NVIDIA A100 40GB SXM4 vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA L40S wins 74 of 149 benchmarks, averaging 10.7% faster.
NVIDIA A100 40GB SXM4's numbers are anchored estimates calibrated against our measured cards, pending first-party measurement. Treat small gaps as ties.
What the numbers say
The gap is widest in MusicGen Small, where NVIDIA L40S leads by 125% (1.08 vs 2.43 x realtime); the closest fight is Nemotron-3-Nano-30B-A3B (1% apart); VRAM decides part of this one: NVIDIA L40S runs 152 of our 12 AI workloads while the other card runs 135, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA A100 40GB SXM4 | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 193.14 | 212.05 | -9% |
| Llama-3.1-8B tok/s | 157.26 | 135.51 | +16% |
| Qwen2.5-Coder-14B tok/s | 86.2 | 74.6 | +16% |
| Qwen3-32B tok/s | 43.8 | 34.44 | +27% |
| Llama-3.3-70B tok/s | 0 | 16.49 | n/a |
| Stable Diffusion XL images/min | 18.576 | 17.02 | +9% |
| Z-Image Turbo images/min | 9.975 | 8.325 | +20% |
| FLUX.1 dev images/min | 4.264 | 3.836 | +11% |
| FLUX.1 Kontext dev images/min | 1.993 | 1.693 | +18% |
| Qwen-Image-Edit images/min | 0 | 1.06 | n/a |
| LTX-Video (distilled) frames/s | 8.95 | 8.01 | +12% |
| Wan 2.2 5B (720p) frames/s | 0.66 | 0.49 | +35% |
| DeepSeek-R1 Distill Llama 8B tok/s | 157.16 | 135.55 | +16% |
| DeepSeek-R1 Distill 1.5B tok/s | 320.78 | 426.38 | -25% |
| DeepSeek-R1 Distill 14B tok/s | 86.39 | 74.6 | +16% |
| DeepSeek-R1 Distill 7B tok/s | 159.09 | 143.97 | +11% |
| Gemma 3 12B tok/s | 88.97 | 80.92 | +10% |
| Gemma 3 4B tok/s | 174.14 | 198.87 | -12% |
| Gemma 4 12B tok/s | 87.14 | 80.82 | +8% |
| Llama 3.2 1B tok/s | 532.08 | 618.87 | -14% |
| Llama 3.2 3B tok/s | 261.24 | 270.85 | -4% |
| Mistral 7B v0.3 tok/s | 165.9 | 144.74 | +15% |
| Mistral Small 24B tok/s | 62.42 | 48.43 | +29% |
| Phi-4 14B tok/s | 98.65 | 74.98 | +32% |
| Phi-4 Mini 3.8B tok/s | 236.71 | 228.93 | +3% |
| Qwen2.5-Coder 7B tok/s | 159.9 | 143.93 | +11% |
| Qwen3 0.6B tok/s | 417.88 | 655.75 | -36% |
| Qwen3 1.7B tok/s | 340.76 | 405.52 | -16% |
| Qwen3 14B tok/s | 90.01 | 75.12 | +20% |
| Qwen3 30B A3B tok/s | 169.1 | 213.66 | -21% |
| Qwen3 8B tok/s | 146.21 | 130.27 | +12% |
| SmolLM3 3B tok/s | 246.25 | 272.88 | -10% |
| Codestral 22B tok/s | 63.74 | 50.06 | +27% |
| DeepSeek-R1 Distill 32B tok/s | 43.13 | 34.6 | +25% |
| Devstral Small 24B tok/s | 62.4 | 48.43 | +29% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 42.97 | 33.08 | +30% |
| Dolphin Mistral 24B Venice tok/s | 62.46 | 48.43 | +29% |
| Dolphin X1 8B tok/s | 156.3 | 135.48 | +15% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 156.55 | 135.55 | +15% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 62.4 | 48.4 | +29% |
| Gemma 3 27B tok/s | 47.47 | 38.48 | +23% |
| Qwen2.5-Coder 32B tok/s | 43.14 | 34.59 | +25% |
