L4 vs T4
L4 has 1.5× the memory; T4 rents for 3.2× less.
The NVIDIA L4 (Ada Lovelace, 2023) and the NVIDIA T4 (Turing, 2018) are 5 years and a silicon generation apart, which makes the price gap the heart of the story. On capacity, the L4 carries 1.5x the memory (24 vs 16 GB), which sets what each can hold at all. Bandwidth is nearly even (0.3 vs 0.32 TB/s), so serving speed per GPU will be similar. Raw compute favors the L4 by 1.9x, which matters for training, prefill-heavy traffic, and diffusion work more than for chat-style decoding. The L4 also speaks native FP8 while the T4 tops out at FP16, a generational gap explained in CUDA cores vs Tensor Cores.
Price is where it settles: the T4 rents from $0.10/hr against $0.32/hr for the L4, a 3.2x gap that the performance numbers above only partly close. Per gigabyte of VRAM, the T4 is the cheaper rental, worth knowing if your model is capacity-bound rather than speed-bound. For how these numbers translate into tokens per dollar, the inference cost estimator runs both cards against any model. Everything below is the underlying data.
Head to head
■ L4 ■ T4, bars share one scale across the whole catalog.
| L4 | T4 | |
|---|---|---|
| Memory | 24 GB GDDR6 | 16 GB GDDR6 |
| Bandwidth | 0.3 TB/s | 0.32 TB/s |
| FP16 dense | 121 TF | 65 TF |
| FP8 dense | 242 TF | - |
| TDP | 72 W | 70 W |
| Interconnect | PCIe Gen4 | PCIe Gen3 |
| Cheapest rental | $0.32/hr | $0.10/hr |
| $/hr per GB VRAM | $1.3¢ | $0.6¢ |
What one can run that the other can't
Only on L4
- Gemma 2 27B (4-bit)
- gpt-oss-20b (4-bit)
Only on T4
Nothing, L4 runs everything T4 does.
Rule of thumb: for LLM serving, prefer the GPU with more memory bandwidth per dollar; for training and prefill-heavy work, prefer FLOPS per dollar; and if the model doesn't fit in VRAM, none of the other numbers matter. Sanity-check with the VRAM calculator.