NVIDIA H100 80GB HBM3 vs NVIDIA H200, AI & Machine Learning Comparison
NVIDIA H200 wins 136 of 152 benchmarks, averaging 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 MusicGen Small, where NVIDIA H100 80GB HBM3 leads by 32% (2.47 vs 1.87 x realtime); the closest fight is DeepSeek-R1 Distill 32B (0% apart).
Benchmark results head-to-head
| Benchmark | NVIDIA H100 80GB HBM3 | NVIDIA H200 | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 310.26 | 318.84 | -3% |
| Llama 3.1 8B tok/s | 261.83 | 268.31 | -2% |
| Qwen2.5-Coder 14B tok/s | 144.84 | 148.43 | -2% |
| Qwen3 32B tok/s | 74.07 | 76.58 | -3% |
| Llama 3.3 70B tok/s | 41 | 42.66 | -4% |
| Stable Diffusion XL images/min | 34.58 | 37.16 | -7% |
| Z-Image Turbo images/min | 22.125 | 23.175 | -5% |
| FLUX.1 dev images/min | 9.064 | 9.514 | -5% |
| FLUX.1 Kontext dev images/min | 4.264 | 4.393 | -3% |
| Qwen-Image-Edit images/min | 3.68 | 3.76 | -2% |
| LTX-Video (distilled) frames/s | 17.47 | 17.87 | -2% |
| Wan 2.2 5B (720p) frames/s | 1.33 | 1.41 | -6% |
| DeepSeek-R1 Distill Llama 8B tok/s | 261.15 | 265.25 | -2% |
| DeepSeek-R1 Distill 1.5B tok/s | 537.32 | 542.95 | -1% |
| DeepSeek-R1 Distill 7B tok/s | 264.42 | 267.1 | -1% |
| Gemma 3 12B tok/s | 150.58 | 153.26 | -2% |
| Gemma 3 4B tok/s | 289.61 | 291.76 | -1% |
| Llama 3.2 1B tok/s | 880.64 | 875.53 | +1% |
| Llama 3.2 3B tok/s | 422.87 | 424.61 | -0% |
| Mistral 7B v0.3 tok/s | 275.85 | 278.5 | -1% |
| Phi-4 Mini 3.8B tok/s | 388.86 | 392.38 | -1% |
| Qwen2.5-Coder 7B tok/s | 263.91 | 266.36 | -1% |
| Qwen3 0.6B tok/s | 713.53 | 718.67 | -1% |
| Qwen3 1.7B tok/s | 556.48 | 563.52 | -1% |
| Qwen3 14B tok/s | 151.16 | 154.51 | -2% |
| Qwen3 8B tok/s | 244.22 | 247.98 | -2% |
| SmolLM3 3B tok/s | 397.99 | 402.24 | -1% |
| DeepSeek-R1 Distill 14B tok/s | 144.82 | 146.82 | -1% |
| Gemma 4 12B tok/s | 148.17 | 151.05 | -2% |
| Mistral Small 24B tok/s | 105.62 | 107.84 | -2% |
| Phi-4 14B tok/s | 165.17 | 168.2 | -2% |
| Qwen3 30B A3B tok/s | 283.78 | 292.11 | -3% |
| TRELLIS Image-to-3D assets/hour | 864 | 818.8 | +6% |
| Codestral 22B tok/s | 110.05 | 111.34 | -1% |
| DeepSeek-R1 Distill 32B tok/s | 75.8 | 75.8 | 0% |
| Devstral Small 24B tok/s | 107.79 | 109.05 | -1% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 76.76 | 76.7 | 0% |
| Dolphin Mistral 24B Venice tok/s | 109.13 | 109.14 | -0% |
| Dolphin X1 8B tok/s | 266 | 268.36 | -1% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 266.19 | 267.84 | -1% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 107.83 | 109.06 | -1% |
| Gemma 3 27B tok/s | 84.77 | 84.75 | 0% |
| Qwen2.5-Coder 32B tok/s | 75.81 | 75.77 | 0% |
| Qwen3-Coder 30B A3B tok/s | 296.22 | 297.9 | -1% |
| QwQ 32B tok/s | 75.82 | 75.77 | 0% |
| StarCoder2 15B tok/s | 134.92 | 136.29 | -1% |
| FLUX.1 Schnell images/min | 58.36 | 60.51 | -4% |
| Z-Image Turbo (1024px) images/min | 39.42 | 40.74 | -3% |