| Qwen3-Coder 30B A3B tok/s | 173.47 | 219.04 | -21% |
| QwQ 32B tok/s | 43.02 | 34.59 | +24% |
| StarCoder2 15B tok/s | 80.25 | 66.88 | +20% |
| FLUX.1 Schnell images/min | 28.44 | 26.24 | +8% |
| Z-Image Turbo (1024px) images/min | 18.7 | 18.2 | +3% |
| BiRefNet images/min | 611.56 | 863.81 | -29% |
| Depth Anything V2 Large images/min | 536.74 | 959.33 | -44% |
| Depth Anything V2 Small images/min | 613.42 | 970.61 | -37% |
| SAM ViT-Base images/min | 696.77 | 998.79 | -30% |
| SAM ViT-Huge images/min | 146.81 | 238.08 | -38% |
| Swin2SR 4x Upscaler images/min | 17.25 | 34.14 | -49% |
| Kokoro TTS 82M x realtime | 142.03 | 243.99 | -42% |
| MusicGen Small x realtime | 1.08 | 2.43 | -56% |
| Whisper large-v3 x realtime | 102.47 | 193.2 | -47% |
| Dolphin X1 Trinity Nano 6B tok/s | 153.27 | 246.45 | -38% |
| gpt-oss-20b tok/s | 205.1 | 233.82 | -12% |
| Olmo-3.1-32B-Think tok/s | 43.67 | 34.11 | +28% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 62.56 | 48.43 | +29% |
| DeepSeek-Coder-V2-Lite tok/s | 205.42 | 257.64 | -20% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 145.84 | 130.23 | +12% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 86.48 | 74.63 | +16% |
| Qwen2.5 1.5B LoRA train tok/s | 6349.7 | 12120.1 | -48% |
| Qwen2.5 7B LoRA train tok/s | 3670.8 | 3736 | -2% |
| SmolLM2 1.7B LoRA train tok/s | 6938.8 | 12052.1 | -42% |
| TinyLlama 1.1B LoRA train tok/s | 7471 | 16408.1 | -54% |
| gemma-2-2b-it-abliterated tok/s | 232.27 | 278.14 | -16% |
| gemma-2-2b tok/s | 234.68 | 278.1 | -16% |
| gemma-2-9b tok/s | 101.97 | 88.87 | +15% |
| gemma-3-1b tok/s | 310.69 | 426.49 | -27% |
| gemma-3-270m tok/s | 574.24 | 822.52 | -30% |
| GLM-4.7-Flash tok/s | 119.1 | 157.89 | -25% |
| SmolLM3-3B tok/s | 247.66 | 272.96 | -9% |
| KAT-Coder-V2.5-Dev tok/s | 137.31 | 173.06 | -21% |
| LFM2.5-1.2B tok/s | 572.69 | 661.54 | -13% |
| Llama-2-7B tok/s | 172.53 | 150.35 | +15% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 260.83 | 271.01 | -4% |
| Meta-Llama-3.1-8B tok/s | 156.68 | 135.46 | +16% |
| Phi-4-mini tok/s | 237.38 | 229.28 | +4% |
| Mistral-7B-Instruct-v0.1 tok/s | 166.39 | 144.82 | +15% |
| Mistral-7B-Instruct-v0.2 tok/s | 166.8 | 144.82 | +15% |
| Mistral-7B-Instruct-v0.3 tok/s | 167.06 | 144.68 | +15% |
| Ornith-1.0-35B tok/s | 131.77 | 157.39 | -16% |
| Ornith-1.0-9B tok/s | 126.79 | 114.57 | +11% |
| Phi-3.5-mini tok/s | 222.53 | 229.06 | -3% |
| Qwen-AgentWorld-35B-A3B tok/s | 131.16 | 156.98 | -16% |
| Qwen3-0.6B tok/s | 408.39 | 656.26 | -38% |
| Qwen3-1.7B tok/s | 341.85 | 405.3 | -16% |
| Qwen3-14B tok/s | 90.07 | 75.01 | +20% |
| Qwen3-8B tok/s | 145.96 | 130.28 | +12% |
| Qwen2.5-0.5B tok/s | 517.09 | 738.06 | -30% |
| Qwen2.5-1.5B tok/s | 315.59 | 427.23 | -26% |
| Qwen2.5-14B tok/s | 86.07 | 74.62 | +15% |
| Qwen2.5-32B tok/s | 43.23 | 34.59 | +25% |
