NVIDIA B200 vs NVIDIA H100 80GB HBM3, AI & Machine Learning Comparison
NVIDIA B200 wins 95 of 147 benchmarks, averaging 5.2% 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 H100 80GB HBM3 leads by 113% (11.97 vs 25.48 images/min); the closest fight is SmolLM3-3B (0% apart).
Benchmark results head-to-head
| Benchmark | NVIDIA B200 | NVIDIA H100 80GB HBM3 | Difference |
|---|---|---|---|
| SmolLM2-135M tok/s | 815.3 | 898.53 | -9% |
| gemma-3-270m tok/s | 889.52 | 961.29 | -7% |
| Qwen1.5-0.5B tok/s | 699.73 | 847.02 | -17% |
| Qwen2.5-0.5B tok/s | 818.27 | 890 | -8% |
| Qwen2.5-Coder-0.5B tok/s | 828.48 | 892.81 | -7% |
| Qwen3 0.6B tok/s | 599.63 | 713.53 | -16% |
| Qwen3-0.6B tok/s | 596.73 | 710.97 | -16% |
| Llama 3.2 1B tok/s | 881.78 | 880.64 | 0% |
| gemma-3-1b tok/s | 519.19 | 504.17 | +3% |
| LFM2.5-1.2B tok/s | 916.79 | 926.61 | -1% |
| DeepSeek-R1 Distill 1.5B tok/s | 456.7 | 537.32 | -15% |
| Qwen2-1.5B tok/s | 454.55 | 536.52 | -15% |
| Qwen2.5-1.5B tok/s | 454.69 | 537.24 | -15% |
| Qwen2.5-Coder-1.5B tok/s | 455.35 | 538.44 | -15% |
| Qwen3 1.7B tok/s | 551.76 | 556.48 | -1% |
| Qwen3-1.7B tok/s | 551.93 | 557.09 | -1% |
| gemma-2-2b tok/s | 391.94 | 392.26 | -0% |
| gemma-2-2b-it-abliterated tok/s | 392.32 | 392.65 | -0% |
| AI21-Jamba-Reasoning-3B tok/s | 360.54 | 363.09 | -1% |
| Hermes-3-Llama-3.2-3B tok/s | 420.6 | 423.72 | -1% |
| Llama 3.2 3B tok/s | 422.09 | 422.87 | -0% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 419.24 | 424.25 | -1% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| Qwen2.5-3B tok/s | 388.14 | 395.52 | -2% |
| Qwen2.5-Coder-3B tok/s | 388.17 | 394 | -1% |
| SmolLM3 3B tok/s | 400.73 | 397.99 | +1% |
| SmolLM3-3B tok/s | 399.05 | 399.14 | -0% |
| Phi-4 Mini 3.8B tok/s | 375.74 | 388.86 | -3% |
| Gemma 3 4B tok/s | 291.82 | 289.61 | +1% |
| Qwen3 4B tok/s | 317.93 | 310.26 | +2% |
| Qwen3-4B-Instruct-2507 tok/s | 316.66 | 310.58 | +2% |
| Qwen3-4B-Instruct-2507 tok/s | 316.7 | 311.47 | +2% |
| Qwen3-4B-Thinking-2507 tok/s | 317.03 | 310.84 | +2% |
| phi-2 tok/s | 348.72 | 340.09 | +3% |
| Dolphin X1 Trinity Nano 6B tok/s | 209.31 | 247.94 | -16% |
| Phi-3.5-mini tok/s | 325 | 345.07 | -6% |
| Phi-4-mini tok/s | 375.06 | 381.52 | -2% |
| DeepSeek-R1 Distill 7B tok/s | 277.02 | 264.42 | +5% |
| Llama-2-7B tok/s | 292 | 284.62 | +3% |
| Mistral 7B v0.3 tok/s | 287.42 | 275.85 | +4% |
| Mistral-7B-Instruct-v0.1 tok/s | 286.89 | 276.29 | +4% |
| Mistral-7B-Instruct-v0.2 tok/s | 286.37 | 276.73 | +3% |
| Mistral-7B-Instruct-v0.3 tok/s | 286.48 | 276.61 | +4% |
| Qwen2.5-7B tok/s | 277.17 | 264.17 | +5% |
| Qwen2.5-Coder 7B tok/s | 277.98 | 263.91 | +5% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 276.57 | 264.24 | +5% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 274.3 | 262.76 | +4% |
