NVIDIA A10G vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA L40S wins 149 of 159 benchmarks, averaging 72.4% 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 260% (14.13 vs 50.81 images/min); the closest fight is Qwen2.5 1.5B + 0.5B draft (3% apart); VRAM decides part of this one: NVIDIA L40S runs 152 of our 12 AI workloads while the other card runs 137, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA A10G | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 129.87 | 212.05 | -39% |
| Llama-3.1-8B tok/s | 86.62 | 135.51 | -36% |
| Qwen2.5-Coder-14B tok/s | 47.07 | 74.6 | -37% |
| Qwen3-32B tok/s | 22.04 | 34.44 | -36% |
| Llama 3.3 70B tok/s | 0 | 16.49 | n/a |
| Stable Diffusion XL images/min | 6.38 | 17.02 | -63% |
| Z-Image Turbo images/min | 3.45 | 8.325 | -59% |
| 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 | 0.18 | 0.49 | -63% |
| Llama 3.2 1B tok/s | 400.6 | 618.87 | -35% |
| Qwen3 0.6B tok/s | 451.38 | 655.75 | -31% |
| Qwen3 1.7B tok/s | 267.08 | 405.52 | -34% |
| DeepSeek-R1 Distill 1.5B tok/s | 265.75 | 426.38 | -38% |
| Gemma 3 4B tok/s | 124.85 | 198.87 | -37% |
| Llama 3.2 3B tok/s | 170.47 | 270.85 | -37% |
| Mistral 7B v0.3 tok/s | 92.82 | 144.74 | -36% |
| Phi-4 Mini 3.8B tok/s | 144.42 | 228.93 | -37% |
| Qwen2.5-Coder 7B tok/s | 90.05 | 143.93 | -37% |
| SmolLM3 3B tok/s | 172.15 | 272.88 | -37% |
| DeepSeek-R1 Distill Llama 8B tok/s | 86.83 | 135.55 | -36% |
| DeepSeek-R1 Distill 14B tok/s | 46.96 | 74.6 | -37% |
| DeepSeek-R1 Distill 7B tok/s | 89.79 | 143.97 | -38% |
| Gemma 3 12B tok/s | 52.07 | 80.92 | -36% |
| Gemma 4 12B tok/s | 52.67 | 80.82 | -35% |
| Mistral Small 24B tok/s | 30.94 | 48.43 | -36% |
| Phi-4 14B tok/s | 48.18 | 74.98 | -36% |
| Qwen3 14B tok/s | 48.12 | 75.12 | -36% |
| Qwen3 30B A3B tok/s | 140.12 | 213.66 | -34% |
| Qwen3 8B tok/s | 84.09 | 130.27 | -35% |
| TRELLIS Image-to-3D assets/hour | 293.6 | 776.1 | -62% |
| Codestral 22B tok/s | 32.09 | 50.06 | -36% |
| DeepSeek-R1 Distill 32B tok/s | 21.85 | 34.6 | -37% |
| Devstral Small 24B tok/s | 30.98 | 48.43 | -36% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 21.12 | 33.08 | -36% |
| Dolphin Mistral 24B Venice tok/s | 30.95 | 48.43 | -36% |
| Dolphin X1 8B tok/s | 86.61 | 135.48 | -36% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 86.54 | 135.55 | -36% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 30.9 | 48.4 | -36% |
| Gemma 3 27B tok/s | 24.8 | 38.48 | -36% |
| Qwen2.5-Coder 32B tok/s | 21.92 | 34.59 | -37% |
| Qwen3-Coder 30B A3B tok/s | 142.77 | 219.04 | -35% |
| QwQ 32B tok/s | 21.87 | 34.59 | -37% |
| StarCoder2 15B tok/s | 42.19 | 66.88 | -37% |
| Z-Image Turbo (1024px) images/min | 6.32 | 18.2 | -65% |
| BiRefNet images/min | 432.33 | 863.81 | -50% |
| Depth Anything V2 Large images/min | 740.74 | 959.33 | -23% |
| Depth Anything V2 Small images/min | 941.18 | 970.61 | -3% |
| SAM ViT-Base images/min | 398.72 | 998.79 | -60% |
| SAM ViT-Huge images/min | 84.22 | 238.08 | -65% |
| Swin2SR 4x Upscaler images/min | 22.31 | 34.14 | -35% |
| Kokoro TTS 82M x realtime | 101.11 | 243.99 | -59% |
| MusicGen Small x realtime | 1.2 | 2.43 | -51% |
| Whisper large-v3 x realtime | 86.07 | 193.2 | -55% |
| Florence-2 Base images/min | 210.79 | 184.2 | +14% |
| Florence-2 Large images/min | 117.94 | 106.54 | +11% |
| Dolphin X1 Trinity Nano 6B tok/s | 171.23 | 246.45 | -31% |
| gpt-oss-20b tok/s | 144.61 | 233.82 | -38% |
