NVIDIA B200 vs NVIDIA H200, AI & Machine Learning Comparison
NVIDIA B200 wins 88 of 147 benchmarks, averaging 2.6% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Swin2SR 4x Upscaler, where NVIDIA H200 leads by 116% (11.97 vs 25.85 images/min); the closest fight is Gemma 3 4B (0% apart).
Benchmark results head-to-head
| Benchmark | NVIDIA B200 | NVIDIA H200 | Difference |
|---|---|---|---|
| SmolLM2-135M tok/s | 815.3 | 905.38 | -10% |
| gemma-3-270m tok/s | 889.52 | 967.35 | -8% |
| Qwen1.5-0.5B tok/s | 699.73 | 855.22 | -18% |
| Qwen2.5-0.5B tok/s | 818.27 | 914.58 | -11% |
| Qwen2.5-Coder-0.5B tok/s | 828.48 | 920.36 | -10% |
| Qwen3 0.6B tok/s | 599.63 | 718.67 | -17% |
| Qwen3-0.6B tok/s | 596.73 | 724.67 | -18% |
| Llama 3.2 1B tok/s | 881.78 | 875.53 | +1% |
| gemma-3-1b tok/s | 519.19 | 513.58 | +1% |
| LFM2.5-1.2B tok/s | 916.79 | 946.53 | -3% |
| DeepSeek-R1 Distill 1.5B tok/s | 456.7 | 542.95 | -16% |
| Qwen2-1.5B tok/s | 454.55 | 546.47 | -17% |
| Qwen2.5-1.5B tok/s | 454.69 | 549.36 | -17% |
| Qwen2.5-Coder-1.5B tok/s | 455.35 | 545.54 | -17% |
| Qwen3 1.7B tok/s | 551.76 | 563.52 | -2% |
| Qwen3-1.7B tok/s | 551.93 | 568.85 | -3% |
| gemma-2-2b tok/s | 391.94 | 398.48 | -2% |
| gemma-2-2b-it-abliterated tok/s | 392.32 | 397.43 | -1% |
| AI21-Jamba-Reasoning-3B tok/s | 360.54 | 370.66 | -3% |
| Hermes-3-Llama-3.2-3B tok/s | 420.6 | 430.79 | -2% |
| Llama 3.2 3B tok/s | 422.09 | 424.61 | -1% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 419.24 | 429.78 | -2% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| Qwen2.5-3B tok/s | 388.14 | 400.48 | -3% |
| Qwen2.5-Coder-3B tok/s | 388.17 | 401.11 | -3% |
| SmolLM3 3B tok/s | 400.73 | 402.24 | -0% |
| SmolLM3-3B tok/s | 399.05 | 404.27 | -1% |
| Phi-4 Mini 3.8B tok/s | 375.74 | 392.38 | -4% |
| Gemma 3 4B tok/s | 291.82 | 291.76 | 0% |
| Qwen3 4B tok/s | 317.93 | 318.84 | -0% |
| Qwen3-4B-Instruct-2507 tok/s | 316.66 | 318.67 | -1% |
| Qwen3-4B-Instruct-2507 tok/s | 316.7 | 318.27 | -0% |
| Qwen3-4B-Thinking-2507 tok/s | 317.03 | 318.94 | -1% |
| phi-2 tok/s | 348.72 | 347.08 | 0% |
| Dolphin X1 Trinity Nano 6B tok/s | 209.31 | 259.81 | -19% |
| Phi-3.5-mini tok/s | 325 | 348.14 | -7% |
| Phi-4-mini tok/s | 375.06 | 395.3 | -5% |
| DeepSeek-R1 Distill 7B tok/s | 277.02 | 267.1 | +4% |
| Llama-2-7B tok/s | 292 | 290.81 | 0% |
| Mistral 7B v0.3 tok/s | 287.42 | 278.5 | +3% |
| Mistral-7B-Instruct-v0.1 tok/s | 286.89 | 282.64 | +2% |
| Mistral-7B-Instruct-v0.2 tok/s | 286.37 | 282.32 | +1% |
| Mistral-7B-Instruct-v0.3 tok/s | 286.48 | 282.39 | +1% |
| Qwen2.5-7B tok/s | 277.17 | 265.15 | +5% |
| Qwen2.5-Coder 7B tok/s | 277.98 | 266.36 | +4% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 276.57 | 270.92 | +2% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 274.3 | 267.82 | +2% |