| DeepSeek-R1 Distill Llama 8B tok/s | 273.9 | 261.15 | +5% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 252.98 | 245.26 | +3% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 274.05 | 266.19 | +3% |
| Dolphin X1 8B tok/s | 273.85 | 266 | +3% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 253.28 | 245 | +3% |
| L3-8B-Stheno-v3.2 tok/s | 274.48 | 261.83 | +5% |
| LFM2.5-8B-A1B tok/s | 575.15 | 581.98 | -1% |
| Llama 3.1 8B tok/s | 274.41 | 261.83 | +5% |
| Meta-Llama-3.1-8B tok/s | 274.17 | 261.7 | +5% |
| Qwen3 8B tok/s | 253.45 | 244.22 | +4% |
| Qwen3-8B tok/s | 253.14 | 244.2 | +4% |
| dolphin-2.9-llama3-8b tok/s | 274.48 | 262.47 | +5% |
| Ornith-1.0-9B tok/s | 226.2 | 216.89 | +4% |
| gemma-2-9b tok/s | 187.45 | 174.7 | +7% |
| Gemma 3 12B tok/s | 151.41 | 150.58 | +1% |
| Gemma 3 12B (Q3_K_M) tok/s | 125.31 | 125.92 | -0% |
| Gemma 4 12B tok/s | 146.36 | 148.17 | -1% |
| NemoMix-Unleashed-12B tok/s | 183.93 | 177.28 | +4% |
| DeepSeek-R1 Distill 14B tok/s | 150.69 | 144.82 | +4% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 120.47 | 118.43 | +2% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 150.67 | 145.15 | +4% |
| Hermes-4-14B tok/s | 158.22 | 152 | +4% |
| Phi-4 14B tok/s | 163.85 | 165.17 | -1% |
| Phi-4 14B (Q3_K_M) tok/s | 135.21 | 138.87 | -3% |
| Qwen2.5-14B tok/s | 150.57 | 144.7 | +4% |
| Qwen2.5-Coder 14B tok/s | 150.96 | 144.84 | +4% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 150.67 | 145.01 | +4% |
| Qwen3 14B tok/s | 158.34 | 151.16 | +5% |
| Qwen3-14B tok/s | 158.19 | 152.2 | +4% |
| Uncensored tok/s | 150.51 | 144.92 | +4% |
| StarCoder2 15B tok/s | 138.64 | 134.92 | +3% |
| Mistral-Nemo-Instruct-2407 tok/s | 183.92 | 177.38 | +4% |
| gpt-oss-20b tok/s | 351.42 | 346.8 | +1% |
| Codestral 22B tok/s | 112.16 | 110.05 | +2% |
| Codestral 22B (Q3_K_M) tok/s | 89.9 | 87.21 | +3% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 167.76 | 168.41 | -0% |
| DeepSeek-Coder-V2-Lite tok/s | 325.85 | 309.14 | +5% |
| Cydonia-24B-v4.3 tok/s | 113.81 | 105.49 | +8% |
| Devstral Small 24B tok/s | 113.91 | 107.79 | +6% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 113.95 | 107.83 | +6% |
| Dolphin Mistral 24B Venice tok/s | 113.93 | 109.13 | +4% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 113.8 | 105.63 | +8% |
| Mistral Small 24B tok/s | 114.04 | 105.62 | +8% |
| Mistral Small 24B (Q3_K_M) tok/s | 87.72 | 84 | +4% |
| Gemma 3 27B tok/s | 86.08 | 84.77 | +2% |
| Nemotron-3-Nano-30B-A3B tok/s | 345.93 | 316.51 | +9% |
| Qwen3 30B A3B tok/s | 270.84 | 283.78 | -5% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 229.58 | 242.6 | -5% |
| Qwen3-30B-A3B tok/s | 271.69 | 285.44 | -5% |
| Qwen3-Coder 30B A3B tok/s | 277.68 | 296.22 | -6% |
| DeepSeek-R1 Distill 32B tok/s | 78.55 | 75.8 | +4% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 78.48 | 73.48 | +7% |
