NVIDIA B300 vs NVIDIA L40S, AI & Machine Learning Comparison
NVIDIA B300 wins 138 of 141 benchmarks, averaging 105.6% 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 B300 leads by 668% (8.14 vs 1.06 images/min); the closest fight is Dolphin X1 Trinity Nano 6B (5% apart); VRAM decides part of this one: NVIDIA B300 runs 152 of our 12 AI workloads while the other card runs 140, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA B300 | NVIDIA L40S | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 333.34 | 212.05 | +57% |
| Llama 3.1 8B tok/s | 287.23 | 135.51 | +112% |
| Qwen2.5-Coder 14B tok/s | 158.49 | 74.6 | +112% |
| Qwen3 32B tok/s | 83.68 | 34.44 | +143% |
| Llama 3.3 70B tok/s | 47.97 | 16.49 | +191% |
| Stable Diffusion XL images/min | 29.2 | 17.02 | +72% |
| Z-Image Turbo images/min | 39.675 | 8.325 | +377% |
| FLUX.1 dev images/min | 20.807 | 3.836 | +442% |
| FLUX.1 Kontext dev images/min | 10.329 | 1.693 | +510% |
| Qwen-Image-Edit images/min | 8.14 | 1.06 | +668% |
| LTX-Video (distilled) frames/s | 31.87 | 8.01 | +298% |
| Wan 2.2 5B (720p) frames/s | 2.94 | 0.49 | +500% |
| DeepSeek-R1 Distill Llama 8B tok/s | 283.77 | 135.55 | +109% |
| DeepSeek-R1 Distill 1.5B tok/s | 562.14 | 426.38 | +32% |
| DeepSeek-R1 Distill 14B tok/s | 156.33 | 74.6 | +110% |
| DeepSeek-R1 Distill 7B tok/s | 284.77 | 143.97 | +98% |
| Gemma 3 12B tok/s | 161.27 | 80.92 | +99% |
| Gemma 3 4B tok/s | 301.07 | 198.87 | +51% |
| Gemma 4 12B tok/s | 155.96 | 80.82 | +93% |
| Llama 3.2 1B tok/s | 889.19 | 618.87 | +44% |
| Llama 3.2 3B tok/s | 432.24 | 270.85 | +60% |
| Mistral 7B v0.3 tok/s | 298.95 | 144.74 | +107% |
| Mistral Small 24B tok/s | 119.3 | 48.43 | +146% |
| Phi-4 14B tok/s | 177.41 | 74.98 | +137% |
| Phi-4 Mini 3.8B tok/s | 398.16 | 228.93 | +74% |
| Qwen2.5-Coder 7B tok/s | 286.5 | 143.93 | +99% |
| Qwen3 0.6B tok/s | 718.58 | 655.75 | +10% |
| Qwen3 1.7B tok/s | 555.1 | 405.52 | +37% |
| Qwen3 14B tok/s | 165.44 | 75.12 | +120% |
| Qwen3 30B A3B tok/s | 284.66 | 213.66 | +33% |
| Qwen3 8B tok/s | 261.4 | 130.27 | +101% |
| SmolLM3 3B tok/s | 411.25 | 272.88 | +51% |
| Codestral 22B tok/s | 119.81 | 50.06 | +139% |
| DeepSeek-R1 Distill 32B tok/s | 82.9 | 34.6 | +140% |
| Devstral Small 24B tok/s | 121.18 | 48.43 | +150% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 85.15 | 33.08 | +157% |
| Dolphin Mistral 24B Venice tok/s | 120.93 | 48.43 | +150% |
| Dolphin X1 8B tok/s | 287.02 | 135.48 | +112% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 285.88 | 135.55 | +111% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 120.95 | 48.4 | +150% |
| Gemma 3 27B tok/s | 93.44 | 38.48 | +143% |
| Qwen2.5-Coder 32B tok/s | 82.9 | 34.59 | +140% |
| Qwen3-Coder 30B A3B tok/s | 284.88 | 219.04 | +30% |
| QwQ 32B tok/s | 82.91 | 34.59 | +140% |
| StarCoder2 15B tok/s | 144.05 | 66.88 | +115% |
| FLUX.1 Schnell images/min | 59.54 | 26.24 | +127% |
| Z-Image Turbo (1024px) images/min | 48.63 | 18.2 | +167% |
| BiRefNet images/min | 1064.73 | 863.81 | +23% |
| Depth Anything V2 Large images/min | 1069.95 | 959.33 | +12% |
| Depth Anything V2 Small images/min | 1491.6 | 970.61 | +54% |
| SAM ViT-Base images/min | 1719.59 | 998.79 | +72% |
| SAM ViT-Huge images/min | 399.37 | 238.08 | +68% |
| Swin2SR 4x Upscaler images/min | 58.29 | 34.14 | +71% |
| Qwen2.5 1.5B LoRA train tok/s | 22351.1 | 12120.1 | +84% |
| Qwen2.5 7B LoRA train tok/s | 14323.1 | 3736 | +283% |
| SmolLM2 1.7B LoRA train tok/s | 26114.3 | 12052.1 | +117% |
| TinyLlama 1.1B LoRA train tok/s | 26938.2 | 16408.1 | +64% |
| AI21-Jamba-Reasoning-3B tok/s | 380.88 | 258.41 | +47% |