| DeepSeek-R1 Distill Llama 8B tok/s | 273.9 | 265.25 | +3% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 252.98 | 250.61 | +1% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 274.05 | 267.84 | +2% |
| Dolphin X1 8B tok/s | 273.85 | 268.36 | +2% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 253.28 | 250.88 | +1% |
| L3-8B-Stheno-v3.2 tok/s | 274.48 | 267.92 | +2% |
| LFM2.5-8B-A1B tok/s | 575.15 | 583.96 | -2% |
| Llama 3.1 8B tok/s | 274.41 | 268.31 | +2% |
| Meta-Llama-3.1-8B tok/s | 274.17 | 262.93 | +4% |
| Qwen3 8B tok/s | 253.45 | 247.98 | +2% |
| Qwen3-8B tok/s | 253.14 | 251.04 | +1% |
| dolphin-2.9-llama3-8b tok/s | 274.48 | 267.96 | +2% |
| Ornith-1.0-9B tok/s | 226.2 | 222.94 | +1% |
| gemma-2-9b tok/s | 187.45 | 180.47 | +4% |
| Gemma 3 12B tok/s | 151.41 | 153.26 | -1% |
| Gemma 3 12B (Q3_K_M) tok/s | 125.31 | 127.04 | -1% |
| Gemma 4 12B tok/s | 146.36 | 151.05 | -3% |
| NemoMix-Unleashed-12B tok/s | 183.93 | 181.53 | +1% |
| DeepSeek-R1 Distill 14B tok/s | 150.69 | 146.82 | +3% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 120.47 | 119.66 | +1% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 150.67 | 148.74 | +1% |
| Hermes-4-14B tok/s | 158.22 | 156.26 | +1% |
| Phi-4 14B tok/s | 163.85 | 168.2 | -3% |
| Phi-4 14B (Q3_K_M) tok/s | 135.21 | 139.97 | -3% |
| Qwen2.5-14B tok/s | 150.57 | 148.76 | +1% |
| Qwen2.5-Coder 14B tok/s | 150.96 | 148.43 | +2% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 150.67 | 148.48 | +1% |
| Qwen3 14B tok/s | 158.34 | 154.51 | +2% |
| Qwen3-14B tok/s | 158.19 | 156.3 | +1% |
| Uncensored tok/s | 150.51 | 148.7 | +1% |
| StarCoder2 15B tok/s | 138.64 | 136.29 | +2% |
| Mistral-Nemo-Instruct-2407 tok/s | 183.92 | 181.3 | +1% |
| gpt-oss-20b tok/s | 351.42 | 354.23 | -1% |
| Codestral 22B tok/s | 112.16 | 111.34 | +1% |
| Codestral 22B (Q3_K_M) tok/s | 89.9 | 88.03 | +2% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 167.76 | 171.45 | -2% |
| DeepSeek-Coder-V2-Lite tok/s | 325.85 | 312.1 | +4% |
| Cydonia-24B-v4.3 tok/s | 113.81 | 109.2 | +4% |
| Devstral Small 24B tok/s | 113.91 | 109.05 | +4% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 113.95 | 109.06 | +4% |
| Dolphin Mistral 24B Venice tok/s | 113.93 | 109.14 | +4% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 113.8 | 109.2 | +4% |
| Mistral Small 24B tok/s | 114.04 | 107.84 | +6% |
| Mistral Small 24B (Q3_K_M) tok/s | 87.72 | 85.1 | +3% |
| Gemma 3 27B tok/s | 86.08 | 84.75 | +2% |
| Nemotron-3-Nano-30B-A3B tok/s | 345.93 | 327.48 | +6% |
| Qwen3 30B A3B tok/s | 270.84 | 292.11 | -7% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 229.58 | 246.35 | -7% |
| Qwen3-30B-A3B tok/s | 271.69 | 293.47 | -7% |
| Qwen3-Coder 30B A3B tok/s | 277.68 | 297.9 | -7% |