| Olmo-3.1-32B-Think tok/s | 79.75 | 75.53 | +6% |
| QwQ 32B tok/s | 78.47 | 75.82 | +3% |
| Qwen2.5-32B tok/s | 78.55 | 73.5 | +7% |
| Qwen2.5-Coder 32B tok/s | 78.53 | 75.81 | +4% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 60.24 | 58.05 | +4% |
| Qwen3 32B tok/s | 78.56 | 74.07 | +6% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 79.35 | 76.76 | +3% |
| Ornith-1.0-35B tok/s | 221.13 | 221.65 | -0% |
| Qwen-AgentWorld-35B-A3B tok/s | 221.78 | 219.1 | +1% |
| GLM-4.7-Flash tok/s | 183.82 | 186.05 | -1% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 230.68 | 231.83 | -0% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 44.49 | 41.17 | +8% |
| Hermes-4-70B tok/s | 44.49 | 41.2 | +8% |
| Llama 3.3 70B tok/s | 44.54 | 41 | +9% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 44.48 | 41.15 | +8% |
| Meta-Llama-3.1-70B tok/s | 44.48 | 41.19 | +8% |
| Qwen2.5-72B tok/s | 45.34 | 40.78 | +11% |
| Qwen3-Next-80B-A3B tok/s | 179.71 | 186.22 | -3% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 186.03 | 188.14 | -1% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 188.78 | 191.08 | -1% |
| Qwen3-Coder-Next tok/s | 186.39 | 188.21 | -1% |
| Qwen3-Coder-Next tok/s | 189.76 | 193.79 | -2% |
| Qwen3-Coder-Next-abliterated tok/s | 186.57 | 190.32 | -2% |
| BiRefNet images/min | 1082.03 | 1310.97 | -17% |
| Swin2SR 4x Upscaler images/min | 11.97 | 25.48 | -53% |
| Stable Diffusion XL images/min | 46.12 | 34.58 | +33% |
| Z-Image Turbo (1024px) images/min | 62.47 | 39.42 | +58% |
| Z-Image Turbo images/min | 32.325 | 22.125 | +46% |
| FLUX.1 Schnell images/min | 81.41 | 58.36 | +39% |
| FLUX.1 Kontext dev images/min | 5.721 | 4.264 | +34% |
| FLUX.1 dev images/min | 13.136 | 9.064 | +45% |
| Krea 2 Turbo images/min | 9.5 | 6.85 | +39% |
| Qwen-Image-Edit images/min | 4.86 | 3.68 | +32% |
| Wan 2.2 5B (720p) frames/s | 1.88 | 1.33 | +41% |
| LTX-Video (distilled) frames/s | 26.8 | 17.47 | +53% |
| Depth Anything V2 Small images/min | 1137.49 | 1081.54 | +5% |
| SAM ViT-Base images/min | 1662.34 | 1431.27 | +16% |
| Depth Anything V2 Large images/min | 1039.47 | 918.71 | +13% |
| SAM ViT-Huge images/min | 388.61 | 320.64 | +21% |
| TinyLlama 1.1B LoRA train tok/s | 17993.5 | 17522 | +3% |
| TinyLlama 1.1B served serve tok/s | 11562.2 | 8336.7 | +39% |
| Qwen2.5 1.5B LoRA train tok/s | 15322 | 16034.9 | -4% |
| Qwen2.5 1.5B served serve tok/s | 9186.4 | 6794.8 | +35% |
| SmolLM2 1.7B LoRA train tok/s | 18064.7 | 18910.3 | -4% |
| SmolLM2 1.7B served serve tok/s | 9672.9 | 6509.9 | +49% |
| Qwen2.5 7B LoRA train tok/s | 13722 | 8514.3 | +61% |
| Qwen2.5 7B served serve tok/s | 5799.1 | 3601.1 | +61% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA B200 is faster on 12 of 14; NVIDIA H100 80GB HBM3 on 2.