| Olmo-3.1-32B-Think tok/s | 22.13 | 34.11 | -35% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 31 | 48.43 | -36% |
| DeepSeek-Coder-V2-Lite tok/s | 160.13 | 257.64 | -38% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 83.22 | 130.23 | -36% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 47 | 74.63 | -37% |
| gemma-2-2b-it-abliterated tok/s | 174.56 | 278.14 | -37% |
| gemma-2-2b tok/s | 175.03 | 278.1 | -37% |
| gemma-2-9b tok/s | 57.37 | 88.87 | -35% |
| gemma-3-1b tok/s | 284.78 | 426.49 | -33% |
| gemma-3-270m tok/s | 597.49 | 822.52 | -27% |
| GLM-4.7-Flash tok/s | 103.24 | 157.89 | -35% |
| SmolLM3-3B tok/s | 171.66 | 272.96 | -37% |
| KAT-Coder-V2.5-Dev tok/s | 113.6 | 173.06 | -34% |
| L3-8B-Stheno-v3.2 tok/s | 86.8 | 135.53 | -36% |
| LFM2.5-1.2B tok/s | 429.96 | 661.54 | -35% |
| Llama-2-7B tok/s | 96.69 | 150.35 | -36% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 167.63 | 271.01 | -38% |
| Meta-Llama-3.1-8B tok/s | 86.69 | 135.46 | -36% |
| Phi-4-mini tok/s | 144.27 | 229.28 | -37% |
| Mistral-7B-Instruct-v0.1 tok/s | 92.45 | 144.82 | -36% |
| Mistral-7B-Instruct-v0.2 tok/s | 92.74 | 144.82 | -36% |
| Mistral-7B-Instruct-v0.3 tok/s | 92.45 | 144.68 | -36% |
| Ornith-1.0-35B tok/s | 104.9 | 157.39 | -33% |
| Ornith-1.0-9B tok/s | 73.78 | 114.57 | -36% |
| Phi-3.5-mini tok/s | 145.97 | 229.06 | -36% |
| Qwen-AgentWorld-35B-A3B tok/s | 104.36 | 156.98 | -34% |
| Qwen3-0.6B tok/s | 446.74 | 656.26 | -32% |
| Qwen3-1.7B tok/s | 262.49 | 405.3 | -35% |
| Qwen3-14B tok/s | 47.8 | 75.01 | -36% |
| Qwen3-8B tok/s | 83.04 | 130.28 | -36% |
| Qwen2.5-0.5B tok/s | 508.78 | 738.06 | -31% |
| Qwen2.5-1.5B tok/s | 265.7 | 427.23 | -38% |
| Qwen2.5-14B tok/s | 47.02 | 74.62 | -37% |
| Qwen2.5-32B tok/s | 21.98 | 34.59 | -36% |
| Qwen2.5-3B tok/s | 168.23 | 270.04 | -38% |
| Qwen2.5-7B tok/s | 89.68 | 143.94 | -38% |
| Qwen2.5-Coder-1.5B tok/s | 264.79 | 427.12 | -38% |
| Qwen2.5-Coder-3B tok/s | 168 | 269.9 | -38% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 97.48 | 146.51 | -33% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 83.26 | 130.25 | -36% |
| LFM2.5-8B-A1B tok/s | 268.04 | 394.48 | -32% |
| Mistral-Nemo-Instruct-2407 tok/s | 57 | 89.72 | -36% |
| Hermes-4-14B tok/s | 47.8 | 75.1 | -36% |
| phi-2 tok/s | 172.45 | 283.26 | -39% |
| Qwen3-30B-A3B tok/s | 140.14 | 214.05 | -35% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 46.91 | 74.63 | -37% |
| Qwen3-4B-Instruct-2507 tok/s | 129.9 | 212.05 | -39% |
| Qwen3-4B-Thinking-2507 tok/s | 129.99 | 211.88 | -39% |
| SmolLM2-135M tok/s | 680.37 | 916.56 | -26% |
| Cydonia-24B-v4.3 tok/s | 31.04 | 48.43 | -36% |
| CogVideoX-5B I2V clips/min | 0 | 0.383 | n/a |
| Qwen2.5 1.5B LoRA train tok/s | 4214.4 | 12120.1 | -65% |
| Qwen2.5 7B LoRA train tok/s | 1239 | 3736 | -67% |
| SmolLM2 1.7B LoRA train tok/s | 3668.3 | 12052.1 | -70% |
| TinyLlama 1.1B LoRA train tok/s | 5499.6 | 16408.1 | -66% |
| 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 | 2702.5 | 4686.8 | -42% |
| Qwen2.5 7B served serve tok/s | 858.7 | 1406.8 | -39% |
| SmolLM2 1.7B served serve tok/s | 2242.7 | 3525.9 | -36% |
| TinyLlama 1.1B served serve tok/s | 3558.2 | 6278.4 | -43% |
| AuraFlow v0.3 images/min | 1.39 | 3.58 | -61% |
| FLUX.1 Schnell images/min | 0 | 26.24 | n/a |
| PixArt-Sigma XL images/min | 9.93 | 26.95 | -63% |
| Sana 1.6B images/min | 14.13 | 50.81 | -72% |