| DeepSeek-R1 Distill 32B tok/s | 78.55 | 75.8 | +4% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 78.48 | 75.71 | +4% |
| Olmo-3.1-32B-Think tok/s | 79.75 | 78.46 | +2% |
| QwQ 32B tok/s | 78.47 | 75.77 | +4% |
| Qwen2.5-32B tok/s | 78.55 | 75.76 | +4% |
| Qwen2.5-Coder 32B tok/s | 78.53 | 75.77 | +4% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 60.24 | 58.87 | +2% |
| Qwen3 32B tok/s | 78.56 | 76.58 | +3% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 79.35 | 76.7 | +3% |
| Ornith-1.0-35B tok/s | 221.13 | 208.51 | +6% |
| Qwen-AgentWorld-35B-A3B tok/s | 221.78 | 225.31 | -2% |
| GLM-4.7-Flash tok/s | 183.82 | 188.57 | -3% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 230.68 | 236.31 | -2% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 44.49 | 42.76 | +4% |
| Hermes-4-70B tok/s | 44.49 | 42.77 | +4% |
| Llama 3.3 70B tok/s | 44.54 | 42.66 | +4% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 44.48 | 42.75 | +4% |
| Meta-Llama-3.1-70B tok/s | 44.48 | 42.72 | +4% |
| Qwen2.5-72B tok/s | 45.34 | 43.4 | +4% |
| Qwen3-Next-80B-A3B tok/s | 179.71 | 192.52 | -7% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 186.03 | 194.82 | -5% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 188.78 | 201.12 | -6% |
| Qwen3-Coder-Next tok/s | 186.39 | 195.21 | -5% |
| Qwen3-Coder-Next tok/s | 189.76 | 196.65 | -4% |
| Qwen3-Coder-Next-abliterated tok/s | 186.57 | 197.14 | -5% |
| BiRefNet images/min | 1082.03 | 1457.73 | -26% |
| Swin2SR 4x Upscaler images/min | 11.97 | 25.85 | -54% |
| Stable Diffusion XL images/min | 46.12 | 37.16 | +24% |
| Z-Image Turbo (1024px) images/min | 62.47 | 40.74 | +53% |
| Z-Image Turbo images/min | 32.325 | 23.175 | +39% |
| FLUX.1 Schnell images/min | 81.41 | 60.51 | +35% |
| FLUX.1 Kontext dev images/min | 5.721 | 4.393 | +30% |
| FLUX.1 dev images/min | 13.136 | 9.514 | +38% |
| Krea 2 Turbo images/min | 9.5 | 7.25 | +31% |
| Qwen-Image-Edit images/min | 4.86 | 3.76 | +29% |
| Wan 2.2 5B (720p) frames/s | 1.88 | 1.41 | +33% |
| LTX-Video (distilled) frames/s | 26.8 | 17.87 | +50% |
| Depth Anything V2 Small images/min | 1137.49 | 1198.58 | -5% |
| SAM ViT-Base images/min | 1662.34 | 1469.34 | +13% |
| Depth Anything V2 Large images/min | 1039.47 | 1040.81 | -0% |
| SAM ViT-Huge images/min | 388.61 | 325.19 | +20% |
| TinyLlama 1.1B LoRA train tok/s | 17993.5 | 16222.4 | +11% |
| TinyLlama 1.1B served serve tok/s | 11562.2 | 9137.7 | +27% |
| Qwen2.5 1.5B LoRA train tok/s | 15322 | 14931.7 | +3% |
| Qwen2.5 1.5B served serve tok/s | 9186.4 | 7541.6 | +22% |
| SmolLM2 1.7B LoRA train tok/s | 18064.7 | 17650.4 | +2% |
| SmolLM2 1.7B served serve tok/s | 9672.9 | 7107.8 | +36% |
| Qwen2.5 7B LoRA train tok/s | 13722 | 8855.4 | +55% |
| Qwen2.5 7B served serve tok/s | 5799.1 | 4394.8 | +32% |
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 H200 on 1.