| Workflow | NVIDIA B200 | NVIDIA H100 80GB HBM3 | Difference | Cost per run |
|---|---|---|---|---|
| Depth pass on a batch | 6 s | 5 s | NVIDIA H100 80GB HBM3 1.04x faster | $0.009 vs $0.004 |
| Masking run | 85 s | 1.6 min | NVIDIA B200 1.14x faster | $0.141 vs $0.072 |
| 24-frame storyboard | 89 s | 1.8 min | NVIDIA B200 1.22x faster | $0.149 vs $0.081 |
| 60-second AI short film | 1.9 min | 2.5 min | NVIDIA B200 1.33x faster | $0.186 vs $0.111 |
| 6-panel comic page | 2.3 min | 3 min | NVIDIA B200 1.27x faster | $0.233 vs $0.133 |
| Character sheet, 12 poses | 2.9 min | 3.7 min | NVIDIA B200 1.28x faster | $0.285 vs $0.164 |
| Short social clips | 5.6 min | 8 min | NVIDIA B200 1.42x faster | $0.558 vs $0.358 |
| Full codebase review | 6.6 min | 6.9 min | NVIDIA B200 1.04x faster | $0.660 vs $0.310 |
| Product photo shoot | 8.4 min | 11.1 min | NVIDIA B200 1.33x faster | $0.836 vs $0.499 |
| Long-form article batch | 10.5 min | 11.4 min | NVIDIA B200 1.09x faster | $1.044 vs $0.510 |
| Product shoot, start to finish | 12.1 min | 12.9 min | NVIDIA B200 1.07x faster | $1.201 vs $0.577 |
| Product catalogue cutout | 17.2 min | 8.1 min | NVIDIA H100 80GB HBM3 2.11x faster | $1.712 vs $0.365 |
| Photo restoration batch | 17.8 min | 23.8 min | NVIDIA B200 1.34x faster | $1.773 vs $1.068 |
| Restore and enlarge photos | 26.4 min | 27.8 min | NVIDIA B200 1.05x faster | $2.628 vs $1.245 |
Renting by the hour, NVIDIA H100 80GB HBM3 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 H100 80GB HBM3 | $2.690 |
NVIDIA H100 80GB HBM3 is 2.22x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA B200 | NVIDIA H100 80GB HBM3 | |
|---|---|---|
| VRAM | 192GB | 80GB |
| Transistors | 208,000M | 80,000M |
| Process node | 4 nm | 4 nm |
| Architecture | Blackwell | Hopper |
| Memory bandwidth | 8 TB/s | 3350 GB/s |
| Boost clock | n/a | 1,980 MHz |
| TDP | 1000 W | 700 W |
| Launch MSRP | $40,000 | $30,000 |
| Release | 2025-02-01 | 2022-09-20 |
FAQ
- Which is better for ai & machine learning: NVIDIA B200 or NVIDIA H100 80GB HBM3?
- NVIDIA B200 performs better for ai & machine learning, winning 95 of 147 benchmarks in our suite with an average 5.2% advantage.
- What are the main hardware differences between NVIDIA B200 and NVIDIA H100 80GB HBM3?
- NVIDIA B200 has 192GB VRAM and a 1000W TDP, while NVIDIA H100 80GB HBM3 has 80GB VRAM and a 700W TDP.
- Where is the biggest performance difference between NVIDIA B200 and NVIDIA H100 80GB HBM3?
- Swin2SR 4x Upscaler: NVIDIA H100 80GB HBM3 leads by roughly 113% (11.97 vs 25.48 images/min) in our testing.
NVIDIA B200 full review · NVIDIA H100 80GB HBM3 full review · All AI & Machine Learning rankings