NVIDIA L40S vs NVIDIA RTX PRO 6000 Blackwell Workstation Edition, AI & Machine Learning Comparison
NVIDIA RTX PRO 6000 Blackwell Workstation Edition wins 141 of 144 benchmarks, averaging 66% 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 RTX PRO 6000 Blackwell Workstation Edition leads by 149% (1.06 vs 2.64 images/min); the closest fight is SAM ViT-Base (8% apart); VRAM decides part of this one: NVIDIA RTX PRO 6000 Blackwell Workstation Edition runs 152 of our 12 AI workloads while the other card runs 144, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA L40S | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 212.05 | 362.81 | -42% |
| Llama-3.1-8B tok/s | 135.51 | 258.37 | -48% |
| Qwen2.5-Coder-14B tok/s | 74.6 | 147.74 | -50% |
| Qwen3-32B tok/s | 34.44 | 70.13 | -51% |
| Llama-3.3-70B tok/s | 16.49 | 34.87 | -53% |
| Stable Diffusion XL images/min | 17.02 | 27.94 | -39% |
| Z-Image Turbo images/min | 8.325 | 15.6 | -47% |
| FLUX.1 dev images/min | 3.836 | 6.921 | -45% |
| FLUX.1 Kontext dev images/min | 1.693 | 3.086 | -45% |
| Qwen-Image-Edit images/min | 1.06 | 2.64 | -60% |
| LTX-Video (distilled) frames/s | 8.01 | 16.36 | -51% |
| DeepSeek-R1 Distill Llama 8B tok/s | 135.55 | 235.07 | -42% |
| DeepSeek-R1 Distill 1.5B tok/s | 426.38 | 635.1 | -33% |
| DeepSeek-R1 Distill 14B tok/s | 74.6 | 135.41 | -45% |
| DeepSeek-R1 Distill 7B tok/s | 143.97 | 256.04 | -44% |
| Gemma 3 12B tok/s | 80.92 | 138.09 | -41% |
| Gemma 3 4B tok/s | 198.87 | 297.73 | -33% |
| Gemma 4 12B tok/s | 80.82 | 137.73 | -41% |
| Llama 3.2 1B tok/s | 618.87 | 892.74 | -31% |
| Llama 3.2 3B tok/s | 270.85 | 435.63 | -38% |
| Mistral 7B v0.3 tok/s | 144.74 | 247.33 | -41% |
| Mistral Small 24B tok/s | 48.43 | 93.55 | -48% |
| Phi-4 14B tok/s | 74.98 | 142.68 | -47% |
| Phi-4 Mini 3.8B tok/s | 228.93 | 391.04 | -41% |
| Qwen2.5-Coder 7B tok/s | 143.93 | 256.11 | -44% |
| Qwen3 0.6B tok/s | 655.75 | 788.97 | -17% |
| Qwen3 1.7B tok/s | 405.52 | 575.77 | -30% |
| Qwen3 14B tok/s | 75.12 | 139.01 | -46% |
| Qwen3 30B A3B tok/s | 213.66 | 308.74 | -31% |
| Qwen3 8B tok/s | 130.27 | 225.95 | -42% |
| SmolLM3 3B tok/s | 272.88 | 406.45 | -33% |
| Codestral 22B tok/s | 50.06 | 92.7 | -46% |
| DeepSeek-R1 Distill 32B tok/s | 34.6 | 66.06 | -48% |
| Devstral Small 24B tok/s | 48.43 | 93.52 | -48% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 33.08 | 63.74 | -48% |
| Dolphin Mistral 24B Venice tok/s | 48.43 | 93.67 | -48% |
| Dolphin X1 8B tok/s | 135.48 | 239.4 | -43% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 135.55 | 239.28 | -43% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 48.4 | 93.62 | -48% |
| Gemma 3 27B tok/s | 38.48 | 72.1 | -47% |
| Qwen2.5-Coder 32B tok/s | 34.59 | 66.07 | -48% |
| Qwen3-Coder 30B A3B tok/s | 219.04 | 317.68 | -31% |
| QwQ 32B tok/s | 34.59 | 66.07 | -48% |
| StarCoder2 15B tok/s | 66.88 | 122.09 | -45% |
| FLUX.1 Schnell images/min | 26.24 | 42.41 | -38% |
| Z-Image Turbo (1024px) images/min | 18.2 | 31.43 | -42% |
| BiRefNet images/min | 863.81 | 1445.92 | -40% |
| Depth Anything V2 Large images/min | 959.33 | 1328.11 | -28% |
| Depth Anything V2 Small images/min | 970.61 | 1374.58 | -29% |
| SAM ViT-Base images/min | 998.79 | 1079.89 | -8% |
| SAM ViT-Huge images/min | 238.08 | 193.47 | +23% |
| Swin2SR 4x Upscaler images/min | 34.14 | 52.83 | -35% |
| Dolphin X1 Trinity Nano 6B tok/s | 246.45 | 282.02 | -13% |
| gpt-oss-20b tok/s | 233.82 | 375.09 | -38% |
| Meta-Llama-3.1-8B tok/s | 135.46 | 237.69 | -43% |
| Ornith-1.0-35B tok/s | 157.39 | 237.7 | -34% |
| Qwen-AgentWorld-35B-A3B tok/s | 156.98 | 234.82 | -33% |
| Qwen2.5-3B tok/s | 270.04 | 398.72 | -32% |
| Qwen2.5-7B tok/s | 143.94 | 256.63 | -44% |
| Olmo-3.1-32B-Think tok/s | 34.11 | 65.71 | -48% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 48.43 | 93.71 | -48% |