| Workflow | NVIDIA B200 | NVIDIA H200 | Difference | Cost per run |
|---|---|---|---|---|
| Depth pass on a batch | 6 s | 6 s | NVIDIA B200 1.09x faster | $0.009 vs $0.006 |
| Masking run | 85 s | 1.7 min | NVIDIA B200 1.18x faster | $0.141 vs $0.100 |
| 24-frame storyboard | 89 s | 3.9 min | NVIDIA B200 2.62x faster | $0.149 vs $0.234 |
| 60-second AI short film | 1.9 min | 3.9 min | NVIDIA B200 2.09x faster | $0.186 vs $0.233 |
| 6-panel comic page | 2.3 min | 3.2 min | NVIDIA B200 1.37x faster | $0.233 vs $0.192 |
| Character sheet, 12 poses | 2.9 min | 3.9 min | NVIDIA B200 1.37x faster | $0.285 vs $0.234 |
| Short social clips | 5.6 min | 10 min | NVIDIA B200 1.78x faster | $0.558 vs $0.598 |
| Full codebase review | 6.6 min | 6.7 min | NVIDIA B200 1.02x faster | $0.660 vs $0.403 |
| Product photo shoot | 8.4 min | 11.2 min | NVIDIA B200 1.33x faster | $0.836 vs $0.668 |
| Long-form article batch | 10.5 min | 10.9 min | NVIDIA B200 1.04x faster | $1.044 vs $0.655 |
| Product shoot, start to finish | 12.1 min | 12.9 min | NVIDIA B200 1.07x faster | $1.201 vs $0.770 |
| Product catalogue cutout | 17.2 min | 8 min | NVIDIA H200 2.15x faster | $1.712 vs $0.479 |
| Photo restoration batch | 17.8 min | 23.3 min | NVIDIA B200 1.31x faster | $1.773 vs $1.396 |
| Restore and enlarge photos | 26.4 min | 27.2 min | NVIDIA B200 1.03x faster | $2.628 vs $1.629 |
Renting by the hour, NVIDIA H200 finishes 11 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 H200 | $3.590 |
NVIDIA H200 is 1.67x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA B200 | NVIDIA H200 | |
|---|---|---|
| VRAM | 192GB | 141GB |
| Transistors | 208,000M | 80,000M |
| Process node | 4 nm | 4 nm |
| Architecture | Blackwell | Hopper |
| Memory bandwidth | 8 TB/s | 4800 GB/s |
| Boost clock | n/a | 1,980 MHz |
| TDP | 1000 W | 700 W |
| Launch MSRP | $40,000 | $31,000 |
| Release | 2025-02-01 | 2024-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA B200 or NVIDIA H200?
- NVIDIA B200 performs better for ai & machine learning, winning 88 of 147 benchmarks in our suite with an average 2.6% advantage.
- What are the main hardware differences between NVIDIA B200 and NVIDIA H200?
- NVIDIA B200 has 192GB VRAM and a 1000W TDP, while NVIDIA H200 has 141GB VRAM and a 700W TDP.
- Where is the biggest performance difference between NVIDIA B200 and NVIDIA H200?
- Swin2SR 4x Upscaler: NVIDIA H200 leads by roughly 116% (11.97 vs 25.85 images/min) in our testing.
NVIDIA B200 full review · NVIDIA H200 full review · All AI & Machine Learning rankings