NVIDIA H100 80GB HBM3 vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA H100 80GB HBM3 wins 147 of 155 benchmarks, averaging 72.6% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Qwen-Image-Edit, where NVIDIA H100 80GB HBM3 leads by 247% (3.68 vs 1.06 images/min); the closest fight is Dolphin X1 Trinity Nano 6B (1% apart); VRAM decides part of this one: NVIDIA H100 80GB HBM3 runs 156 of our 12 AI workloads while the other card runs 152, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA H100 80GB HBM3 | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 310.26 | 212.05 | +46% |
| Llama 3.1 8B tok/s | 261.83 | 135.51 | +93% |
| Qwen2.5-Coder 14B tok/s | 144.84 | 74.6 | +94% |
| Qwen3 32B tok/s | 74.07 | 34.44 | +115% |
| Llama 3.3 70B tok/s | 41 | 16.49 | +149% |
| Stable Diffusion XL images/min | 34.58 | 17.02 | +103% |
| Z-Image Turbo images/min | 22.125 | 8.325 | +166% |
| FLUX.1 dev images/min | 9.064 | 3.836 | +136% |
| FLUX.1 Kontext dev images/min | 4.264 | 1.693 | +152% |
| Qwen-Image-Edit images/min | 3.68 | 1.06 | +247% |
| LTX-Video (distilled) frames/s | 17.47 | 8.01 | +118% |
| Wan 2.2 5B (720p) frames/s | 1.33 | 0.49 | +171% |
| DeepSeek-R1 Distill Llama 8B tok/s | 261.15 | 135.55 | +93% |
| DeepSeek-R1 Distill 1.5B tok/s | 537.32 | 426.38 | +26% |
| DeepSeek-R1 Distill 7B tok/s | 264.42 | 143.97 | +84% |
| Gemma 3 12B tok/s | 150.58 | 80.92 | +86% |
| Gemma 3 4B tok/s | 289.61 | 198.87 | +46% |
| Llama 3.2 1B tok/s | 880.64 | 618.87 | +42% |
| Llama 3.2 3B tok/s | 422.87 | 270.85 | +56% |
| Mistral 7B v0.3 tok/s | 275.85 | 144.74 | +91% |
| Phi-4 Mini 3.8B tok/s | 388.86 | 228.93 | +70% |
| Qwen2.5-Coder 7B tok/s | 263.91 | 143.93 | +83% |
| Qwen3 0.6B tok/s | 713.53 | 655.75 | +9% |
| Qwen3 1.7B tok/s | 556.48 | 405.52 | +37% |
| Qwen3 14B tok/s | 151.16 | 75.12 | +101% |
| Qwen3 8B tok/s | 244.22 | 130.27 | +87% |
| SmolLM3 3B tok/s | 397.99 | 272.88 | +46% |
| DeepSeek-R1 Distill 14B tok/s | 144.82 | 74.6 | +94% |
| Gemma 4 12B tok/s | 148.17 | 80.82 | +83% |
| Mistral Small 24B tok/s | 105.62 | 48.43 | +118% |
| Phi-4 14B tok/s | 165.17 | 74.98 | +120% |
| Qwen3 30B A3B tok/s | 283.78 | 213.66 | +33% |
| TRELLIS Image-to-3D assets/hour | 864 | 776.1 | +11% |
| Codestral 22B tok/s | 110.05 | 50.06 | +120% |
| DeepSeek-R1 Distill 32B tok/s | 75.8 | 34.6 | +119% |
| Devstral Small 24B tok/s | 107.79 | 48.43 | +123% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 76.76 | 33.08 | +132% |
| Dolphin Mistral 24B Venice tok/s | 109.13 | 48.43 | +125% |
| Dolphin X1 8B tok/s | 266 | 135.48 | +96% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 266.19 | 135.55 | +96% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 107.83 | 48.4 | +123% |
| Gemma 3 27B tok/s | 84.77 | 38.48 | +120% |
| Qwen2.5-Coder 32B tok/s | 75.81 | 34.59 | +119% |
| Qwen3-Coder 30B A3B tok/s | 296.22 | 219.04 | +35% |
| QwQ 32B tok/s | 75.82 | 34.59 | +119% |
| StarCoder2 15B tok/s | 134.92 | 66.88 | +102% |
| FLUX.1 Schnell images/min | 58.36 | 26.24 | +122% |
| Z-Image Turbo (1024px) images/min | 39.42 | 18.2 | +117% |