| DeepSeek-Coder-V2-Lite tok/s | 257.64 | 352.52 | -27% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 130.23 | 226.85 | -43% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 16.49 | 32.52 | -49% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 74.63 | 135.83 | -45% |
| Qwen2.5 1.5B LoRA train tok/s | 12120.1 | 15753.9 | -23% |
| Qwen2.5 7B LoRA train tok/s | 3736 | 6041.8 | -38% |
| SmolLM2 1.7B LoRA train tok/s | 12052.1 | 16255.4 | -26% |
| TinyLlama 1.1B LoRA train tok/s | 16408.1 | 18556.4 | -12% |
| gemma-2-2b-it-abliterated tok/s | 278.14 | 416.68 | -33% |
| gemma-2-2b tok/s | 278.1 | 417.07 | -33% |
| gemma-2-9b tok/s | 88.87 | 156.29 | -43% |
| gemma-3-1b tok/s | 426.49 | 567.1 | -25% |
| gemma-3-270m tok/s | 822.52 | 965 | -15% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 146.51 | 199.14 | -26% |
| GLM-4.7-Flash tok/s | 157.89 | 215.61 | -27% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 130.25 | 227.15 | -43% |
| SmolLM3-3B tok/s | 272.96 | 409.73 | -33% |
| KAT-Coder-V2.5-Dev tok/s | 173.06 | 253.35 | -32% |
| L3-8B-Stheno-v3.2 tok/s | 135.53 | 239.84 | -43% |
| LFM2.5-1.2B tok/s | 661.54 | 989.91 | -33% |
| LFM2.5-8B-A1B tok/s | 394.48 | 617.09 | -36% |
| Llama-2-7B tok/s | 150.35 | 263.84 | -43% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 271.01 | 440.15 | -38% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 16.49 | 32.53 | -49% |
| Meta-Llama-3.1-70B tok/s | 16.49 | 32.52 | -49% |
| Phi-4-mini tok/s | 229.28 | 393.81 | -42% |
| Mistral-7B-Instruct-v0.1 tok/s | 144.82 | 252.55 | -43% |
| Mistral-7B-Instruct-v0.2 tok/s | 144.82 | 253.34 | -43% |
| Mistral-7B-Instruct-v0.3 tok/s | 144.68 | 253.2 | -43% |
| Mistral-Nemo-Instruct-2407 tok/s | 89.72 | 164.28 | -45% |
| Ornith-1.0-9B tok/s | 114.57 | 201.69 | -43% |
| phi-2 tok/s | 283.26 | 410.12 | -31% |
| Phi-3.5-mini tok/s | 229.06 | 362.79 | -37% |
| Qwen3-0.6B tok/s | 656.26 | 806.77 | -19% |
| Qwen3-1.7B tok/s | 405.3 | 593.5 | -32% |
| Qwen3-14B tok/s | 75.01 | 139.45 | -46% |
| Qwen3-8B tok/s | 130.28 | 226.88 | -43% |
| Qwen2.5-0.5B tok/s | 738.06 | 905.27 | -18% |
| Qwen2.5-1.5B tok/s | 427.23 | 636.39 | -33% |
| Qwen2.5-14B tok/s | 74.62 | 135.66 | -45% |
| Qwen2.5-32B tok/s | 34.59 | 66 | -48% |
| Qwen2.5-Coder-1.5B tok/s | 427.12 | 650.4 | -34% |
| Qwen2.5-Coder-3B tok/s | 269.9 | 401.24 | -33% |
| Qwen3-4B-Instruct-2507 tok/s | 212.05 | 334.71 | -37% |
| SmolLM2-135M tok/s | 916.56 | 1252.77 | -27% |
| Cydonia-24B-v4.3 tok/s | 48.43 | 93.76 | -48% |
| Hermes-4-14B tok/s | 75.1 | 139.56 | -46% |
| Qwen3-30B-A3B tok/s | 214.05 | 311.76 | -31% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 74.63 | 135.88 | -45% |
| Qwen3-4B-Thinking-2507 tok/s | 211.88 | 334.65 | -37% |
| Qwen2.5 1.5B served serve tok/s | 4686.8 | 5786.7 | -19% |
| Qwen2.5 7B served serve tok/s | 1406.8 | 2282.8 | -38% |
| SmolLM2 1.7B served serve tok/s | 3525.9 | 5642 | -38% |
| TinyLlama 1.1B served serve tok/s | 6278.4 | 7678.7 | -18% |
| Codestral 22B (Q3_K_M) tok/s | 59.53 | 107.67 | -45% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 224.95 | 309.61 | -27% |
| AI21-Jamba-Reasoning-3B tok/s | 258.41 | 400.02 | -35% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 134.78 | 239.45 | -44% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 34.43 | 66.03 | -48% |
| dolphin-2.9-llama3-8b tok/s | 134.94 | 239.29 | -44% |
| Hermes-3-Llama-3.2-3B tok/s | 265.87 | 439.65 | -40% |
| Hermes-4-70B tok/s | 16.45 | 32.51 | -49% |
| Qwen3-Coder-Next-abliterated tok/s | 0 | 210.47 | n/a |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 88.85 | 164.21 | -46% |
| Nemotron-3-Nano-30B-A3B tok/s | 190.96 | 325.43 | -41% |
| Qwen3-4B-Instruct-2507 tok/s | 210.24 | 333.54 | -37% |
| Qwen3-Coder-Next tok/s | 0 | 205.69 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 203.29 | n/a |
