NVIDIA A100 40GB SXM4 vs NVIDIA L4, AI & Machine Learning Comparison
NVIDIA A100 40GB SXM4 wins 128 of 149 benchmarks, averaging 152.9% 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 Wan 2.2 5B (720p), where NVIDIA A100 40GB SXM4 leads by 340% (0.66 vs 0.15 frames/s); the closest fight is Swin2SR 4x Upscaler (6% apart); VRAM decides part of this one: NVIDIA A100 40GB SXM4 runs 138 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 L4 | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 193.14 | 84 | +130% |
| Llama-3.1-8B tok/s | 157.26 | 50.45 | +212% |
| Qwen2.5-Coder-14B tok/s | 86.2 | 27.46 | +214% |
| Qwen3-32B tok/s | 43.8 | 12.42 | +253% |
| Llama-3.3-70B tok/s | 0 | 0 | n/a |
| Stable Diffusion XL images/min | 18.576 | 5.18 | +259% |
| Z-Image Turbo images/min | 9.975 | 2.475 | +303% |
| FLUX.1 dev images/min | 4.264 | 0 | n/a |
| FLUX.1 Kontext dev images/min | 1.993 | 0 | n/a |
| Qwen-Image-Edit images/min | 0 | 0 | n/a |
| Wan 2.2 5B (720p) frames/s | 0.66 | 0.15 | +340% |
| DeepSeek-R1 Distill Llama 8B tok/s | 157.16 | 50.39 | +212% |
| DeepSeek-R1 Distill 1.5B tok/s | 320.78 | 191.38 | +68% |
| DeepSeek-R1 Distill 14B tok/s | 86.39 | 27.39 | +215% |
| DeepSeek-R1 Distill 7B tok/s | 159.09 | 53.27 | +199% |
| Gemma 3 12B tok/s | 88.97 | 31.01 | +187% |
| Gemma 3 4B tok/s | 174.14 | 82.03 | +112% |
| Gemma 4 12B tok/s | 87.14 | 31.49 | +177% |
| Llama 3.2 1B tok/s | 532.08 | 259.48 | +105% |
| Llama 3.2 3B tok/s | 261.24 | 107.1 | +144% |
| Mistral 7B v0.3 tok/s | 165.9 | 54.02 | +207% |
| Mistral Small 24B tok/s | 62.42 | 17.33 | +260% |
| Phi-4 14B tok/s | 98.65 | 27.24 | +262% |
| Phi-4 Mini 3.8B tok/s | 236.71 | 88.48 | +168% |
| Qwen2.5-Coder 7B tok/s | 159.9 | 53.26 | +200% |
| Qwen3 0.6B tok/s | 417.88 | 357.79 | +17% |
| Qwen3 1.7B tok/s | 340.76 | 177.5 | +92% |
| Qwen3 14B tok/s | 90.01 | 27.61 | +226% |
| Qwen3 30B A3B tok/s | 169.1 | 96.15 | +76% |
| Qwen3 8B tok/s | 146.21 | 48.95 | +199% |
| SmolLM3 3B tok/s | 246.25 | 112.07 | +120% |
| Codestral 22B tok/s | 63.74 | 18.14 | +251% |
| DeepSeek-R1 Distill 32B tok/s | 43.13 | 12.28 | +251% |
| Devstral Small 24B tok/s | 62.4 | 17.34 | +260% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 42.97 | 11.92 | +260% |
| Dolphin Mistral 24B Venice tok/s | 62.46 | 17.34 | +260% |
| Dolphin X1 8B tok/s | 156.3 | 50.12 | +212% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 156.55 | 50.24 | +212% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 62.4 | 17.32 | +260% |
| Gemma 3 27B tok/s | 47.47 | 14.18 | +235% |
| Qwen2.5-Coder 32B tok/s | 43.14 | 12.29 | +251% |
| Qwen3-Coder 30B A3B tok/s | 173.47 | 99.14 | +75% |
| QwQ 32B tok/s | 43.02 | 12.28 | +250% |
| StarCoder2 15B tok/s | 80.25 | 24.4 | +229% |
| FLUX.1 Schnell images/min | 28.44 | 0 | n/a |
| Z-Image Turbo (1024px) images/min | 18.7 | 4.6 | +307% |
| BiRefNet images/min | 611.56 | 305.81 | +100% |
| Depth Anything V2 Large images/min | 536.74 | 770.37 | -30% |
| Depth Anything V2 Small images/min | 613.42 | 916.66 | -33% |
| SAM ViT-Base images/min | 696.77 | 315.89 | +121% |
| SAM ViT-Huge images/min | 146.81 | 70.46 | +108% |
| Swin2SR 4x Upscaler images/min | 17.25 | 18.23 | -5% |
| Kokoro TTS 82M x realtime | 142.03 | 97.35 | +46% |
| MusicGen Small x realtime | 1.08 | 1.01 | +7% |
| Whisper large-v3 x realtime | 102.47 | 70.11 | +46% |
| Dolphin X1 Trinity Nano 6B tok/s | 153.27 | 166.6 | -8% |
| gpt-oss-20b tok/s | 205.1 | 91.78 | +123% |