| Krea 2 Turbo images/min | 6.85 | 7.25 | -6% |
| TRELLIS.2 Image-to-3D (1536³ max quality) assets/hour | 62.7 | 65.8 | -5% |
| BiRefNet images/min | 1310.97 | 1457.73 | -10% |
| Depth Anything V2 Large images/min | 918.71 | 1040.81 | -12% |
| Depth Anything V2 Small images/min | 1081.54 | 1198.58 | -10% |
| SAM ViT-Base images/min | 1431.27 | 1469.34 | -3% |
| SAM ViT-Huge images/min | 320.64 | 325.19 | -1% |
| Swin2SR 4x Upscaler images/min | 25.48 | 25.85 | -1% |
| Kokoro TTS 82M x realtime | 222.02 | 170.82 | +30% |
| MusicGen Small x realtime | 2.47 | 1.87 | +32% |
| Whisper large-v3 x realtime | 181.47 | 163.78 | +11% |
| Qwen2.5 1.5B LoRA train tok/s | 16034.9 | 14931.7 | +7% |
| Qwen2.5 7B LoRA train tok/s | 8514.3 | 8855.4 | -4% |
| SmolLM2 1.7B LoRA train tok/s | 18910.3 | 17650.4 | +7% |
| TinyLlama 1.1B LoRA train tok/s | 17522 | 16222.4 | +8% |
| Qwen2.5 1.5B served serve tok/s | 6794.8 | 7541.6 | -10% |
| Qwen2.5 7B served serve tok/s | 3601.1 | 4394.8 | -18% |
| SmolLM2 1.7B served serve tok/s | 6509.9 | 7107.8 | -8% |
| TinyLlama 1.1B served serve tok/s | 8336.7 | 9137.7 | -9% |
| AI21-Jamba-Reasoning-3B tok/s | 363.09 | 370.66 | -2% |
| Olmo-3.1-32B-Think tok/s | 75.53 | 78.46 | -4% |
| Codestral 22B (Q3_K_M) tok/s | 87.21 | 88.03 | -1% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 105.63 | 109.2 | -3% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 262.76 | 267.82 | -2% |
| DeepSeek-Coder-V2-Lite tok/s | 309.14 | 312.1 | -1% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 245.26 | 250.61 | -2% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 41.17 | 42.76 | -4% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 118.43 | 119.66 | -1% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 73.48 | 75.71 | -3% |
| dolphin-2.9-llama3-8b tok/s | 262.47 | 267.96 | -2% |
| Dolphin X1 Trinity Nano 6B tok/s | 247.94 | 259.81 | -5% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 145.15 | 148.74 | -2% |
| gemma-2-2b-it-abliterated tok/s | 392.65 | 397.43 | -1% |
| gemma-2-2b tok/s | 392.26 | 398.48 | -2% |
| gemma-2-9b tok/s | 174.7 | 180.47 | -3% |
| Gemma 3 12B (Q3_K_M) tok/s | 125.92 | 127.04 | -1% |
| gemma-3-1b tok/s | 504.17 | 513.58 | -2% |
| gemma-3-270m tok/s | 961.29 | 967.35 | -1% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 168.41 | 171.45 | -2% |
| GLM-4.7-Flash tok/s | 186.05 | 188.57 | -1% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 245 | 250.88 | -2% |
| gpt-oss-20b tok/s | 346.8 | 354.23 | -2% |
| Hermes-3-Llama-3.2-3B tok/s | 423.72 | 430.79 | -2% |
| Hermes-4-70B tok/s | 41.2 | 42.77 | -4% |
| SmolLM3-3B tok/s | 399.14 | 404.27 | -1% |
| Qwen3-Coder-Next-abliterated tok/s | 190.32 | 197.14 | -3% |
| KAT-Coder-V2.5-Dev tok/s | 231.83 | 236.31 | -2% |
| L3-8B-Stheno-v3.2 tok/s | 261.83 | 267.92 | -2% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 926.61 | 946.53 | -2% |
| LFM2.5-8B-A1B tok/s | 581.98 | 583.96 | -0% |
