NVIDIA L40S vs NVIDIA T4, AI & Machine Learning Comparison
NVIDIA L40S wins 145 of 153 benchmarks, averaging 303.7% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Sana 1.6B, where NVIDIA L40S leads by 4396% (50.81 vs 1.13 images/min); the closest fight is Qwen2.5 1.5B + 0.5B draft (5% apart); VRAM decides part of this one: NVIDIA L40S runs 152 of our 12 AI workloads while the other card runs 111, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA L40S | NVIDIA T4 | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 212.05 | 67.71 | +213% |
| Llama-3.1-8B tok/s | 135.51 | 35.01 | +287% |
| Qwen2.5-Coder-14B tok/s | 74.6 | 19.76 | +278% |
| Qwen3-32B tok/s | 34.44 | 0 | n/a |
| Llama-3.3-70B tok/s | 16.49 | 0 | n/a |
| Stable Diffusion XL images/min | 17.02 | 2.36 | +621% |
| FLUX.1 dev images/min | 3.836 | 0 | n/a |
| FLUX.1 Kontext dev images/min | 1.693 | 0 | n/a |
| Qwen-Image-Edit images/min | 1.06 | 0 | n/a |
| Wan 2.2 5B (720p) frames/s | 0.49 | 0 | n/a |
| DeepSeek-R1 Distill Llama 8B tok/s | 135.55 | 36.36 | +273% |
| DeepSeek-R1 Distill 1.5B tok/s | 426.38 | 148.72 | +187% |
| DeepSeek-R1 Distill 14B tok/s | 74.6 | 19.21 | +288% |
| DeepSeek-R1 Distill 7B tok/s | 143.97 | 37.58 | +283% |
| Gemma 3 12B tok/s | 80.92 | 23.12 | +250% |
| Gemma 3 4B tok/s | 198.87 | 65.12 | +205% |
| Gemma 4 12B tok/s | 80.82 | 24.03 | +236% |
| Llama 3.2 1B tok/s | 618.87 | 207.71 | +198% |
| Llama 3.2 3B tok/s | 270.85 | 85.94 | +215% |
| Mistral 7B v0.3 tok/s | 144.74 | 40.01 | +262% |
| Mistral Small 24B tok/s | 48.43 | 11.33 | +327% |
| Phi-4 14B tok/s | 74.98 | 17.46 | +329% |
| Phi-4 Mini 3.8B tok/s | 228.93 | 64.15 | +257% |
| Qwen2.5-Coder 7B tok/s | 143.93 | 37.54 | +283% |
| Qwen3 0.6B tok/s | 655.75 | 263.47 | +149% |
| Qwen3 1.7B tok/s | 405.52 | 139.92 | +190% |
| Qwen3 14B tok/s | 75.12 | 19.31 | +289% |
| Qwen3 30B A3B tok/s | 213.66 | 0 | n/a |
| Qwen3 8B tok/s | 130.27 | 36.51 | +257% |
| SmolLM3 3B tok/s | 272.88 | 86.08 | +217% |
| Codestral 22B tok/s | 50.06 | 12.59 | +298% |
| DeepSeek-R1 Distill 32B tok/s | 34.6 | 0 | n/a |
| Devstral Small 24B tok/s | 48.43 | 11.4 | +325% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 33.08 | 0 | n/a |
| Dolphin Mistral 24B Venice tok/s | 48.43 | 11.36 | +326% |
| Dolphin X1 8B tok/s | 135.48 | 35.66 | +280% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 135.55 | 35.68 | +280% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 48.4 | 11.32 | +328% |
| Gemma 3 27B tok/s | 38.48 | 0 | n/a |
| Qwen2.5-Coder 32B tok/s | 34.59 | 0 | n/a |
| Qwen3-Coder 30B A3B tok/s | 219.04 | 0 | n/a |
| QwQ 32B tok/s | 34.59 | 0 | n/a |
| StarCoder2 15B tok/s | 66.88 | 15.66 | +327% |
| FLUX.1 Schnell images/min | 26.24 | 0 | n/a |
| BiRefNet images/min | 863.81 | 183.48 | +371% |
| Depth Anything V2 Large images/min | 959.33 | 402.34 | +138% |
| Depth Anything V2 Small images/min | 970.61 | 612.25 | +59% |
| SAM ViT-Base images/min | 998.79 | 180.73 | +453% |
| SAM ViT-Huge images/min | 238.08 | 34.91 | +582% |
| Swin2SR 4x Upscaler images/min | 34.14 | 10.53 | +224% |
| Kokoro TTS 82M x realtime | 243.99 | 42.88 | +469% |
| Whisper large-v3 x realtime | 193.2 | 44.36 | +336% |
| MusicGen Small x realtime | 2.43 | 0.92 | +164% |
| Florence-2 Base images/min | 184.2 | 148.61 | +24% |
| Florence-2 Large images/min | 106.54 | 83.43 | +28% |
| Dolphin X1 Trinity Nano 6B tok/s | 246.45 | 114.21 | +116% |
