NVIDIA L4 vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA L40S wins 150 of 159 benchmarks, averaging 159% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Z-Image Turbo (1024px), where NVIDIA L40S leads by 296% (4.6 vs 18.2 images/min); the closest fight is Qwen2.5 1.5B + 0.5B draft (2% apart); VRAM decides part of this one: NVIDIA L40S runs 152 of our 12 AI workloads while the other card runs 138, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA L4 | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 84 | 212.05 | -60% |
| Llama 3.1 8B tok/s | 50.45 | 135.51 | -63% |
| Qwen2.5-Coder 14B tok/s | 27.46 | 74.6 | -63% |
| Qwen3-32B tok/s | 12.42 | 34.44 | -64% |
| Llama 3.3 70B tok/s | 0 | 16.49 | n/a |
| Stable Diffusion XL images/min | 5.18 | 17.02 | -70% |
| Z-Image Turbo images/min | 2.475 | 8.325 | -70% |
| FLUX.1 dev images/min | 0 | 3.836 | n/a |
| FLUX.1 Kontext dev images/min | 0 | 1.693 | n/a |
| Qwen-Image-Edit images/min | 0 | 1.06 | n/a |
| Wan 2.2 5B (720p) frames/s | 0.15 | 0.49 | -69% |
| DeepSeek-R1 Distill Llama 8B tok/s | 50.39 | 135.55 | -63% |
| DeepSeek-R1 Distill 1.5B tok/s | 191.38 | 426.38 | -55% |
| DeepSeek-R1 Distill 14B tok/s | 27.39 | 74.6 | -63% |
| DeepSeek-R1 Distill 7B tok/s | 53.27 | 143.97 | -63% |
| Gemma 3 12B tok/s | 31.01 | 80.92 | -62% |
| Gemma 3 4B tok/s | 82.03 | 198.87 | -59% |
| Gemma 4 12B tok/s | 31.49 | 80.82 | -61% |
| Llama 3.2 1B tok/s | 259.48 | 618.87 | -58% |
| Llama 3.2 3B tok/s | 107.1 | 270.85 | -60% |
| Mistral 7B v0.3 tok/s | 54.02 | 144.74 | -63% |
| Mistral Small 24B tok/s | 17.33 | 48.43 | -64% |
| Phi-4 14B tok/s | 27.24 | 74.98 | -64% |
| Phi-4 Mini 3.8B tok/s | 88.48 | 228.93 | -61% |
| Qwen2.5-Coder 7B tok/s | 53.26 | 143.93 | -63% |
| Qwen3 0.6B tok/s | 357.79 | 655.75 | -45% |
| Qwen3 1.7B tok/s | 177.5 | 405.52 | -56% |
| Qwen3 14B tok/s | 27.61 | 75.12 | -63% |
| Qwen3 30B A3B tok/s | 96.15 | 213.66 | -55% |
| Qwen3 8B tok/s | 48.95 | 130.27 | -62% |
| SmolLM3 3B tok/s | 112.07 | 272.88 | -59% |
| TRELLIS Image-to-3D assets/hour | 284.8 | 776.1 | -63% |
| Codestral 22B tok/s | 18.14 | 50.06 | -64% |
| DeepSeek-R1 Distill 32B tok/s | 12.28 | 34.6 | -65% |
| Devstral Small 24B tok/s | 17.34 | 48.43 | -64% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 11.92 | 33.08 | -64% |
| Dolphin Mistral 24B Venice tok/s | 17.34 | 48.43 | -64% |
| Dolphin X1 8B tok/s | 50.12 | 135.48 | -63% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 50.24 | 135.55 | -63% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 17.32 | 48.4 | -64% |
| Gemma 3 27B tok/s | 14.18 | 38.48 | -63% |
| Qwen2.5-Coder 32B tok/s | 12.29 | 34.59 | -64% |
| Qwen3-Coder 30B A3B tok/s | 99.14 | 219.04 | -55% |
| QwQ 32B tok/s | 12.28 | 34.59 | -64% |
| StarCoder2 15B tok/s | 24.4 | 66.88 | -64% |