| BiRefNet images/min | 1310.97 | 863.81 | +52% |
| Depth Anything V2 Large images/min | 918.71 | 959.33 | -4% |
| Depth Anything V2 Small images/min | 1081.54 | 970.61 | +11% |
| SAM ViT-Base images/min | 1431.27 | 998.79 | +43% |
| SAM ViT-Huge images/min | 320.64 | 238.08 | +35% |
| Swin2SR 4x Upscaler images/min | 25.48 | 34.14 | -25% |
| Kokoro TTS 82M x realtime | 222.02 | 243.99 | -9% |
| MusicGen Small x realtime | 2.47 | 2.43 | +2% |
| Whisper large-v3 x realtime | 181.47 | 193.2 | -6% |
| Florence-2 Base images/min | 249.62 | 184.2 | +36% |
| Florence-2 Large images/min | 156.15 | 106.54 | +47% |
| Qwen2.5 1.5B LoRA train tok/s | 16034.9 | 12120.1 | +32% |
| Qwen2.5 7B LoRA train tok/s | 8514.3 | 3736 | +128% |
| SmolLM2 1.7B LoRA train tok/s | 18910.3 | 12052.1 | +57% |
| TinyLlama 1.1B LoRA train tok/s | 17522 | 16408.1 | +7% |
| CogVideoX-5B I2V clips/min | 0.879 | 0.383 | +130% |
| Qwen2.5 1.5B served serve tok/s | 6794.8 | 4686.8 | +45% |
| Qwen2.5 7B served serve tok/s | 3601.1 | 1406.8 | +156% |
| SmolLM2 1.7B served serve tok/s | 6509.9 | 3525.9 | +85% |
| TinyLlama 1.1B served serve tok/s | 8336.7 | 6278.4 | +33% |
| Stable Video Diffusion XT clips/min | 2.882 | 1.166 | +147% |
| Wan 2.2 TI2V-5B (image to video) clips/min | 4.126 | 1.606 | +157% |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.725 | 0.946 | -23% |
| AI21-Jamba-Reasoning-3B tok/s | 363.09 | 258.41 | +41% |
| Olmo-3.1-32B-Think tok/s | 75.53 | 34.11 | +121% |
| Codestral 22B (Q3_K_M) tok/s | 87.21 | 59.53 | +46% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 105.63 | 48.43 | +118% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 262.76 | 134.78 | +95% |
| DeepSeek-Coder-V2-Lite tok/s | 309.14 | 257.64 | +20% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 245.26 | 130.23 | +88% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 41.17 | 16.49 | +150% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 118.43 | 85.85 | +38% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 73.48 | 34.43 | +113% |
| dolphin-2.9-llama3-8b tok/s | 262.47 | 134.94 | +95% |
| Dolphin X1 Trinity Nano 6B tok/s | 247.94 | 246.45 | +1% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 145.15 | 74.63 | +94% |
| gemma-2-2b-it-abliterated tok/s | 392.65 | 278.14 | +41% |
| gemma-2-2b tok/s | 392.26 | 278.1 | +41% |
| gemma-2-9b tok/s | 174.7 | 88.87 | +97% |
| Gemma 3 12B (Q3_K_M) tok/s | 125.92 | 93.05 | +35% |
| gemma-3-1b tok/s | 504.17 | 426.49 | +18% |
| gemma-3-270m tok/s | 961.29 | 822.52 | +17% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 168.41 | 146.51 | +15% |
| GLM-4.7-Flash tok/s | 186.05 | 157.89 | +18% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 245 | 130.25 | +88% |
| gpt-oss-20b tok/s | 346.8 | 233.82 | +48% |
| Hermes-3-Llama-3.2-3B tok/s | 423.72 | 265.87 | +59% |
| Hermes-4-70B tok/s | 41.2 | 16.45 | +150% |
| SmolLM3-3B tok/s | 399.14 | 272.96 | +46% |
| Qwen3-Coder-Next-abliterated tok/s | 190.32 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 231.83 | 173.06 | +34% |
