L40S vs T4
NVIDIA L40S (Ada Lovelace, 48 GB) against NVIDIA T4 (Turing, 16 GB): memory, compute, power and rental price, compared for LLM inference and training.
Pick two GPUs to compare
Side-by-Side Specifications
| Spec | L40S | T4 |
|---|---|---|
| Architecture | Ada Lovelace | Turing |
| Memory | 48 GB GDDR6 | 16 GB GDDR6 |
| Memory bandwidth | 864 GB/s | 320 GB/s |
| FP16 tensor compute | 733 TFLOPS | 65 TFLOPS |
| INT8 tensor compute | 1,466 TOPS | 130 TOPS |
| Interconnect | PCIe 4.0 · 64 GB/s | PCIe 3.0 · 32 GB/s |
| TDP | 350 W | 70 W |
| Est. on-demand price | ~$3.50/h | ~$0.50/h |
| FP16 TFLOPS per $/h | 209 | 130 |
Highlighted values indicate the stronger spec. Hourly rates are indicative on-demand estimates.
Verdict
Raw performance: The L40S leads on FP16 tensor compute (11.3x advantage), which translates directly into higher token throughput for inference and shorter training steps.
Memory: With 48 GB per card, the L40S fits larger models on fewer GPUs - fewer cards means less inter-GPU communication and simpler deployments.
Value: At current on-demand rates, the L40S delivers more compute per dollar (209 vs 130 FP16 TFLOPS per $/h). If your model fits in its VRAM budget, it is usually the more economical choice.
GPUs Needed for Popular LLMs
Cards required to serve each model at 8-bit quantization (with 20% overhead for activations and KV cache).
| Model | VRAM (8-bit) | L40S | T4 |
|---|---|---|---|
| GPT-5.6 Sol | 2682 GB | 56x | 168x |
| GLM 5.3 (743B) | 830 GB | 18x | 52x |
| Claude 4.8 Opus (300B) | 335 GB | 7x | 21x |
| Grok 3 Mini | 212 GB | 5x | 14x |
| Qwen 3.6 Plus (110B) | 123 GB | 3x | 8x |
| Nova Core (34B) | 38 GB | 1x | 3x |
| Phi 4 (14B) | 16 GB | 1x | 1x |
| Mage VL | 5 GB | 1x | 1x |
| LFM 2.5 2.6B | 3 GB | 1x | 1x |
Frequently Asked Questions
Which is better for LLM inference: L40S or T4?
The L40S delivers more raw FP16 compute (733 TFLOPS) and the L40S offers the most memory per card (48 GB). For cost-efficiency, the L40S currently gives more FP16 TFLOPS per dollar of on-demand rental (209 vs 130 TFLOPS per $/h).
How much more memory does the L40S have?
The L40S has 48 GB of GDDR6 versus 16 GB of GDDR6 for the T4 - a ratio of 3.00x in favor of the L40S. More VRAM per card means fewer GPUs to fit a given model.
Is the L40S or the T4 cheaper to rent?
Estimated on-demand rates are ~$3.50/h for the L40S and ~$0.50/h for the T4. Raw hourly price is only part of the story: normalize by throughput (TFLOPS per $/h) and by how many cards you need for your model's VRAM.
How do the L40S and T4 compare on power?
The L40S has a TDP of 350W versus 70W for the T4. FP16 compute per watt: 2.1 vs 0.9 TFLOPS/W.
Deploy on a GPU cloud
Rent the L40S or T4 by the hour instead of buying hardware.