| Z-Image Turbo (1024px) images/min | 4.6 | 18.2 | -75% |
| BiRefNet images/min | 305.81 | 863.81 | -65% |
| Depth Anything V2 Large images/min | 770.37 | 959.33 | -20% |
| Depth Anything V2 Small images/min | 916.66 | 970.61 | -6% |
| Dolphin X1 Trinity Nano 6B tok/s | 166.6 | 246.45 | -32% |
| SAM ViT-Base images/min | 315.89 | 998.79 | -68% |
| SAM ViT-Huge images/min | 70.46 | 238.08 | -70% |
| Swin2SR 4x Upscaler images/min | 18.23 | 34.14 | -47% |
| Sana 1.6B images/min | 13.84 | 50.81 | -73% |
| FLUX.1 Schnell images/min | 0 | 26.24 | n/a |
| PixArt-Sigma XL images/min | 8.08 | 26.95 | -70% |
| Qwen2.5-1.5B-Instruct tok/s | 191.39 | 427.23 | -55% |
| AuraFlow v0.3 images/min | 1.02 | 3.58 | -72% |
| gpt-oss-20b tok/s | 91.78 | 233.82 | -61% |
| Kokoro TTS 82M x realtime | 97.35 | 243.99 | -60% |
| MusicGen Small x realtime | 1.01 | 2.43 | -58% |
| Phi-3.5-mini-instruct tok/s | 90.56 | 229.06 | -60% |
| Qwen2.5-3B-Instruct tok/s | 110.84 | 270.04 | -59% |
| Stable Diffusion 3.5 Large images/min | 0 | 4.54 | n/a |
| Stable Diffusion 3.5 Medium images/min | 2.96 | 11.58 | -74% |
| Whisper large-v3 x realtime | 70.11 | 193.2 | -64% |
| Florence-2 Base images/min | 174.65 | 184.2 | -5% |
| Florence-2 Large images/min | 93.77 | 106.54 | -12% |
| Qwen-AgentWorld-35B-A3B tok/s | 70.23 | 156.98 | -55% |
| Meta-Llama-3.1-8B-Instruct tok/s | 50.36 | 135.46 | -63% |
| Qwen2.5-7B-Instruct tok/s | 53.32 | 143.94 | -63% |
| Ornith-1.0-35B tok/s | 70.31 | 157.39 | -55% |
| Kwaipilot_KAT-Coder-V2.5-Dev tok/s | 79.01 | 173.06 | -54% |
| Qwen2.5-0.5B-Instruct tok/s | 399.4 | 738.06 | -46% |
| GLM-4.7-Flash tok/s | 75.03 | 157.89 | -52% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 49.07 | 130.23 | -62% |
| gemma-2-2b-it tok/s | 114.17 | 278.1 | -59% |
| gemma-3-1b-it tok/s | 216.27 | 426.49 | -49% |
| gemma-3-270m-it tok/s | 493.31 | 822.52 | -40% |
| Ornith-1.0-9B tok/s | 43.35 | 114.57 | -62% |
| Qwen_Qwen3-0.6B tok/s | 356.93 | 656.26 | -46% |
| Qwen2.5-14B-Instruct tok/s | 27.37 | 74.62 | -63% |
| CogVideoX-5B I2V clips/min | 0 | 0.383 | n/a |
| Qwen2.5 1.5B LoRA train tok/s | 3720.9 | 12120.1 | -69% |
| Qwen2.5 7B LoRA train tok/s | 1063.9 | 3736 | -72% |
| SmolLM2 1.7B LoRA train tok/s | 3240.2 | 12052.1 | -73% |
| TinyLlama 1.1B LoRA train tok/s | 4756 | 16408.1 | -71% |
| Stable Video Diffusion XT clips/min | 0 | 1.166 | n/a |
| Wan 2.2 TI2V-5B (image to video) clips/min | 0 | 1.606 | n/a |
| Qwen2.5 1.5B served serve tok/s | 1827 | 4686.8 | -61% |
| Qwen2.5 7B served serve tok/s | 469.9 | 1406.8 | -67% |
| SmolLM2 1.7B served serve tok/s | 1524.9 | 3525.9 | -57% |
| TinyLlama 1.1B served serve tok/s | 2584.3 | 6278.4 | -59% |
| AI21-Jamba-Reasoning-3B tok/s | 107.59 | 258.41 | -58% |