| Olmo-3.1-32B-Think tok/s | 87.85 | 34.11 | +158% |
| Codestral 22B (Q3_K_M) tok/s | 107.98 | 59.53 | +81% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 121.87 | 48.43 | +152% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 292.55 | 134.78 | +117% |
| DeepSeek-Coder-V2-Lite tok/s | 340.53 | 257.64 | +32% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 270.35 | 130.23 | +108% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 48.11 | 16.49 | +192% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 144.74 | 85.85 | +69% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 83.19 | 34.43 | +142% |
| dolphin-2.9-llama3-8b tok/s | 290.47 | 134.94 | +115% |
| Dolphin X1 Trinity Nano 6B tok/s | 257.93 | 246.45 | +5% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 159.46 | 74.63 | +114% |
| gemma-2-2b-it-abliterated tok/s | 418.18 | 278.14 | +50% |
| gemma-2-2b tok/s | 417.84 | 278.1 | +50% |
| gemma-2-9b tok/s | 202.22 | 88.87 | +128% |
| Gemma 3 12B (Q3_K_M) tok/s | 149.61 | 93.05 | +61% |
| gemma-3-1b tok/s | 533.32 | 426.49 | +25% |
| gemma-3-270m tok/s | 926.67 | 822.52 | +13% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 178.49 | 146.51 | +22% |
| GLM-4.7-Flash tok/s | 196.52 | 157.89 | +24% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 270.09 | 130.25 | +107% |
| gpt-oss-20b tok/s | 355.56 | 233.82 | +52% |
| Hermes-3-Llama-3.2-3B tok/s | 446.78 | 265.87 | +68% |
| Hermes-4-70B tok/s | 48.1 | 16.45 | +192% |
| SmolLM3-3B tok/s | 420.3 | 272.96 | +54% |
| Qwen3-Coder-Next-abliterated tok/s | 199.12 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 239.66 | 173.06 | +38% |
| L3-8B-Stheno-v3.2 tok/s | 293 | 135.53 | +116% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 955.13 | 661.54 | +44% |
| LFM2.5-8B-A1B tok/s | 596.16 | 394.48 | +51% |
| Llama-2-7B tok/s | 315.9 | 150.35 | +110% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 448.94 | 271.01 | +66% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 48.1 | 16.49 | +192% |
| Meta-Llama-3.1-70B tok/s | 48.12 | 16.49 | +192% |
| Meta-Llama-3.1-8B tok/s | 293 | 135.46 | +116% |
| Phi-4-mini tok/s | 409.58 | 229.28 | +79% |
| Mistral-7B-Instruct-v0.1 tok/s | 306.44 | 144.82 | +112% |
| Mistral-7B-Instruct-v0.2 tok/s | 306.79 | 144.82 | +112% |
| Mistral-7B-Instruct-v0.3 tok/s | 306.46 | 144.68 | +112% |
| Mistral-Nemo-Instruct-2407 tok/s | 196.92 | 89.72 | +119% |
| Mistral Small 24B (Q3_K_M) tok/s | 108.37 | 58.21 | +86% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 197.06 | 88.85 | +122% |
| Nemotron-3-Nano-30B-A3B tok/s | 345.31 | 190.96 | +81% |
| Hermes-4-14B tok/s | 169.51 | 75.1 | +126% |
| Ornith-1.0-35B tok/s | 232.36 | 157.39 | +48% |
| Ornith-1.0-9B tok/s | 240.67 | 114.57 | +110% |
| phi-2 tok/s | 356.65 | 283.26 | +26% |
| Phi-3.5-mini tok/s | 352.25 | 229.06 | +54% |
| Phi-4 14B (Q3_K_M) tok/s | 165.06 | 89.18 | +85% |
| Qwen-AgentWorld-35B-A3B tok/s | 233.04 | 156.98 | +48% |
| Qwen3-0.6B tok/s | 742.64 | 656.26 | +13% |
| Qwen3-1.7B tok/s | 575.97 | 405.3 | +42% |
| Qwen3-14B tok/s | 169.52 | 75.01 | +126% |
| Qwen3-30B-A3B tok/s | 291.82 | 214.05 | +36% |
| Qwen3-4B-Instruct-2507 tok/s | 339.78 | 210.24 | +62% |
| Qwen3-8B tok/s | 270.35 | 130.28 | +108% |
| Qwen3-Coder-Next tok/s | 196.34 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 194.98 | 0 | n/a |
| Qwen1.5-0.5B tok/s | 894.84 | 774.78 | +15% |
| Qwen2-1.5B tok/s | 583.79 | 416.57 | +40% |
| Qwen2.5-0.5B tok/s | 843.4 | 738.06 | +14% |
| Qwen2.5-1.5B tok/s | 584.96 | 427.23 | +37% |
| Uncensored tok/s | 159.48 | 74.31 | +115% |
| Qwen2.5-14B tok/s | 159.33 | 74.62 | +114% |