| Olmo-3.1-32B-Think tok/s | 43.67 | 12.27 | +256% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 62.56 | 17.2 | +264% |
| DeepSeek-Coder-V2-Lite tok/s | 205.42 | 110.11 | +87% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 145.84 | 49.07 | +197% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 86.48 | 27.3 | +217% |
| Qwen2.5 1.5B LoRA train tok/s | 6349.7 | 3720.9 | +71% |
| Qwen2.5 7B LoRA train tok/s | 3670.8 | 1063.9 | +245% |
| SmolLM2 1.7B LoRA train tok/s | 6938.8 | 3240.2 | +114% |
| TinyLlama 1.1B LoRA train tok/s | 7471 | 4756 | +57% |
| gemma-2-2b-it-abliterated tok/s | 232.27 | 113 | +106% |
| gemma-2-2b tok/s | 234.68 | 114.17 | +106% |
| gemma-2-9b tok/s | 101.97 | 33.8 | +202% |
| gemma-3-1b tok/s | 310.69 | 216.27 | +44% |
| gemma-3-270m tok/s | 574.24 | 493.31 | +16% |
| GLM-4.7-Flash tok/s | 119.1 | 75.03 | +59% |
| SmolLM3-3B tok/s | 247.66 | 110.77 | +124% |
| KAT-Coder-V2.5-Dev tok/s | 137.31 | 79.01 | +74% |
| LFM2.5-1.2B tok/s | 572.69 | 278.53 | +106% |
| Llama-2-7B tok/s | 172.53 | 56.06 | +208% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 260.83 | 101.16 | +158% |
| Meta-Llama-3.1-8B tok/s | 156.68 | 50.36 | +211% |
| Phi-4-mini tok/s | 237.38 | 87.21 | +172% |
| Mistral-7B-Instruct-v0.1 tok/s | 166.39 | 53.6 | +210% |
| Mistral-7B-Instruct-v0.2 tok/s | 166.8 | 53.31 | +213% |
| Mistral-7B-Instruct-v0.3 tok/s | 167.06 | 53.66 | +211% |
| Ornith-1.0-35B tok/s | 131.77 | 70.31 | +87% |
| Ornith-1.0-9B tok/s | 126.79 | 43.35 | +192% |
| Phi-3.5-mini tok/s | 222.53 | 90.56 | +146% |
| Qwen-AgentWorld-35B-A3B tok/s | 131.16 | 70.23 | +87% |
| Qwen3-0.6B tok/s | 408.39 | 356.93 | +14% |
| Qwen3-1.7B tok/s | 341.85 | 176.08 | +94% |
| Qwen3-14B tok/s | 90.07 | 27.48 | +228% |
| Qwen3-8B tok/s | 145.96 | 48.34 | +202% |
| Qwen2.5-0.5B tok/s | 517.09 | 399.4 | +29% |
| Qwen2.5-1.5B tok/s | 315.59 | 191.39 | +65% |
| Qwen2.5-14B tok/s | 86.07 | 27.37 | +214% |
| Qwen2.5-32B tok/s | 43.23 | 12.24 | +253% |
| Qwen2.5-3B tok/s | 240.48 | 110.84 | +117% |
| Qwen2.5-7B tok/s | 159.81 | 53.32 | +200% |
| Qwen2.5-Coder-1.5B tok/s | 322.96 | 184.6 | +75% |
| Qwen2.5-Coder-3B tok/s | 241.51 | 109.42 | +121% |
| AI21-Jamba-Reasoning-3B tok/s | 231.15 | 107.59 | +115% |
| Codestral 22B (Q3_K_M) tok/s | 48.37 | 17.96 | +169% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 156.62 | 48.82 | +221% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 65.33 | 27.87 | +134% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 43.07 | 12.26 | +251% |
| dolphin-2.9-llama3-8b tok/s | 153.34 | 50.04 | +206% |
| Gemma 3 12B (Q3_K_M) tok/s | 69.7 | 32.29 | +116% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 109.1 | 70.87 | +54% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 146.8 | 48.72 | +201% |
| Hermes-3-Llama-3.2-3B tok/s | 261 | 106.14 | +146% |
| L3-8B-Stheno-v3.2 tok/s | 156.55 | 50.07 | +213% |
| LFM2.5-8B-A1B tok/s | 351.47 | 170.2 | +107% |
| Mistral-Nemo-Instruct-2407 tok/s | 106.33 | 33.01 | +222% |
| Mistral Small 24B (Q3_K_M) tok/s | 45.34 | 16.62 | +173% |
| NemoMix-Unleashed-12B tok/s | 105.05 | 33.02 | +218% |
| Nemotron-3-Nano-30B-A3B tok/s | 192.44 | 0 | n/a |
| Hermes-4-14B tok/s | 89.97 | 27.48 | +227% |
| phi-2 tok/s | 228.1 | 114.99 | +98% |
| Phi-4 14B (Q3_K_M) tok/s | 80.71 | 27.86 | +190% |
| Qwen3-30B-A3B tok/s | 167.1 | 95.02 | +76% |
| Qwen3-4B-Instruct-2507 tok/s | 192.89 | 83.03 | +132% |