| Olmo-3.1-32B-Think tok/s | 12.27 | 34.11 | -64% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 17.2 | 48.43 | -64% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 48.82 | 134.78 | -64% |
| DeepSeek-Coder-V2-Lite tok/s | 110.11 | 257.64 | -57% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 27.87 | 85.85 | -68% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 12.26 | 34.43 | -64% |
| dolphin-2.9-llama3-8b tok/s | 50.04 | 134.94 | -63% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 27.3 | 74.63 | -63% |
| gemma-2-2b-it-abliterated tok/s | 113 | 278.14 | -59% |
| gemma-2-9b tok/s | 33.8 | 88.87 | -62% |
| Gemma 3 12B (Q3_K_M) tok/s | 32.29 | 93.05 | -65% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 70.87 | 146.51 | -52% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 48.72 | 130.25 | -63% |
| Hermes-3-Llama-3.2-3B tok/s | 106.14 | 265.87 | -60% |
| SmolLM3-3B tok/s | 110.77 | 272.96 | -59% |
| L3-8B-Stheno-v3.2 tok/s | 50.07 | 135.53 | -63% |
| LFM2.5-1.2B tok/s | 278.53 | 661.54 | -58% |
| LFM2.5-8B-A1B tok/s | 170.2 | 394.48 | -57% |
| Llama-2-7B tok/s | 56.06 | 150.35 | -63% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 101.16 | 271.01 | -63% |
| Phi-4-mini tok/s | 87.21 | 229.28 | -62% |
| Mistral-7B-Instruct-v0.1 tok/s | 53.6 | 144.82 | -63% |
| Mistral-7B-Instruct-v0.2 tok/s | 53.31 | 144.82 | -63% |
| Mistral-7B-Instruct-v0.3 tok/s | 53.66 | 144.68 | -63% |
| Mistral-Nemo-Instruct-2407 tok/s | 33.01 | 89.72 | -63% |
| NemoMix-Unleashed-12B tok/s | 33.02 | 88.85 | -63% |
| Hermes-4-14B tok/s | 27.48 | 75.1 | -63% |
| phi-2 tok/s | 114.99 | 283.26 | -59% |
| Phi-4 14B (Q3_K_M) tok/s | 27.86 | 89.18 | -69% |
| Qwen3-1.7B tok/s | 176.08 | 405.3 | -57% |
| Qwen3-14B tok/s | 27.48 | 75.01 | -63% |
| Qwen3-30B-A3B tok/s | 95.02 | 214.05 | -56% |
| Qwen3-4B-Instruct-2507 tok/s | 83.03 | 210.24 | -61% |
| Qwen3-8B tok/s | 48.34 | 130.28 | -63% |
| Qwen1.5-0.5B tok/s | 431.41 | 774.78 | -44% |
| Qwen2-1.5B tok/s | 189.18 | 416.57 | -55% |
| Uncensored tok/s | 27.29 | 74.31 | -63% |
| Qwen2.5-32B tok/s | 12.24 | 34.59 | -65% |
| Qwen2.5-Coder-0.5B tok/s | 397.1 | 715.81 | -45% |
| Qwen2.5-Coder-1.5B tok/s | 184.6 | 427.12 | -57% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 27.29 | 74.63 | -63% |
| Qwen2.5-Coder-3B tok/s | 109.42 | 269.9 | -59% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 53.02 | 142.97 | -63% |
| Qwen3-4B-Instruct-2507 tok/s | 83.13 | 212.05 | -61% |
| Qwen3-4B-Thinking-2507 tok/s | 83.12 | 211.88 | -61% |
| SmolLM2-135M tok/s | 652.15 | 916.56 | -29% |
| Cydonia-24B-v4.3 tok/s | 17.3 | 48.43 | -64% |
| Codestral 22B (Q3_K_M) tok/s | 17.96 | 59.53 | -70% |