| L3-8B-Stheno-v3.2 tok/s | 261.83 | 135.53 | +93% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 926.61 | 661.54 | +40% |
| LFM2.5-8B-A1B tok/s | 581.98 | 394.48 | +48% |
| Llama-2-7B tok/s | 284.62 | 150.35 | +89% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 424.25 | 271.01 | +57% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 41.15 | 16.49 | +150% |
| Meta-Llama-3.1-70B tok/s | 41.19 | 16.49 | +150% |
| Meta-Llama-3.1-8B tok/s | 261.7 | 135.46 | +93% |
| Phi-4-mini tok/s | 381.52 | 229.28 | +66% |
| Mistral-7B-Instruct-v0.1 tok/s | 276.29 | 144.82 | +91% |
| Mistral-7B-Instruct-v0.2 tok/s | 276.73 | 144.82 | +91% |
| Mistral-7B-Instruct-v0.3 tok/s | 276.61 | 144.68 | +91% |
| Mistral-Nemo-Instruct-2407 tok/s | 177.38 | 89.72 | +98% |
| Mistral Small 24B (Q3_K_M) tok/s | 84 | 58.21 | +44% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 177.28 | 88.85 | +100% |
| Nemotron-3-Nano-30B-A3B tok/s | 316.51 | 190.96 | +66% |
| Hermes-4-14B tok/s | 152 | 75.1 | +102% |
| Ornith-1.0-35B tok/s | 221.65 | 157.39 | +41% |
| Ornith-1.0-9B tok/s | 216.89 | 114.57 | +89% |
| phi-2 tok/s | 340.09 | 283.26 | +20% |
| Phi-3.5-mini tok/s | 345.07 | 229.06 | +51% |
| Phi-4 14B (Q3_K_M) tok/s | 138.87 | 89.18 | +56% |
| Qwen-AgentWorld-35B-A3B tok/s | 219.1 | 156.98 | +40% |
| Qwen3-0.6B tok/s | 710.97 | 656.26 | +8% |
| Qwen3-1.7B tok/s | 557.09 | 405.3 | +37% |
| Qwen3-14B tok/s | 152.2 | 75.01 | +103% |
| Qwen3-30B-A3B tok/s | 285.44 | 214.05 | +33% |
| Qwen3-4B-Instruct-2507 tok/s | 310.58 | 210.24 | +48% |
| Qwen3-8B tok/s | 244.2 | 130.28 | +87% |
| Qwen3-Coder-Next tok/s | 188.21 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 188.14 | 0 | n/a |
| Qwen1.5-0.5B tok/s | 847.02 | 774.78 | +9% |
| Qwen2-1.5B tok/s | 536.52 | 416.57 | +29% |
| Qwen2.5-0.5B tok/s | 890 | 738.06 | +21% |
| Qwen2.5-1.5B tok/s | 537.24 | 427.23 | +26% |
| Uncensored tok/s | 144.92 | 74.31 | +95% |
| Qwen2.5-14B tok/s | 144.7 | 74.62 | +94% |
| Qwen2.5-32B tok/s | 73.5 | 34.59 | +112% |
| Qwen2.5-3B tok/s | 395.52 | 270.04 | +46% |
| Qwen2.5-7B tok/s | 264.17 | 143.94 | +84% |
| Qwen2.5-Coder-0.5B tok/s | 892.81 | 715.81 | +25% |
| Qwen2.5-Coder-1.5B tok/s | 538.44 | 427.12 | +26% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 145.01 | 74.63 | +94% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 58.05 | 41.06 | +41% |
| Qwen2.5-Coder-3B tok/s | 394 | 269.9 | +46% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 264.24 | 142.97 | +85% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 242.6 | 224.95 | +8% |
| Qwen3-4B-Instruct-2507 tok/s | 311.47 | 212.05 | +47% |
| Qwen3-4B-Thinking-2507 tok/s | 310.84 | 211.88 | +47% |
| Qwen3-Coder-Next tok/s | 193.79 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 191.08 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 186.22 | 0 | n/a |
| SmolLM2-135M tok/s | 898.53 | 916.56 | -2% |
| Cydonia-24B-v4.3 tok/s | 105.49 | 48.43 | +118% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA H100 80GB HBM3 is faster on 16 of 17; NVIDIA L40S on 1.
