← Back to GPUs
NVIDIA · RTX 40
NVIDIA GeForce RTX 4070 Ti
$600$799 MSRP
The original RTX 4070 Ti shipped with only 12GB VRAM, which was widely criticized. The SUPER variant fixed this with 16GB. On the used market, the 12GB 4070 Ti is a good gaming card but limited for AI — you're restricted to 7B-8B models. If AI matters, get the 16GB SUPER instead.
Best ForUsed market 1440p/4K gaming
VerdictGood gaming card used, but the 12GB VRAM kills its AI potential — get the SUPER.
AI
5/10
Gaming
8/10
Specifications
VRAM12GB GDDR6X
Memory Bandwidth504 GB/s
CUDA Cores7,680
Boost Clock2610 MHz
TDP285W
Power Connector1x 16-pin
Length290mm
Form FactorDual Slot
Release Year2023
AI Capabilities
Entry Level12GB VRAM
Limited to small models with heavy quantization. Fine for experimenting.
Can run (Q4 quantized)
Llama 3.1 8BQwen 2.5 14BMistral 7BFLUX.1 DevStable Diffusion XLStable Diffusion 3.5 LargeCogVideoX-5BMochi 1LTX VideoStable Video DiffusionWan Video 14BAlphaFold 2ESMFold (ESM-2 15B)ESM-2 3BscGPTRFdiffusionFine-tune Llama 8BTrain SDXL LoRA
Tight fit (may need CPU offload)
HunyuanVideo (14GB Q4)Codestral 22B (13GB Q4)Train FLUX LoRA (16GB Q4)
Recommended system RAM for AI: 24GB+ (2x GPU VRAM for model overflow)
Performance Estimates
Estimated tokens/sec for LLM inference based on 504 GB/s memory bandwidth — not hardware benchmarks. Methodology · What is Q4/Q8?
Llama 3.1 8B8B
Q8~37-45 tok/sFastQwen 2.5 14B14B
Q4~34-42 tok/sFastMistral 7B7B
Q8~42-52 tok/sFastCodestral 22B22B
Offload~1-3 tok/sVery slowPros
- +Strong gaming performance
- +Good used prices
- +DLSS 3 support
Cons
- -Only 12GB VRAM — AI-limited
- -Superseded by 16GB SUPER
- -16-pin connector
gaming