NVIDIA B200 vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA B200 wins 137 of 145 benchmarks, averaging 87.4% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Qwen-Image-Edit, where NVIDIA B200 leads by 358% (4.86 vs 1.06 images/min); the closest fight is Qwen3 30B A3B (Q3_K_M) (2% apart); VRAM decides part of this one: NVIDIA B200 runs 152 of our 12 AI workloads while the other card runs 145, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA B200 | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 317.93 | 212.05 | +50% |
| Llama 3.1 8B tok/s | 274.41 | 135.51 | +103% |
| Qwen2.5-Coder 14B tok/s | 150.96 | 74.6 | +102% |
| Qwen3 32B tok/s | 78.56 | 34.44 | +128% |
| Llama 3.3 70B tok/s | 44.54 | 16.49 | +170% |
| Stable Diffusion XL images/min | 46.12 | 17.02 | +171% |
| Z-Image Turbo images/min | 32.325 | 8.325 | +288% |
| FLUX.1 dev images/min | 13.136 | 3.836 | +242% |
| FLUX.1 Kontext dev images/min | 5.721 | 1.693 | +238% |
| Qwen-Image-Edit images/min | 4.86 | 1.06 | +358% |
| LTX-Video (distilled) frames/s | 26.8 | 8.01 | +235% |
| Wan 2.2 5B (720p) frames/s | 1.88 | 0.49 | +284% |
| DeepSeek-R1 Distill Llama 8B tok/s | 273.9 | 135.55 | +102% |
| DeepSeek-R1 Distill 1.5B tok/s | 456.7 | 426.38 | +7% |
| DeepSeek-R1 Distill 14B tok/s | 150.69 | 74.6 | +102% |
| DeepSeek-R1 Distill 7B tok/s | 277.02 | 143.97 | +92% |
| Gemma 3 12B tok/s | 151.41 | 80.92 | +87% |
| Gemma 3 4B tok/s | 291.82 | 198.87 | +47% |
| Gemma 4 12B tok/s | 146.36 | 80.82 | +81% |
| Llama 3.2 1B tok/s | 881.78 | 618.87 | +42% |
| Llama 3.2 3B tok/s | 422.09 | 270.85 | +56% |
| Mistral 7B v0.3 tok/s | 287.42 | 144.74 | +99% |
| Mistral Small 24B tok/s | 114.04 | 48.43 | +135% |
| Phi-4 14B tok/s | 163.85 | 74.98 | +119% |
| Phi-4 Mini 3.8B tok/s | 375.74 | 228.93 | +64% |
| Qwen2.5-Coder 7B tok/s | 277.98 | 143.93 | +93% |
| Qwen3 0.6B tok/s | 599.63 | 655.75 | -9% |
| Qwen3 1.7B tok/s | 551.76 | 405.52 | +36% |
| Qwen3 14B tok/s | 158.34 | 75.12 | +111% |
| Qwen3 30B A3B tok/s | 270.84 | 213.66 | +27% |
| Qwen3 8B tok/s | 253.45 | 130.27 | +95% |
| SmolLM3 3B tok/s | 400.73 | 272.88 | +47% |
| Codestral 22B tok/s | 112.16 | 50.06 | +124% |
| DeepSeek-R1 Distill 32B tok/s | 78.55 | 34.6 | +127% |
| Devstral Small 24B tok/s | 113.91 | 48.43 | +135% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 79.35 | 33.08 | +140% |
| Dolphin Mistral 24B Venice tok/s | 113.93 | 48.43 | +135% |
| Dolphin X1 8B tok/s | 273.85 | 135.48 | +102% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 274.05 | 135.55 | +102% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 113.95 | 48.4 | +135% |
| Gemma 3 27B tok/s | 86.08 | 38.48 | +124% |
| Qwen2.5-Coder 32B tok/s | 78.53 | 34.59 | +127% |
| Qwen3-Coder 30B A3B tok/s | 277.68 | 219.04 | +27% |
| QwQ 32B tok/s | 78.47 | 34.59 | +127% |
| StarCoder2 15B tok/s | 138.64 | 66.88 | +107% |
| FLUX.1 Schnell images/min | 81.41 | 26.24 | +210% |
| Z-Image Turbo (1024px) images/min | 62.47 | 18.2 | +243% |
| BiRefNet images/min | 1082.03 | 863.81 | +25% |
| Depth Anything V2 Large images/min | 1039.47 | 959.33 | +8% |
| Depth Anything V2 Small images/min | 1137.49 | 970.61 | +17% |
| SAM ViT-Base images/min | 1662.34 | 998.79 | +66% |
| SAM ViT-Huge images/min | 388.61 | 238.08 | +63% |
| Swin2SR 4x Upscaler images/min | 11.97 | 34.14 | -65% |
| Qwen2.5 1.5B LoRA train tok/s | 15322 | 12120.1 | +26% |
| Qwen2.5 7B LoRA train tok/s | 13722 | 3736 | +267% |
| SmolLM2 1.7B LoRA train tok/s | 18064.7 | 12052.1 | +50% |
| TinyLlama 1.1B LoRA train tok/s | 17993.5 | 16408.1 | +10% |
| TinyLlama 1.1B served serve tok/s | 11562.2 | 6278.4 | +84% |
| Qwen2.5 1.5B served serve tok/s | 9186.4 | 4686.8 | +96% |