| Qwen2.5-3B tok/s | 240.48 | 270.04 | -11% |
| Qwen2.5-7B tok/s | 159.81 | 143.94 | +11% |
| Qwen2.5-Coder-1.5B tok/s | 322.96 | 427.12 | -24% |
| Qwen2.5-Coder-3B tok/s | 241.51 | 269.9 | -11% |
| AI21-Jamba-Reasoning-3B tok/s | 231.15 | 258.41 | -11% |
| Codestral 22B (Q3_K_M) tok/s | 48.37 | 59.53 | -19% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 156.62 | 134.78 | +16% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 65.33 | 85.85 | -24% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 43.07 | 34.43 | +25% |
| dolphin-2.9-llama3-8b tok/s | 153.34 | 134.94 | +14% |
| Gemma 3 12B (Q3_K_M) tok/s | 69.7 | 93.05 | -25% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 109.1 | 146.51 | -26% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 146.8 | 130.25 | +13% |
| Hermes-3-Llama-3.2-3B tok/s | 261 | 265.87 | -2% |
| L3-8B-Stheno-v3.2 tok/s | 156.55 | 135.53 | +16% |
| LFM2.5-8B-A1B tok/s | 351.47 | 394.48 | -11% |
| Mistral-Nemo-Instruct-2407 tok/s | 106.33 | 89.72 | +19% |
| Mistral Small 24B (Q3_K_M) tok/s | 45.34 | 58.21 | -22% |
| NemoMix-Unleashed-12B tok/s | 105.05 | 88.85 | +18% |
| Nemotron-3-Nano-30B-A3B tok/s | 192.44 | 190.96 | +1% |
| Hermes-4-14B tok/s | 89.97 | 75.1 | +20% |
| phi-2 tok/s | 228.1 | 283.26 | -19% |
| Phi-4 14B (Q3_K_M) tok/s | 80.71 | 89.18 | -9% |
| Qwen3-30B-A3B tok/s | 167.1 | 214.05 | -22% |
| Qwen3-4B-Instruct-2507 tok/s | 192.89 | 210.24 | -8% |
| Qwen1.5-0.5B tok/s | 497.73 | 774.78 | -36% |
| Qwen2-1.5B tok/s | 322.09 | 416.57 | -23% |
| Uncensored tok/s | 85.82 | 74.31 | +15% |
| Qwen2.5-Coder-0.5B tok/s | 542.36 | 715.81 | -24% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 86.34 | 74.63 | +16% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 31.04 | 41.06 | -24% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 159.76 | 142.97 | +12% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 142.78 | 224.95 | -37% |
| Qwen3-4B-Instruct-2507 tok/s | 193.39 | 212.05 | -9% |
| Qwen3-4B-Thinking-2507 tok/s | 189.89 | 211.88 | -10% |
| SmolLM2-135M tok/s | 544.06 | 916.56 | -41% |
| Cydonia-24B-v4.3 tok/s | 62.41 | 48.43 | +29% |
| Qwen2.5 1.5B served serve tok/s | 4111.1 | 4686.8 | -12% |
| Qwen2.5 7B served serve tok/s | 1769.6 | 1406.8 | +26% |
| SmolLM2 1.7B served serve tok/s | 3865.2 | 3525.9 | +10% |
| TinyLlama 1.1B served serve tok/s | 3221.3 | 6278.4 | -49% |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.694 | 0.946 | -27% |
| 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 |
| 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 |
| DeepSeek-R1-Distill-Llama-70B tok/s | 0 | 16.49 | n/a |
| Hermes-4-70B tok/s | 0 | 16.45 | 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 |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA A100 40GB SXM4 is faster on 12 of 15; NVIDIA L40S on 3.
