NVIDIA L4 vs NVIDIA RTX PRO 6000 Blackwell Workstation Edition, AI & Machine Learning Comparison
NVIDIA RTX PRO 6000 Blackwell Workstation Edition wins 143 of 145 benchmarks, averaging 327.6% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Z-Image Turbo (1024px), where NVIDIA RTX PRO 6000 Blackwell Workstation Edition leads by 583% (4.6 vs 31.43 images/min); the closest fight is Depth Anything V2 Small (50% apart); VRAM decides part of this one: NVIDIA RTX PRO 6000 Blackwell Workstation Edition runs 144 of our 12 AI workloads while the other card runs 138, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA L4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 84 | 362.81 | -77% |
| Llama 3.1 8B tok/s | 50.45 | 258.37 | -80% |
| Qwen2.5-Coder 14B tok/s | 27.46 | 147.74 | -81% |
| Qwen3-32B tok/s | 12.42 | 70.13 | -82% |
| Llama 3.3 70B tok/s | 0 | 34.87 | n/a |
| Stable Diffusion XL images/min | 5.18 | 27.94 | -81% |
| Z-Image Turbo images/min | 2.475 | 15.6 | -84% |
| FLUX.1 dev images/min | 0 | 6.921 | n/a |
| FLUX.1 Kontext dev images/min | 0 | 3.086 | n/a |
| Qwen-Image-Edit images/min | 0 | 2.64 | n/a |
| DeepSeek-R1 Distill Llama 8B tok/s | 50.39 | 235.07 | -79% |
| DeepSeek-R1 Distill 1.5B tok/s | 191.38 | 635.1 | -70% |
| DeepSeek-R1 Distill 14B tok/s | 27.39 | 135.41 | -80% |
| DeepSeek-R1 Distill 7B tok/s | 53.27 | 256.04 | -79% |
| Gemma 3 12B tok/s | 31.01 | 138.09 | -78% |
| Gemma 3 4B tok/s | 82.03 | 297.73 | -72% |
| Gemma 4 12B tok/s | 31.49 | 137.73 | -77% |
| Llama 3.2 1B tok/s | 259.48 | 892.74 | -71% |
| Llama 3.2 3B tok/s | 107.1 | 435.63 | -75% |
| Mistral 7B v0.3 tok/s | 54.02 | 247.33 | -78% |
| Mistral Small 24B tok/s | 17.33 | 93.55 | -81% |
| Phi-4 14B tok/s | 27.24 | 142.68 | -81% |
| Phi-4 Mini 3.8B tok/s | 88.48 | 391.04 | -77% |
| Qwen2.5-Coder 7B tok/s | 53.26 | 256.11 | -79% |
| Qwen3 0.6B tok/s | 357.79 | 788.97 | -55% |
| Qwen3 1.7B tok/s | 177.5 | 575.77 | -69% |
| Qwen3 14B tok/s | 27.61 | 139.01 | -80% |
| Qwen3 30B A3B tok/s | 96.15 | 308.74 | -69% |
| Qwen3 8B tok/s | 48.95 | 225.95 | -78% |
| SmolLM3 3B tok/s | 112.07 | 406.45 | -72% |
| Codestral 22B tok/s | 18.14 | 92.7 | -80% |
| DeepSeek-R1 Distill 32B tok/s | 12.28 | 66.06 | -81% |
| Devstral Small 24B tok/s | 17.34 | 93.52 | -81% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 11.92 | 63.74 | -81% |
| Dolphin Mistral 24B Venice tok/s | 17.34 | 93.67 | -81% |
| Dolphin X1 8B tok/s | 50.12 | 239.4 | -79% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 50.24 | 239.28 | -79% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 17.32 | 93.62 | -81% |
| Gemma 3 27B tok/s | 14.18 | 72.1 | -80% |
| Qwen2.5-Coder 32B tok/s | 12.29 | 66.07 | -81% |
| Qwen3-Coder 30B A3B tok/s | 99.14 | 317.68 | -69% |
| QwQ 32B tok/s | 12.28 | 66.07 | -81% |
| StarCoder2 15B tok/s | 24.4 | 122.09 | -80% |
| Z-Image Turbo (1024px) images/min | 4.6 | 31.43 | -85% |
| BiRefNet images/min | 305.81 | 1445.92 | -79% |
| Depth Anything V2 Large images/min | 770.37 | 1328.11 | -42% |
| Depth Anything V2 Small images/min | 916.66 | 1374.58 | -33% |
| Dolphin X1 Trinity Nano 6B tok/s | 166.6 | 282.02 | -41% |
| SAM ViT-Base images/min | 315.89 | 1079.89 | -71% |
| SAM ViT-Huge images/min | 70.46 | 193.47 | -64% |
| Swin2SR 4x Upscaler images/min | 18.23 | 52.83 | -65% |
| FLUX.1 Schnell images/min | 0 | 42.41 | n/a |
| Krea 2 Turbo images/min | 0 | 4.71 | n/a |
| Qwen2.5-1.5B-Instruct tok/s | 191.39 | 636.39 | -70% |
| gpt-oss-20b tok/s | 91.78 | 375.09 | -76% |
| Phi-3.5-mini-instruct tok/s | 90.56 | 362.79 | -75% |
