NVIDIA Tesla T4

The Tesla T4 is a Turing-era datacenter GPU with 16 GB of GDDR6 at 320 GB/s in a 70W power envelope - no PCIe power cables needed. Used prices around $300-400 make it an interesting budget option for local AI. It has Turing Tensor cores (FP16) and supports INT8 inference, making it more capable than the older Tesla P40 for modern LLM workloads.

Quick verdict

The Tesla T4 is a compelling budget AI card if you can handle passive cooling. 16 GB VRAM at only 70W with Tensor cores is unique. It runs Mistral Small 3 at Q4_K_M and 7-13B models well. The 320 GB/s bandwidth is slow, and Turing lacks modern precision support, but for $350 the value is solid.

Spec breakdown

• VRAM: 16 GB GDDR6
• Memory bandwidth: 320 GB/s
• TDP: 70 W (no power cables)
• Architecture: Turing TU104
• Tensor cores: 320 (Turing)
• Form factor: Single slot, passive

Real-world AI inference

Model	Tokens/sec	Source
Llama 3.1 8B Q4_K_M	~25 tok/s	Community
Mistral Small 3 Q4_K_M	~10 tok/s	Community
Qwen3-30B	Offload needed	VRAM limited

Honest alternatives

• Used RTX 3060 12GB (~$200): Cheaper, also 12 GB, easier setup
• Tesla P40 24GB (~$150): More VRAM, older architecture, cheaper
• Intel Arc A770 16GB (~$230): Cheaper, faster, easier to use