| Llama-2-7B tok/s | 284.62 | 290.81 | -2% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 424.25 | 429.78 | -1% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 41.15 | 42.75 | -4% |
| Meta-Llama-3.1-70B tok/s | 41.19 | 42.72 | -4% |
| Meta-Llama-3.1-8B tok/s | 261.7 | 262.93 | -0% |
| Phi-4-mini tok/s | 381.52 | 395.3 | -3% |
| Mistral-7B-Instruct-v0.1 tok/s | 276.29 | 282.64 | -2% |
| Mistral-7B-Instruct-v0.2 tok/s | 276.73 | 282.32 | -2% |
| Mistral-7B-Instruct-v0.3 tok/s | 276.61 | 282.39 | -2% |
| Mistral-Nemo-Instruct-2407 tok/s | 177.38 | 181.3 | -2% |
| Mistral Small 24B (Q3_K_M) tok/s | 84 | 85.1 | -1% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 177.28 | 181.53 | -2% |
| Nemotron-3-Nano-30B-A3B tok/s | 316.51 | 327.48 | -3% |
| Hermes-4-14B tok/s | 152 | 156.26 | -3% |
| Ornith-1.0-35B tok/s | 221.65 | 208.51 | +6% |
| Ornith-1.0-9B tok/s | 216.89 | 222.94 | -3% |
| phi-2 tok/s | 340.09 | 347.08 | -2% |
| Phi-3.5-mini tok/s | 345.07 | 348.14 | -1% |
| Phi-4 14B (Q3_K_M) tok/s | 138.87 | 139.97 | -1% |
| Qwen-AgentWorld-35B-A3B tok/s | 219.1 | 225.31 | -3% |
| Qwen3-0.6B tok/s | 710.97 | 724.67 | -2% |
| Qwen3-1.7B tok/s | 557.09 | 568.85 | -2% |
| Qwen3-14B tok/s | 152.2 | 156.3 | -3% |
| Qwen3-30B-A3B tok/s | 285.44 | 293.47 | -3% |
| Qwen3-4B-Instruct-2507 tok/s | 310.58 | 318.67 | -3% |
| Qwen3-8B tok/s | 244.2 | 251.04 | -3% |
| Qwen3-Coder-Next tok/s | 188.21 | 195.21 | -4% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 188.14 | 194.82 | -3% |
| Qwen1.5-0.5B tok/s | 847.02 | 855.22 | -1% |
| Qwen2-1.5B tok/s | 536.52 | 546.47 | -2% |
| Qwen2.5-0.5B tok/s | 890 | 914.58 | -3% |
| Qwen2.5-1.5B tok/s | 537.24 | 549.36 | -2% |
| Uncensored tok/s | 144.92 | 148.7 | -3% |
| Qwen2.5-14B tok/s | 144.7 | 148.76 | -3% |
| Qwen2.5-32B tok/s | 73.5 | 75.76 | -3% |
| Qwen2.5-3B tok/s | 395.52 | 400.48 | -1% |
| Qwen2.5-72B tok/s | 40.78 | 43.4 | -6% |
| Qwen2.5-7B tok/s | 264.17 | 265.15 | -0% |
| Qwen2.5-Coder-0.5B tok/s | 892.81 | 920.36 | -3% |
| Qwen2.5-Coder-1.5B tok/s | 538.44 | 545.54 | -1% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 145.01 | 148.48 | -2% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 58.05 | 58.87 | -1% |
| Qwen2.5-Coder-3B tok/s | 394 | 401.11 | -2% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 264.24 | 270.92 | -2% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 242.6 | 246.35 | -2% |
| Qwen3-4B-Instruct-2507 tok/s | 311.47 | 318.27 | -2% |
| Qwen3-4B-Thinking-2507 tok/s | 310.84 | 318.94 | -3% |
| Qwen3-Coder-Next tok/s | 193.79 | 196.65 | -1% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 191.08 | 201.12 | -5% |
| Qwen3-Next-80B-A3B tok/s | 186.22 | 192.52 | -3% |
| SmolLM2-135M tok/s | 898.53 | 905.38 | -1% |
| Cydonia-24B-v4.3 tok/s | 105.49 | 109.2 | -3% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA H100 80GB HBM3 is faster on 12 of 17; NVIDIA H200 on 5.