| Workflow | NVIDIA H100 80GB HBM3 | NVIDIA L40S | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 5 s | 6 s | NVIDIA H100 80GB HBM3 1.11x faster | $0.002 vs $0.001 |
| 30-minute podcast pass | 34 s | 75 s | NVIDIA H100 80GB HBM3 2.21x faster | $0.013 vs $0.017 |
| 500-image masking run | 1.6 min | 2.2 min | NVIDIA H100 80GB HBM3 1.33x faster | $0.036 vs $0.028 |
| 24-frame storyboard | 1.8 min | 4.1 min | NVIDIA H100 80GB HBM3 2.23x faster | $0.040 vs $0.053 |
| 20-asset 3D game kit | 2.2 min | 2.9 min | NVIDIA H100 80GB HBM3 1.34x faster | $0.049 vs $0.039 |
| 60-second AI short film | 2.5 min | 4.9 min | NVIDIA H100 80GB HBM3 1.99x faster | $0.055 vs $0.065 |
| 60-second AI short film, narrated | 2.6 min | 5.1 min | NVIDIA H100 80GB HBM3 1.93x faster | $0.059 vs $0.067 |
| 6-panel comic page | 3 min | 6.5 min | NVIDIA H100 80GB HBM3 2.20x faster | $0.066 vs $0.086 |
| Character sheet, 12 poses | 3.7 min | 8.4 min | NVIDIA H100 80GB HBM3 2.28x faster | $0.082 vs $0.110 |
| Full codebase review | 6.9 min | 13.5 min | NVIDIA H100 80GB HBM3 1.95x faster | $0.154 vs $0.177 |
| 10 short social clips | 8 min | 19.1 min | NVIDIA H100 80GB HBM3 2.39x faster | $0.178 vs $0.251 |
| 200-product catalogue cutout | 8.1 min | 6.2 min | NVIDIA L40S 1.31x faster | $0.181 vs $0.082 |
| 40-product photo shoot | 11.1 min | 26.7 min | NVIDIA H100 80GB HBM3 2.40x faster | $0.248 vs $0.352 |
| 20 long-form articles | 11.4 min | 28.5 min | NVIDIA H100 80GB HBM3 2.50x faster | $0.253 vs $0.375 |
| 40-product shoot, start to finish | 12.9 min | 28 min | NVIDIA H100 80GB HBM3 2.18x faster | $0.287 vs $0.369 |
| 100-photo restoration batch | 23.8 min | 59.6 min | NVIDIA H100 80GB HBM3 2.50x faster | $0.531 vs $0.784 |
| 100-photo restore and enlarge | 27.8 min | 62.5 min | NVIDIA H100 80GB HBM3 2.25x faster | $0.618 vs $0.823 |
Renting by the hour, NVIDIA H100 80GB HBM3 finishes 13 of 17 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA H100 80GB HBM3 | $1.336 |
| NVIDIA L40S | $0.790 |
NVIDIA L40S is 1.69x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA H100 80GB HBM3 | NVIDIA L40S | |
|---|---|---|
| VRAM | 80GB | 48GB |
| Architecture | Hopper | Ada Lovelace |
| Memory bandwidth | 3350 GB/s | 864 GB/s |
| Boost clock | 1,980 MHz | 2,520 MHz |
| TDP | 700 W | 350 W |
| Launch MSRP | $30,000 | $7,500 |
| Release | 2022-09-20 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA H100 80GB HBM3 or NVIDIA L40S?
- NVIDIA H100 80GB HBM3 performs better for ai & machine learning, winning 147 of 155 benchmarks in our suite with an average 72.6% advantage.
- What are the main hardware differences between NVIDIA H100 80GB HBM3 and NVIDIA L40S?
- NVIDIA H100 80GB HBM3 has 80GB VRAM and a 700W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA H100 80GB HBM3 and NVIDIA L40S?
- For AI, yes, NVIDIA H100 80GB HBM3 fits 6 more of our 12 workloads than NVIDIA L40S. 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 H100 80GB HBM3 and NVIDIA L40S?
- Qwen-Image-Edit: NVIDIA H100 80GB HBM3 leads by roughly 247% (3.68 vs 1.06 images/min) in our testing.
NVIDIA H100 80GB HBM3 full review · NVIDIA L40S full review · All AI & Machine Learning rankings