| gpt-oss-20b tok/s | 233.82 | 63.63 | +267% |
| Meta-Llama-3.1-8B tok/s | 135.46 | 37.8 | +258% |
| Ornith-1.0-35B tok/s | 157.39 | 0 | n/a |
| Qwen-AgentWorld-35B-A3B tok/s | 156.98 | 0 | n/a |
| Qwen2.5-3B tok/s | 270.04 | 86.04 | +214% |
| Qwen2.5-7B tok/s | 143.94 | 39.01 | +269% |
| Olmo-3.1-32B-Think tok/s | 34.11 | 0 | n/a |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 48.43 | 11.33 | +327% |
| DeepSeek-Coder-V2-Lite tok/s | 257.64 | 83.49 | +209% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 130.23 | 37.7 | +245% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 16.49 | 0 | n/a |
| EVA-Qwen2.5-14B-v0.2 tok/s | 74.63 | 19.92 | +275% |
| Qwen2.5 1.5B LoRA train tok/s | 12120.1 | 1434.4 | +745% |
| Qwen2.5 7B LoRA train tok/s | 3736 | 0 | n/a |
| SmolLM2 1.7B LoRA train tok/s | 12052.1 | 1561.9 | +672% |
| TinyLlama 1.1B LoRA train tok/s | 16408.1 | 1520.1 | +979% |
| gemma-2-2b-it-abliterated tok/s | 278.14 | 91.91 | +203% |
| gemma-2-2b tok/s | 278.1 | 92.13 | +202% |
| gemma-2-9b tok/s | 88.87 | 28.62 | +211% |
| gemma-3-1b tok/s | 426.49 | 156.88 | +172% |
| gemma-3-270m tok/s | 822.52 | 362.8 | +127% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 146.51 | 53.83 | +172% |
| GLM-4.7-Flash tok/s | 157.89 | 0 | n/a |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 130.25 | 35.62 | +266% |
| SmolLM3-3B tok/s | 272.96 | 86.86 | +214% |
| KAT-Coder-V2.5-Dev tok/s | 173.06 | 0 | n/a |
| L3-8B-Stheno-v3.2 tok/s | 135.53 | 36.28 | +274% |
| LFM2.5-1.2B tok/s | 661.54 | 221.66 | +198% |
| LFM2.5-8B-A1B tok/s | 394.48 | 139.03 | +184% |
| Llama-2-7B tok/s | 150.35 | 42.98 | +250% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 271.01 | 86.29 | +214% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 16.49 | 0 | n/a |
| Meta-Llama-3.1-70B tok/s | 16.49 | 0 | n/a |
| Phi-4-mini tok/s | 229.28 | 66 | +247% |
| Mistral-7B-Instruct-v0.1 tok/s | 144.82 | 40.63 | +256% |
| Mistral-7B-Instruct-v0.2 tok/s | 144.82 | 41.31 | +251% |
| Mistral-7B-Instruct-v0.3 tok/s | 144.68 | 39.87 | +263% |
| Mistral-Nemo-Instruct-2407 tok/s | 89.72 | 22.98 | +290% |
| Ornith-1.0-9B tok/s | 114.57 | 32.93 | +248% |
| phi-2 tok/s | 283.26 | 89.6 | +216% |
| Phi-3.5-mini tok/s | 229.06 | 72.07 | +218% |
| Qwen3-0.6B tok/s | 656.26 | 259.91 | +152% |
| Qwen3-1.7B tok/s | 405.3 | 135.21 | +200% |
| Qwen3-14B tok/s | 75.01 | 19.44 | +286% |
| Qwen3-8B tok/s | 130.28 | 37.37 | +249% |
| Qwen2.5-0.5B tok/s | 738.06 | 295.47 | +150% |
| Qwen2.5-1.5B tok/s | 427.23 | 148.14 | +188% |
| Qwen2.5-14B tok/s | 74.62 | 19.84 | +276% |
| Qwen2.5-32B tok/s | 34.59 | 0 | n/a |
| Qwen2.5-Coder-1.5B tok/s | 427.12 | 148.51 | +188% |
| Qwen2.5-Coder-3B tok/s | 269.9 | 85.74 | +215% |
| Qwen3-4B-Instruct-2507 tok/s | 212.05 | 67.45 | +214% |
| SmolLM2-135M tok/s | 916.56 | 416.24 | +120% |
| Cydonia-24B-v4.3 tok/s | 48.43 | 11.21 | +332% |
| Hermes-4-14B tok/s | 75.1 | 19.02 | +295% |
| Qwen3-30B-A3B tok/s | 214.05 | 0 | n/a |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 74.63 | 18.86 | +296% |
| Qwen3-4B-Thinking-2507 tok/s | 211.88 | 67 | +216% |
| Qwen2.5 1.5B served serve tok/s | 4686.8 | 1337.1 | +251% |
| Qwen2.5 7B served serve tok/s | 1406.8 | 0 | n/a |
| SmolLM2 1.7B served serve tok/s | 3525.9 | 1222.5 | +188% |
| TinyLlama 1.1B served serve tok/s | 6278.4 | 1897.8 | +231% |
| Codestral 22B (Q3_K_M) tok/s | 59.53 | 11.3 | +427% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 224.95 | 61.82 | +264% |