| Qwen1.5-0.5B tok/s | 497.73 | 431.41 | +15% |
| Qwen2-1.5B tok/s | 322.09 | 189.18 | +70% |
| Uncensored tok/s | 85.82 | 27.29 | +214% |
| Qwen2.5-Coder-0.5B tok/s | 542.36 | 397.1 | +37% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 86.34 | 27.29 | +216% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 31.04 | 12.05 | +158% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 159.76 | 53.02 | +201% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 142.78 | 94.03 | +52% |
| Qwen3-4B-Instruct-2507 tok/s | 193.39 | 83.13 | +133% |
| Qwen3-4B-Thinking-2507 tok/s | 189.89 | 83.12 | +128% |
| SmolLM2-135M tok/s | 544.06 | 652.15 | -17% |
| Cydonia-24B-v4.3 tok/s | 62.41 | 17.3 | +261% |
| Qwen2.5 1.5B served serve tok/s | 4111.1 | 1827 | +125% |
| Qwen2.5 7B served serve tok/s | 1769.6 | 469.9 | +277% |
| SmolLM2 1.7B served serve tok/s | 3865.2 | 1524.9 | +153% |
| TinyLlama 1.1B served serve tok/s | 3221.3 | 2584.3 | +25% |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.694 | 0.968 | -28% |
| 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 |
| Qwen2.5-72B 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 | 0 | n/a |
| Hermes-4-70B tok/s | 0 | 0 | n/a |
| Llama-3.3-70B-Instruct-abliterated tok/s | 0 | 0 | n/a |
| Meta-Llama-3.1-70B tok/s | 0 | 0 | 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 6 of 7; NVIDIA L4 on 1.
| Workflow | NVIDIA A100 40GB SXM4 | NVIDIA L4 | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 10 s | 17 s | NVIDIA A100 40GB SXM4 1.77x faster | $0.003 vs $0.002 |
| 30-minute podcast pass | 59 s | 2 min | NVIDIA A100 40GB SXM4 2.04x faster | $0.016 vs $0.015 |
| 24-frame storyboard | 3 min | 12.2 min | NVIDIA A100 40GB SXM4 4.01x faster | $0.051 vs $0.089 |
| 500-image masking run | 3.5 min | 7.5 min | NVIDIA A100 40GB SXM4 2.14x faster | $0.058 vs $0.055 |
| Full codebase review | 11.7 min | 36.4 min | NVIDIA A100 40GB SXM4 3.12x faster | $0.195 vs $0.267 |
| 200-product catalogue cutout | 12.2 min | 12.1 min | NVIDIA L4 1.01x faster | $0.203 vs $0.089 |
| 10 short social clips | 13.7 min | 60.3 min | NVIDIA A100 40GB SXM4 4.41x faster | $0.228 vs $0.442 |
Renting by the hour, NVIDIA L4 finishes 4 of 7 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 L4 | $0.440 |
NVIDIA L4 is 2.27x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA A100 40GB SXM4 | NVIDIA L4 | |
|---|---|---|
| VRAM | 40GB | 24GB |
| Architecture | Ampere | Ada Lovelace |
| Memory bandwidth | 1555 GB/s | 300 GB/s |
| Boost clock | 1,410 MHz | 2,040 MHz |
| TDP | 400 W | 72 W |
| Launch MSRP | $12,000 | $2,500 |
| Release | 2020-05-14 | 2023-03-21 |
FAQ
- Which is better for ai & machine learning: NVIDIA A100 40GB SXM4 or NVIDIA L4?
- NVIDIA A100 40GB SXM4 performs better for ai & machine learning, winning 128 of 149 benchmarks in our suite with an average 152.9% advantage.
- What are the main hardware differences between NVIDIA A100 40GB SXM4 and NVIDIA L4?
- NVIDIA A100 40GB SXM4 has 40GB VRAM and a 400W TDP, while NVIDIA L4 has 24GB VRAM and a 72W TDP.
- Does VRAM matter more than speed between NVIDIA A100 40GB SXM4 and NVIDIA L4?
- For AI, yes, NVIDIA A100 40GB SXM4 fits 9 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 A100 40GB SXM4 and NVIDIA L4?
- Wan 2.2 5B (720p): NVIDIA A100 40GB SXM4 leads by roughly 51% (0.66 vs 0.15 frames/s) in our testing.
NVIDIA A100 40GB SXM4 full review · NVIDIA L4 full review · All AI & Machine Learning rankings