| Stable Diffusion 3.5 Large images/min | 0 | 4.54 | n/a |
| Stable Diffusion 3.5 Medium images/min | 4.11 | 11.58 | -65% |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.919 | 0.946 | -3% |
| AI21-Jamba-Reasoning-3B tok/s | 166.29 | 258.41 | -36% |
| Codestral 22B (Q3_K_M) tok/s | 28.01 | 59.53 | -53% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 87.84 | 134.78 | -35% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 0 | 16.49 | n/a |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 41.72 | 85.85 | -51% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 22.14 | 34.43 | -36% |
| dolphin-2.9-llama3-8b tok/s | 87.98 | 134.94 | -35% |
| Gemma 3 12B (Q3_K_M) tok/s | 46.29 | 93.05 | -50% |
| Hermes-3-Llama-3.2-3B tok/s | 169.23 | 265.87 | -36% |
| 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 |
| Mistral Small 24B (Q3_K_M) tok/s | 26.8 | 58.21 | -54% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 57.52 | 88.85 | -35% |
| Nemotron-3-Nano-30B-A3B tok/s | 0 | 190.96 | n/a |
| Phi-4 14B (Q3_K_M) tok/s | 43.56 | 89.18 | -51% |
| Qwen3-4B-Instruct-2507 tok/s | 130.83 | 210.24 | -38% |
| 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 | 542.19 | 774.78 | -30% |
| Qwen2-1.5B tok/s | 265.45 | 416.57 | -36% |
| Uncensored tok/s | 47.45 | 74.31 | -36% |
| Qwen2.5-Coder-0.5B tok/s | 513.36 | 715.81 | -28% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 18.78 | 41.06 | -54% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 91.41 | 142.97 | -36% |
| 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 |
| Qwen3 30B A3B (Q3_K_M) tok/s | 120.38 | 224.95 | -46% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA A10G is faster on 0 of 8; NVIDIA L40S on 8.
| Workflow | NVIDIA A10G | NVIDIA L40S | Difference |
|---|---|---|---|
| 50-image depth pass | 7 s | 6 s | NVIDIA L40S 1.08x faster |
| 30-minute podcast pass | 77 s | 75 s | NVIDIA L40S 1.02x faster |
| 500-image masking run | 6 min | 2.2 min | NVIDIA L40S 2.79x faster |
| 20-asset 3D game kit | 7.6 min | 2.9 min | NVIDIA L40S 2.60x faster |
| 24-frame storyboard | 9 min | 4.1 min | NVIDIA L40S 2.21x faster |
| 200-product catalogue cutout | 9.6 min | 6.2 min | NVIDIA L40S 1.54x faster |
| Full codebase review | 21.3 min | 13.5 min | NVIDIA L40S 1.58x faster |
| 10 short social clips | 50.7 min | 19.1 min | NVIDIA L40S 2.66x faster |
Specifications compared
| NVIDIA A10G | NVIDIA L40S | |
|---|---|---|
| VRAM | 24GB | 48GB |
| Architecture | Ampere | Ada Lovelace |
| Memory bandwidth | 600 GB/s | 864 GB/s |
| Boost clock | 1,710 MHz | 2,520 MHz |
| TDP | 150 W | 350 W |
| Launch MSRP | $2,800 | $7,500 |
| Release | 2021-11-01 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA A10G or NVIDIA L40S?
- NVIDIA L40S performs better for ai & machine learning, winning 149 of 159 benchmarks in our suite with an average 72.4% advantage.
- What are the main hardware differences between NVIDIA A10G and NVIDIA L40S?
- NVIDIA A10G has 24GB VRAM and a 150W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA A10G and NVIDIA L40S?
- For AI, yes, NVIDIA L40S fits 15 more of our 12 workloads than NVIDIA A10G. 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 A10G and NVIDIA L40S?
- Sana 1.6B: NVIDIA L40S leads by roughly 260% (14.13 vs 50.81 images/min) in our testing.
NVIDIA A10G full review · NVIDIA L40S full review · All AI & Machine Learning rankings