| AuraFlow v0.3 images/min | 3.58 | 0 | n/a |
| PixArt-Sigma XL images/min | 26.95 | 0 | n/a |
| Sana 1.6B images/min | 50.81 | 1.13 | +4396% |
| Stable Diffusion 3.5 Large images/min | 4.54 | 0 | n/a |
| Stable Diffusion 3.5 Medium images/min | 11.58 | 0 | n/a |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.946 | 0.9 | +5% |
| AI21-Jamba-Reasoning-3B tok/s | 258.41 | 86.91 | +197% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 134.78 | 35.73 | +277% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 34.43 | 0 | n/a |
| dolphin-2.9-llama3-8b tok/s | 134.94 | 35.1 | +284% |
| Hermes-3-Llama-3.2-3B tok/s | 265.87 | 84.99 | +213% |
| Hermes-4-70B tok/s | 16.45 | 0 | n/a |
| Qwen3-Coder-Next-abliterated tok/s | 0 | 0 | n/a |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 88.85 | 23.19 | +283% |
| Nemotron-3-Nano-30B-A3B tok/s | 190.96 | 0 | n/a |
| Qwen3-4B-Instruct-2507 tok/s | 210.24 | 66.89 | +214% |
| Qwen3-Coder-Next tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 0 | n/a |
| Qwen1.5-0.5B tok/s | 774.78 | 328.29 | +136% |
| Qwen2-1.5B tok/s | 416.57 | 147.68 | +182% |
| Uncensored tok/s | 74.31 | 18.94 | +292% |
| Qwen2.5-Coder-0.5B tok/s | 715.81 | 294.96 | +143% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 142.97 | 37.31 | +283% |
| Qwen3-Coder-Next tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 0 | 0 | n/a |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 85.85 | 16.86 | +409% |
| Gemma 3 12B (Q3_K_M) tok/s | 93.05 | 19.79 | +370% |
| Phi-4 14B (Q3_K_M) tok/s | 89.18 | 16.04 | +456% |
| Mistral Small 24B (Q3_K_M) tok/s | 58.21 | 10.13 | +475% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 41.06 | 0 | n/a |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA L40S is faster on 4 of 4; NVIDIA T4 on 0.
| Workflow | NVIDIA L40S | NVIDIA T4 | Difference |
|---|---|---|---|
| 50-image depth pass | 6 s | 13 s | NVIDIA L40S 2.18x faster |
| 500-image masking run | 2.2 min | 14.4 min | NVIDIA L40S 6.69x faster |
| 200-product catalogue cutout | 6.2 min | 20.3 min | NVIDIA L40S 3.26x faster |
| Full codebase review | 13.5 min | 50.7 min | NVIDIA L40S 3.77x faster |
Specifications compared
| NVIDIA L40S | NVIDIA T4 | |
|---|---|---|
| VRAM | 48GB | 16GB |
| Architecture | Ada Lovelace | Turing |
| Memory bandwidth | 864 GB/s | 320 GB/s |
| Boost clock | 2,520 MHz | 1,590 MHz |
| TDP | 350 W | 70 W |
| Launch MSRP | $7,500 | $2,299 |
| Release | 2023-08-08 | 2018-09-13 |
FAQ
- Which is better for ai & machine learning: NVIDIA L40S or NVIDIA T4?
- NVIDIA L40S performs better for ai & machine learning, winning 145 of 153 benchmarks in our suite with an average 303.7% advantage.
- What are the main hardware differences between NVIDIA L40S and NVIDIA T4?
- NVIDIA L40S has 48GB VRAM and a 350W TDP, while NVIDIA T4 has 16GB VRAM and a 70W TDP.
- Does VRAM matter more than speed between NVIDIA L40S and NVIDIA T4?
- For AI, yes, NVIDIA L40S fits 36 more of our 12 workloads than NVIDIA T4. A model that exceeds VRAM doesn't run slower, it doesn't run at all, so the card that fits the model wins that workload outright.
- Where is the biggest performance difference between NVIDIA L40S and NVIDIA T4?
- Sana 1.6B: NVIDIA L40S leads by roughly 4396% (50.81 vs 1.13 images/min) in our testing.
NVIDIA L40S full review · NVIDIA T4 full review · All AI & Machine Learning rankings