| Qwen2.5-32B tok/s | 83.41 | 34.59 | +141% |
| Qwen2.5-3B tok/s | 409.56 | 270.04 | +52% |
| Qwen2.5-7B tok/s | 293.63 | 143.94 | +104% |
| Qwen2.5-Coder-0.5B tok/s | 843.44 | 715.81 | +18% |
| Qwen2.5-Coder-1.5B tok/s | 584.5 | 427.12 | +37% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 159.5 | 74.63 | +114% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 74.28 | 41.06 | +81% |
| Qwen2.5-Coder-3B tok/s | 410.06 | 269.9 | +52% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 292.21 | 142.97 | +104% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 268.71 | 224.95 | +19% |
| Qwen3-4B-Instruct-2507 tok/s | 339.72 | 212.05 | +60% |
| Qwen3-4B-Thinking-2507 tok/s | 339.74 | 211.88 | +60% |
| Qwen3-Coder-Next tok/s | 199.39 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 200.55 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 191.32 | 0 | n/a |
| SmolLM2-135M tok/s | 855.87 | 916.56 | -7% |
| Cydonia-24B-v4.3 tok/s | 121.9 | 48.43 | +152% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA B300 is faster on 14 of 14; NVIDIA L40S on 0.
| Workflow | NVIDIA B300 | NVIDIA L40S | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 5 s | 6 s | NVIDIA B300 1.30x faster | $0.009 vs $0.001 |
| 500-image masking run | 78 s | 2.2 min | NVIDIA B300 1.66x faster | $0.150 vs $0.028 |
| 24-frame storyboard | 1.5 min | 4.1 min | NVIDIA B300 2.70x faster | $0.174 vs $0.053 |
| 6-panel comic page | 2 min | 6.5 min | NVIDIA B300 3.23x faster | $0.234 vs $0.086 |
| 60-second AI short film | 2.2 min | 4.9 min | NVIDIA B300 2.28x faster | $0.251 vs $0.065 |
| Character sheet, 12 poses | 2.2 min | 8.4 min | NVIDIA B300 3.76x faster | $0.257 vs $0.110 |
| 200-product catalogue cutout | 3.7 min | 6.2 min | NVIDIA B300 1.67x faster | $0.430 vs $0.082 |
| 10 short social clips | 4.7 min | 19.1 min | NVIDIA B300 4.02x faster | $0.548 vs $0.251 |
| 40-product photo shoot | 6.1 min | 26.7 min | NVIDIA B300 4.41x faster | $0.701 vs $0.352 |
| Full codebase review | 6.3 min | 13.5 min | NVIDIA B300 2.13x faster | $0.730 vs $0.177 |
| 40-product shoot, start to finish | 6.9 min | 28 min | NVIDIA B300 4.07x faster | $0.797 vs $0.369 |
| 20 long-form articles | 9.7 min | 28.5 min | NVIDIA B300 2.93x faster | $1.125 vs $0.375 |
| 100-photo restoration batch | 10.2 min | 59.6 min | NVIDIA B300 5.85x faster | $1.178 vs $0.784 |
| 100-photo restore and enlarge | 11.9 min | 62.5 min | NVIDIA B300 5.25x faster | $1.378 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 B300 | $6.940 |
| NVIDIA L40S | $0.790 |
NVIDIA L40S is 8.78x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA B300 | NVIDIA L40S | |
|---|---|---|
| VRAM | 288GB | 48GB |
| Architecture | Blackwell Ultra | Ada Lovelace |
| Memory bandwidth | 8 TB/s | 864 GB/s |
| Boost clock | n/a | 2,520 MHz |
| TDP | 1400 W | 350 W |
| Launch MSRP | $40,000 | $7,500 |
| Release | 2025-11-01 | 2023-08-08 |
FAQ
- Which is better for ai & machine learning: NVIDIA B300 or NVIDIA L40S?
- NVIDIA B300 performs better for ai & machine learning, winning 138 of 141 benchmarks in our suite with an average 105.6% advantage.
- What are the main hardware differences between NVIDIA B300 and NVIDIA L40S?
- NVIDIA B300 has 288GB VRAM and a 1400W TDP, while NVIDIA L40S has 48GB VRAM and a 350W TDP.
- Does VRAM matter more than speed between NVIDIA B300 and NVIDIA L40S?
- For AI, yes, NVIDIA B300 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 B300 and NVIDIA L40S?
- Qwen-Image-Edit: NVIDIA B300 leads by roughly 668% (8.14 vs 1.06 images/min) in our testing.
NVIDIA B300 full review · NVIDIA L40S full review · All AI & Machine Learning rankings