| Workflow | NVIDIA H100 80GB HBM3 | NVIDIA H200 | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 5 s | 6 s | NVIDIA H100 80GB HBM3 1.13x faster | $0.002 vs $0.006 |
| 30-minute podcast pass | 34 s | 46 s | NVIDIA H100 80GB HBM3 1.34x faster | $0.013 vs $0.046 |
| 500-image masking run | 1.6 min | 1.7 min | NVIDIA H100 80GB HBM3 1.03x faster | $0.036 vs $0.100 |
| 24-frame storyboard | 1.8 min | 3.9 min | NVIDIA H100 80GB HBM3 2.15x faster | $0.040 vs $0.234 |
| 20-asset 3D game kit | 2.2 min | 2.4 min | NVIDIA H100 80GB HBM3 1.11x faster | $0.049 vs $0.145 |
| 60-second AI short film | 2.5 min | 3.9 min | NVIDIA H100 80GB HBM3 1.57x faster | $0.055 vs $0.233 |
| 60-second AI short film, narrated | 2.6 min | 4.1 min | NVIDIA H100 80GB HBM3 1.54x faster | $0.059 vs $0.243 |
| 6-panel comic page | 3 min | 3.2 min | NVIDIA H100 80GB HBM3 1.08x faster | $0.066 vs $0.192 |
| Character sheet, 12 poses | 3.7 min | 3.9 min | NVIDIA H100 80GB HBM3 1.06x faster | $0.082 vs $0.234 |
| Full codebase review | 6.9 min | 6.7 min | NVIDIA H200 1.02x faster | $0.154 vs $0.403 |
| 10 short social clips | 8 min | 10 min | NVIDIA H100 80GB HBM3 1.25x faster | $0.178 vs $0.598 |
| 200-product catalogue cutout | 8.1 min | 8 min | NVIDIA H200 1.02x faster | $0.181 vs $0.479 |
| 40-product photo shoot | 11.1 min | 11.2 min | NVIDIA H100 80GB HBM3 1.00x faster | $0.248 vs $0.668 |
| 20 long-form articles | 11.4 min | 10.9 min | NVIDIA H200 1.04x faster | $0.253 vs $0.655 |
| 40-product shoot, start to finish | 12.9 min | 12.9 min | NVIDIA H100 80GB HBM3 1.00x faster | $0.287 vs $0.770 |
| 100-photo restoration batch | 23.8 min | 23.3 min | NVIDIA H200 1.02x faster | $0.531 vs $1.396 |
| 100-photo restore and enlarge | 27.8 min | 27.2 min | NVIDIA H200 1.02x faster | $0.618 vs $1.629 |
Renting by the hour, NVIDIA H100 80GB HBM3 finishes 17 of 17 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA H100 80GB HBM3 | $1.336 |
| NVIDIA H200 | $3.590 |
NVIDIA H100 80GB HBM3 is 2.69x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA H100 80GB HBM3 | NVIDIA H200 | |
|---|---|---|
| VRAM | 80GB | 141GB |
| Architecture | Hopper | Hopper |
| Memory bandwidth | 3350 GB/s | 4800 GB/s |
| Boost clock | 1,980 MHz | 1,980 MHz |
| TDP | 700 W | 700 W |
| Launch MSRP | $30,000 | $31,000 |
| Release | 2022-09-20 | 2024-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA H100 80GB HBM3 or NVIDIA H200?
- NVIDIA H200 performs better for ai & machine learning, winning 136 of 152 benchmarks in our suite with an average 2% advantage.
- What are the main hardware differences between NVIDIA H100 80GB HBM3 and NVIDIA H200?
- NVIDIA H100 80GB HBM3 has 80GB VRAM and a 700W TDP, while NVIDIA H200 has 141GB VRAM and a 700W TDP.
- Where is the biggest performance difference between NVIDIA H100 80GB HBM3 and NVIDIA H200?
- MusicGen Small: NVIDIA H100 80GB HBM3 leads by roughly 32% (2.47 vs 1.87 x realtime) in our testing.
NVIDIA H100 80GB HBM3 full review · NVIDIA H200 full review · All AI & Machine Learning rankings