| Qwen2.5-3B-Instruct tok/s | 110.84 | 398.72 | -72% |
| Qwen-AgentWorld-35B-A3B tok/s | 70.23 | 234.82 | -70% |
| Meta-Llama-3.1-8B-Instruct tok/s | 50.36 | 237.69 | -79% |
| Qwen2.5-7B-Instruct tok/s | 53.32 | 256.63 | -79% |
| Ornith-1.0-35B tok/s | 70.31 | 237.7 | -70% |
| Kwaipilot_KAT-Coder-V2.5-Dev tok/s | 79.01 | 253.35 | -69% |
| Qwen2.5-0.5B-Instruct tok/s | 399.4 | 905.27 | -56% |
| GLM-4.7-Flash tok/s | 75.03 | 215.61 | -65% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 49.07 | 226.85 | -78% |
| gemma-2-2b-it tok/s | 114.17 | 417.07 | -73% |
| gemma-3-1b-it tok/s | 216.27 | 567.1 | -62% |
| gemma-3-270m-it tok/s | 493.31 | 965 | -49% |
| Ornith-1.0-9B tok/s | 43.35 | 201.69 | -79% |
| Qwen_Qwen3-0.6B tok/s | 356.93 | 806.77 | -56% |
| Qwen2.5-14B-Instruct tok/s | 27.37 | 135.66 | -80% |
| Qwen2.5 1.5B LoRA train tok/s | 3720.9 | 15753.9 | -76% |
| Qwen2.5 7B LoRA train tok/s | 1063.9 | 6041.8 | -82% |
| SmolLM2 1.7B LoRA train tok/s | 3240.2 | 16255.4 | -80% |
| TinyLlama 1.1B LoRA train tok/s | 4756 | 18556.4 | -74% |
| Qwen2.5 1.5B served serve tok/s | 1827 | 5786.7 | -68% |
| Qwen2.5 7B served serve tok/s | 469.9 | 2282.8 | -79% |
| SmolLM2 1.7B served serve tok/s | 1524.9 | 5642 | -73% |
| TinyLlama 1.1B served serve tok/s | 2584.3 | 7678.7 | -66% |
| AI21-Jamba-Reasoning-3B tok/s | 107.59 | 400.02 | -73% |
| Olmo-3.1-32B-Think tok/s | 12.27 | 65.71 | -81% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 17.2 | 93.71 | -82% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 48.82 | 239.45 | -80% |
| DeepSeek-Coder-V2-Lite tok/s | 110.11 | 352.52 | -69% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 27.87 | 149.81 | -81% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 12.26 | 66.03 | -81% |
| dolphin-2.9-llama3-8b tok/s | 50.04 | 239.29 | -79% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 27.3 | 135.83 | -80% |
| gemma-2-2b-it-abliterated tok/s | 113 | 416.68 | -73% |
| gemma-2-9b tok/s | 33.8 | 156.29 | -78% |
| Gemma 3 12B (Q3_K_M) tok/s | 32.29 | 153.93 | -79% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 70.87 | 199.14 | -64% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 48.72 | 227.15 | -79% |
| Hermes-3-Llama-3.2-3B tok/s | 106.14 | 439.65 | -76% |
| SmolLM3-3B tok/s | 110.77 | 409.73 | -73% |
| L3-8B-Stheno-v3.2 tok/s | 50.07 | 239.84 | -79% |
| LFM2.5-1.2B tok/s | 278.53 | 989.91 | -72% |
| LFM2.5-8B-A1B tok/s | 170.2 | 617.09 | -72% |
| Llama-2-7B tok/s | 56.06 | 263.84 | -79% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 101.16 | 440.15 | -77% |
| Phi-4-mini tok/s | 87.21 | 393.81 | -78% |
| Mistral-7B-Instruct-v0.1 tok/s | 53.6 | 252.55 | -79% |
| Mistral-7B-Instruct-v0.2 tok/s | 53.31 | 253.34 | -79% |
| Mistral-7B-Instruct-v0.3 tok/s | 53.66 | 253.2 | -79% |
| Mistral-Nemo-Instruct-2407 tok/s | 33.01 | 164.28 | -80% |
| NemoMix-Unleashed-12B tok/s | 33.02 | 164.21 | -80% |
| Hermes-4-14B tok/s | 27.48 | 139.56 | -80% |
| phi-2 tok/s | 114.99 | 410.12 | -72% |
| Phi-4 14B (Q3_K_M) tok/s | 27.86 | 165.43 | -83% |
| Qwen3-1.7B tok/s | 176.08 | 593.5 | -70% |
| Qwen3-14B tok/s | 27.48 | 139.45 | -80% |
| Qwen3-30B-A3B tok/s | 95.02 | 311.76 | -70% |
| Qwen3-4B-Instruct-2507 tok/s | 83.03 | 333.54 | -75% |
| Qwen3-8B tok/s | 48.34 | 226.88 | -79% |
| Qwen1.5-0.5B tok/s | 431.41 | 1025.58 | -58% |
| Qwen2-1.5B tok/s | 189.18 | 648.49 | -71% |
| Uncensored tok/s | 27.29 | 135.68 | -80% |
| Qwen2.5-32B tok/s | 12.24 | 66 | -81% |
| Qwen2.5-Coder-0.5B tok/s | 397.1 | 910.13 | -56% |
