B models, and 70B-class models need 48 GB or more (or two cards). The VRAM calculator gives the exact figure for a model and context length.
What is the cheapest way to get 24 GB of VRAM?
Renting: right now the cheapest 24 GB+ card we track is the L4 from $0.045/hr. Buying: a used RTX 3090 is the usual answer; our used-3090 guide covers what to check.