| Qwen2.5 7B served serve tok/s | 5799.1 | 1406.8 | +312% |
| SmolLM2 1.7B served serve tok/s | 9672.9 | 3525.9 | +174% |
| AI21-Jamba-Reasoning-3B tok/s | 360.54 | 258.41 | +40% |
| Olmo-3.1-32B-Think tok/s | 79.75 | 34.11 | +134% |
| Codestral 22B (Q3_K_M) tok/s | 89.9 | 59.53 | +51% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 113.8 | 48.43 | +135% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 274.3 | 134.78 | +104% |
| DeepSeek-Coder-V2-Lite tok/s | 325.85 | 257.64 | +26% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 252.98 | 130.23 | +94% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 44.49 | 16.49 | +170% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 120.47 | 85.85 | +40% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 78.48 | 34.43 | +128% |
| dolphin-2.9-llama3-8b tok/s | 274.48 | 134.94 | +103% |
| Dolphin X1 Trinity Nano 6B tok/s | 209.31 | 246.45 | -15% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 150.67 | 74.63 | +102% |
| gemma-2-2b-it-abliterated tok/s | 392.32 | 278.14 | +41% |
| gemma-2-2b tok/s | 391.94 | 278.1 | +41% |
| gemma-2-9b tok/s | 187.45 | 88.87 | +111% |
| Gemma 3 12B (Q3_K_M) tok/s | 125.31 | 93.05 | +35% |
| gemma-3-1b tok/s | 519.19 | 426.49 | +22% |
| gemma-3-270m tok/s | 889.52 | 822.52 | +8% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 167.76 | 146.51 | +15% |
| GLM-4.7-Flash tok/s | 183.82 | 157.89 | +16% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 253.28 | 130.25 | +94% |
| gpt-oss-20b tok/s | 351.42 | 233.82 | +50% |
| Hermes-3-Llama-3.2-3B tok/s | 420.6 | 265.87 | +58% |
| Hermes-4-70B tok/s | 44.49 | 16.45 | +170% |
| SmolLM3-3B tok/s | 399.05 | 272.96 | +46% |
| Qwen3-Coder-Next-abliterated tok/s | 186.57 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 230.68 | 173.06 | +33% |
| L3-8B-Stheno-v3.2 tok/s | 274.48 | 135.53 | +103% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 916.79 | 661.54 | +39% |
| LFM2.5-8B-A1B tok/s | 575.15 | 394.48 | +46% |
| Llama-2-7B tok/s | 292 | 150.35 | +94% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 419.24 | 271.01 | +55% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 44.48 | 16.49 | +170% |
| Meta-Llama-3.1-70B tok/s | 44.48 | 16.49 | +170% |
| Meta-Llama-3.1-8B tok/s | 274.17 | 135.46 | +102% |
| Phi-4-mini tok/s | 375.06 | 229.28 | +64% |
| Mistral-7B-Instruct-v0.1 tok/s | 286.89 | 144.82 | +98% |
| Mistral-7B-Instruct-v0.2 tok/s | 286.37 | 144.82 | +98% |
| Mistral-7B-Instruct-v0.3 tok/s | 286.48 | 144.68 | +98% |
| Mistral-Nemo-Instruct-2407 tok/s | 183.92 | 89.72 | +105% |
| Mistral Small 24B (Q3_K_M) tok/s | 87.72 | 58.21 | +51% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 183.93 | 88.85 | +107% |
| Nemotron-3-Nano-30B-A3B tok/s | 345.93 | 190.96 | +81% |
| Hermes-4-14B tok/s | 158.22 | 75.1 | +111% |
| Ornith-1.0-35B tok/s | 221.13 | 157.39 | +40% |
| Ornith-1.0-9B tok/s | 226.2 | 114.57 | +97% |
| phi-2 tok/s | 348.72 | 283.26 | +23% |
| Phi-3.5-mini tok/s | 325 | 229.06 | +42% |
| Phi-4 14B (Q3_K_M) tok/s | 135.21 | 89.18 | +52% |
| Qwen-AgentWorld-35B-A3B tok/s | 221.78 | 156.98 | +41% |
| Qwen3-0.6B tok/s | 596.73 | 656.26 | -9% |
| Qwen3-1.7B tok/s | 551.93 | 405.3 | +36% |
| Qwen3-14B tok/s | 158.19 | 75.01 | +111% |
| Qwen3-30B-A3B tok/s | 271.69 | 214.05 | +27% |
| Qwen3-4B-Instruct-2507 tok/s | 316.66 | 210.24 | +51% |
| Qwen3-8B tok/s | 253.14 | 130.28 | +94% |
| Qwen3-Coder-Next tok/s | 186.39 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 186.03 | 0 | n/a |
| Qwen1.5-0.5B tok/s | 699.73 | 774.78 | -10% |
| Qwen2-1.5B tok/s | 454.55 | 416.57 | +9% |
| Qwen2.5-0.5B tok/s | 818.27 | 738.06 | +11% |
| Qwen2.5-1.5B tok/s | 454.69 | 427.23 | +6% |