| Qwen2.5-Coder-1.5B tok/s | 184.6 | 650.4 | -72% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 27.29 | 135.88 | -80% |
| Qwen2.5-Coder-3B tok/s | 109.42 | 401.24 | -73% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 53.02 | 257.07 | -79% |
| Qwen3-4B-Instruct-2507 tok/s | 83.13 | 334.71 | -75% |
| Qwen3-4B-Thinking-2507 tok/s | 83.12 | 334.65 | -75% |
| SmolLM2-135M tok/s | 652.15 | 1252.77 | -48% |
| Cydonia-24B-v4.3 tok/s | 17.3 | 93.76 | -82% |
| Codestral 22B (Q3_K_M) tok/s | 17.96 | 107.67 | -83% |
| Mistral Small 24B (Q3_K_M) tok/s | 16.62 | 108.88 | -85% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 12.05 | 75.78 | -84% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 94.03 | 309.61 | -70% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 0 | 32.52 | n/a |
| Hermes-4-70B tok/s | 0 | 32.51 | n/a |
| Qwen3-Coder-Next-abliterated tok/s | 0 | 210.47 | n/a |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| Llama-3.3-70B-Instruct-abliterated tok/s | 0 | 32.53 | n/a |
| Meta-Llama-3.1-70B tok/s | 0 | 32.52 | n/a |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| Nemotron-3-Nano-30B-A3B tok/s | 0 | 325.43 | n/a |
| Qwen3-Coder-Next tok/s | 0 | 205.69 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 203.29 | n/a |
| Qwen2.5-72B tok/s | 0 | 29.65 | n/a |
| 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 |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA L4 is faster on 0 of 5; NVIDIA RTX PRO 6000 Blackwell Workstation Edition on 5.
| Workflow | NVIDIA L4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | Difference |
|---|---|---|---|
| 50-image depth pass | 17 s | 4 s | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 4.10x faster |
| 500-image masking run | 7.5 min | 2.6 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 2.84x faster |
| 200-product catalogue cutout | 12.1 min | 4 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 3.00x faster |
| 24-frame storyboard | 12.2 min | 3.6 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 3.35x faster |
| Full codebase review | 36.4 min | 6.8 min | NVIDIA RTX PRO 6000 Blackwell Workstation Edition 5.38x faster |
Specifications compared
| NVIDIA L4 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition | |
|---|---|---|
| VRAM | 24GB | 96GB |
| Architecture | Ada Lovelace | Blackwell |
| Memory bandwidth | 300 GB/s | 1792 GB/s |
| Boost clock | 2,040 MHz | 2,617 MHz |
| TDP | 72 W | 600 W |
| Launch MSRP | $2,500 | $8,565 |
| Release | 2023-03-21 | 2025-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA L4 or NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition performs better for ai & machine learning, winning 143 of 145 benchmarks in our suite with an average 327.6% advantage.
- What are the main hardware differences between NVIDIA L4 and NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- NVIDIA L4 has 24GB VRAM and a 72W TDP, while NVIDIA RTX PRO 6000 Blackwell Workstation Edition has 96GB VRAM and a 600W TDP.
- Does VRAM matter more than speed between NVIDIA L4 and NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- For AI, yes, NVIDIA RTX PRO 6000 Blackwell Workstation Edition fits 22 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 L4 and NVIDIA RTX PRO 6000 Blackwell Workstation Edition?
- Z-Image Turbo (1024px): NVIDIA RTX PRO 6000 Blackwell Workstation Edition leads by roughly 583% (4.6 vs 31.43 images/min) in our testing.
NVIDIA L4 full review · NVIDIA RTX PRO 6000 Blackwell Workstation Edition full review · All AI & Machine Learning rankings