| Workflow | NVIDIA A100 40GB SXM4 | NVIDIA L40S | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 10 s | 6 s | NVIDIA L40S 1.58x faster | $0.003 vs $0.001 |
| 30-minute podcast pass | 59 s | 75 s | NVIDIA A100 40GB SXM4 1.27x faster | $0.016 vs $0.017 |
| 24-frame storyboard | 3 min | 4.1 min | NVIDIA A100 40GB SXM4 1.33x faster | $0.051 vs $0.053 |
| 500-image masking run | 3.5 min | 2.2 min | NVIDIA L40S 1.62x faster | $0.058 vs $0.028 |
| 60-second AI short film | 4 min | 4.9 min | NVIDIA A100 40GB SXM4 1.25x faster | $0.066 vs $0.065 |
| 60-second AI short film, narrated | 4.2 min | 5.1 min | NVIDIA A100 40GB SXM4 1.22x faster | $0.070 vs $0.067 |
| 6-panel comic page | 4.8 min | 6.5 min | NVIDIA A100 40GB SXM4 1.36x faster | $0.080 vs $0.086 |
| Character sheet, 12 poses | 6.3 min | 8.4 min | NVIDIA A100 40GB SXM4 1.34x faster | $0.104 vs $0.110 |
| Full codebase review | 11.7 min | 13.5 min | NVIDIA A100 40GB SXM4 1.15x faster | $0.195 vs $0.177 |
| 200-product catalogue cutout | 12.2 min | 6.2 min | NVIDIA L40S 1.95x faster | $0.203 vs $0.082 |
| 10 short social clips | 13.7 min | 19.1 min | NVIDIA A100 40GB SXM4 1.39x faster | $0.228 vs $0.251 |
| 40-product photo shoot | 22.2 min | 26.7 min | NVIDIA A100 40GB SXM4 1.20x faster | $0.370 vs $0.352 |
| 40-product shoot, start to finish | 24.8 min | 28 min | NVIDIA A100 40GB SXM4 1.13x faster | $0.414 vs $0.369 |
| 100-photo restoration batch | 50.2 min | 59.6 min | NVIDIA A100 40GB SXM4 1.19x faster | $0.836 vs $0.784 |
| 100-photo restore and enlarge | 56 min | 62.5 min | NVIDIA A100 40GB SXM4 1.12x faster | $0.934 vs $0.823 |
Renting by the hour, NVIDIA L40S finishes 10 of 15 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA A100 40GB SXM4 | $1.000 |
| NVIDIA L40S | $0.790 |
NVIDIA L40S is 1.27x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA A100 40GB SXM4 | NVIDIA L40S | |
|---|---|---|
| VRAM | 40GB | 48GB |
| Architecture | Ampere | Ada Lovelace |
| Memory bandwidth | 1555 GB/s | 864 GB/s |
| Boost clock | 1,410 MHz | 2,520 MHz |
| TDP | 400 W | 350 W |
| Launch MSRP | $12,000 | $7,500 |
| Release | 2020-05-14 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA A100 40GB SXM4 or NVIDIA L40S?
- NVIDIA L40S performs better for ai & machine learning, winning 74 of 149 benchmarks in our suite with an average 10.7% advantage.
- What are the main hardware differences between NVIDIA A100 40GB SXM4 and NVIDIA L40S?
- NVIDIA A100 40GB SXM4 has 40GB VRAM and a 400W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA A100 40GB SXM4 and NVIDIA L40S?
- For AI, yes, NVIDIA L40S fits 7 more of our 12 workloads than NVIDIA A100 40GB SXM4. 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 A100 40GB SXM4 and NVIDIA L40S?
- MusicGen Small: NVIDIA L40S leads by roughly 125% (1.08 vs 2.43 x realtime) in our testing.
NVIDIA A100 40GB SXM4 full review · NVIDIA L40S full review · All AI & Machine Learning rankings