| Uncensored tok/s | 150.51 | 74.31 | +103% |
| Qwen2.5-14B tok/s | 150.57 | 74.62 | +102% |
| Qwen2.5-32B tok/s | 78.55 | 34.59 | +127% |
| Qwen2.5-3B tok/s | 388.14 | 270.04 | +44% |
| Qwen2.5-7B tok/s | 277.17 | 143.94 | +93% |
| Qwen2.5-Coder-0.5B tok/s | 828.48 | 715.81 | +16% |
| Qwen2.5-Coder-1.5B tok/s | 455.35 | 427.12 | +7% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 150.67 | 74.63 | +102% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 60.24 | 41.06 | +47% |
| Qwen2.5-Coder-3B tok/s | 388.17 | 269.9 | +44% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 276.57 | 142.97 | +93% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 229.58 | 224.95 | +2% |
| Qwen3-4B-Instruct-2507 tok/s | 316.7 | 212.05 | +49% |
| Qwen3-4B-Thinking-2507 tok/s | 317.03 | 211.88 | +50% |
| Qwen3-Coder-Next tok/s | 189.76 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 188.78 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 179.71 | 0 | n/a |
| SmolLM2-135M tok/s | 815.3 | 916.56 | -11% |
| Cydonia-24B-v4.3 tok/s | 113.81 | 48.43 | +135% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA B200 is faster on 13 of 14; NVIDIA L40S on 1.
| Workflow | NVIDIA B200 | NVIDIA L40S | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 6 s | 6 s | NVIDIA B200 1.07x faster | $0.009 vs $0.001 |
| 500-image masking run | 85 s | 2.2 min | NVIDIA B200 1.52x faster | $0.141 vs $0.028 |
| 24-frame storyboard | 89 s | 4.1 min | NVIDIA B200 2.72x faster | $0.149 vs $0.053 |
| 60-second AI short film | 1.9 min | 4.9 min | NVIDIA B200 2.65x faster | $0.186 vs $0.065 |
| 6-panel comic page | 2.3 min | 6.5 min | NVIDIA B200 2.79x faster | $0.233 vs $0.086 |
| Character sheet, 12 poses | 2.9 min | 8.4 min | NVIDIA B200 2.92x faster | $0.285 vs $0.110 |
| 10 short social clips | 5.6 min | 19.1 min | NVIDIA B200 3.40x faster | $0.558 vs $0.251 |
| Full codebase review | 6.6 min | 13.5 min | NVIDIA B200 2.03x faster | $0.660 vs $0.177 |
| 40-product photo shoot | 8.4 min | 26.7 min | NVIDIA B200 3.18x faster | $0.836 vs $0.352 |
| 20 long-form articles | 10.5 min | 28.5 min | NVIDIA B200 2.72x faster | $1.044 vs $0.375 |
| 40-product shoot, start to finish | 12.1 min | 28 min | NVIDIA B200 2.33x faster | $1.201 vs $0.369 |
| 200-product catalogue cutout | 17.2 min | 6.2 min | NVIDIA L40S 2.76x faster | $1.712 vs $0.082 |
| 100-photo restoration batch | 17.8 min | 59.6 min | NVIDIA B200 3.35x faster | $1.773 vs $0.784 |
| 100-photo restore and enlarge | 26.4 min | 62.5 min | NVIDIA B200 2.37x faster | $2.628 vs $0.823 |
Renting by the hour, NVIDIA L40S finishes 14 of 14 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA B200 | $5.980 |
| NVIDIA L40S | $0.790 |
NVIDIA L40S is 7.57x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA B200 | NVIDIA L40S | |
|---|---|---|
| VRAM | 192GB | 48GB |
| Architecture | Blackwell | Ada Lovelace |
| Memory bandwidth | 8 TB/s | 864 GB/s |
| Boost clock | n/a | 2,520 MHz |
| TDP | 1000 W | 350 W |
| Launch MSRP | $40,000 | $7,500 |
| Release | 2025-02-01 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA B200 or NVIDIA L40S?
- NVIDIA B200 performs better for ai & machine learning, winning 137 of 145 benchmarks in our suite with an average 87.4% advantage.
- What are the main hardware differences between NVIDIA B200 and NVIDIA L40S?
- NVIDIA B200 has 192GB VRAM and a 1000W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA B200 and NVIDIA L40S?
- For AI, yes, NVIDIA B200 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 B200 and NVIDIA L40S?
- Qwen-Image-Edit: NVIDIA B200 leads by roughly 358% (4.86 vs 1.06 images/min) in our testing.
NVIDIA B200 full review · NVIDIA L40S full review · All AI & Machine Learning rankings