| Mistral Small 24B (Q3_K_M) tok/s | 16.62 | 58.21 | -71% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 12.05 | 41.06 | -71% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 94.03 | 224.95 | -58% |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.968 | 0.946 | +2% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 0 | 16.49 | n/a |
| Hermes-4-70B tok/s | 0 | 16.45 | n/a |
| Qwen3-Coder-Next-abliterated tok/s | 0 | 0 | n/a |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| Llama-3.3-70B-Instruct-abliterated tok/s | 0 | 16.49 | n/a |
| Meta-Llama-3.1-70B tok/s | 0 | 16.49 | n/a |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| Nemotron-3-Nano-30B-A3B tok/s | 0 | 190.96 | n/a |
| Qwen3-Coder-Next tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 0 | n/a |
| 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 |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA L4 is faster on 0 of 8; NVIDIA L40S on 8.
| Workflow | NVIDIA L4 | NVIDIA L40S | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 17 s | 6 s | NVIDIA L40S 2.80x faster | $0.002 vs $0.001 |
| 30-minute podcast pass | 2 min | 75 s | NVIDIA L40S 1.60x faster | $0.015 vs $0.017 |
| 500-image masking run | 7.5 min | 2.2 min | NVIDIA L40S 3.46x faster | $0.055 vs $0.028 |
| 20-asset 3D game kit | 8.3 min | 2.9 min | NVIDIA L40S 2.85x faster | $0.061 vs $0.039 |
| 200-product catalogue cutout | 12.1 min | 6.2 min | NVIDIA L40S 1.94x faster | $0.089 vs $0.082 |
| 24-frame storyboard | 12.2 min | 4.1 min | NVIDIA L40S 3.01x faster | $0.089 vs $0.053 |
| Full codebase review | 36.4 min | 13.5 min | NVIDIA L40S 2.71x faster | $0.267 vs $0.177 |
| 10 short social clips | 60.3 min | 19.1 min | NVIDIA L40S 3.17x faster | $0.442 vs $0.251 |
Renting by the hour, NVIDIA L40S finishes 7 of 8 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA L4 | $0.440 |
| NVIDIA L40S | $0.790 |
NVIDIA L4 is 1.80x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA L4 | NVIDIA L40S | |
|---|---|---|
| VRAM | 24GB | 48GB |
| Architecture | Ada Lovelace | Ada Lovelace |
| Memory bandwidth | 300 GB/s | 864 GB/s |
| Boost clock | 2,040 MHz | 2,520 MHz |
| TDP | 72 W | 350 W |
| Launch MSRP | $2,500 | $7,500 |
| Release | 2023-03-21 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA L4 or NVIDIA L40S?
- NVIDIA L40S performs better for ai & machine learning, winning 150 of 159 benchmarks in our suite with an average 159% advantage.
- What are the main hardware differences between NVIDIA L4 and NVIDIA L40S?
- NVIDIA L4 has 24GB VRAM and a 72W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA L4 and NVIDIA L40S?
- For AI, yes, NVIDIA L40S fits 16 more of our 12 workloads than NVIDIA L4. 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 L4 and NVIDIA L40S?
- Z-Image Turbo (1024px): NVIDIA L40S leads by roughly 296% (4.6 vs 18.2 images/min) in our testing.
NVIDIA L4 full review · NVIDIA L40S full review · All AI & Machine Learning rankings