| Qwen1.5-0.5B tok/s | 774.78 | 1025.58 | -24% |
| Qwen2-1.5B tok/s | 416.57 | 648.49 | -36% |
| Uncensored tok/s | 74.31 | 135.68 | -45% |
| Qwen2.5-Coder-0.5B tok/s | 715.81 | 910.13 | -21% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 142.97 | 257.07 | -44% |
| Qwen3-Coder-Next tok/s | 0 | 211.5 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 214.04 | n/a |
| Qwen3-Next-80B-A3B tok/s | 0 | 210.17 | n/a |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 85.85 | 149.81 | -43% |
| Gemma 3 12B (Q3_K_M) tok/s | 93.05 | 153.93 | -40% |
| Phi-4 14B (Q3_K_M) tok/s | 89.18 | 165.43 | -46% |
| Mistral Small 24B (Q3_K_M) tok/s | 58.21 | 108.88 | -47% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 41.06 | 75.78 | -46% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA L40S is faster on 1 of 13; NVIDIA RTX PRO 6000 Blackwell Workstation Edition on 12.
| Workflow | NVIDIA L40S | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | Difference |
|---|---|---|---|
| 50-image depth pass | 6 s | 4 s | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.46x faster |
| 500-image masking run | 2.2 min | 2.6 min | NVIDIA L40S 1.22x faster |
| 24-frame storyboard | 4.1 min | 3.6 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.11x faster |
| 60-second AI short film | 4.9 min | 4.5 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.09x faster |
| 200-product catalogue cutout | 6.2 min | 4 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.55x faster |
| 6-panel comic page | 6.5 min | 5.1 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.29x faster |
| Character sheet, 12 poses | 8.4 min | 6.1 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.37x faster |
| Full codebase review | 13.5 min | 6.8 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.99x faster |
| 40-product photo shoot | 26.7 min | 16.7 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.60x faster |
| 40-product shoot, start to finish | 28 min | 17.6 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.60x faster |
| 20 long-form articles | 28.5 min | 13.4 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 2.13x faster |
| 100-photo restoration batch | 59.6 min | 33.4 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.78x faster |
| 100-photo restore and enlarge | 62.5 min | 35.3 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 1.77x faster |
Specifications compared
| NVIDIA L40S | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | |
|---|---|---|
| VRAM | 48GB | 96GB |
| Architecture | Ada Lovelace | Blackwell |
| Memory bandwidth | 864 GB/s | 1792 GB/s |
| Boost clock | 2,520 MHz | 2,617 MHz |
| TDP | 350 W | 600 W |
| Launch MSRP | $7,500 | $8,565 |
| Release | 2023-08-08 | 2025-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA L40S or NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition performs better for ai & machine learning, winning 141 of 144 benchmarks in our suite with an average 66% advantage.
- What are the main hardware differences between NVIDIA L40S and NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- NVIDIA L40S has 48GB VRAM and a 350W TDP, while NVIDIA RTX PRO 6000 Blackwell Workstation Edition has 96GB VRAM and a 600W TDP.
- Does VRAM matter more than speed between NVIDIA L40S and NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- For AI, yes, NVIDIA RTX PRO 6000 Blackwell Workstation Edition 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 L40S and NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- Qwen-Image-Edit: NVIDIA RTX PRO 6000 Blackwell Workstation Edition leads by roughly 149% (1.06 vs 2.64 images/min) in our testing.
NVIDIA L40S full review · NVIDIA RTX PRO 6000 Blackwell Workstation Edition full review · All AI & Machine Learning rankings