NVIDIA A100 80GB SXM4 vs NVIDIA H200, AI & Machine Learning Comparison
NVIDIA H200 wins 149 of 152 benchmarks, averaging 70.8% 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 H200 leads by 153% (0.74 vs 1.87 x realtime); the closest fight is Swin2SR 4x Upscaler (11% apart).
Benchmark results head-to-head
| Benchmark | NVIDIA A100 80GB SXM4 | NVIDIA H200 | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 197 | 318.84 | -38% |
| Llama 3.1 8B tok/s | 162.57 | 268.31 | -39% |
| Qwen2.5-Coder 14B tok/s | 89.45 | 148.43 | -40% |
| Qwen3 32B tok/s | 45.53 | 76.58 | -41% |
| Llama 3.3 70B tok/s | 24.4 | 42.66 | -43% |
| Stable Diffusion XL images/min | 16.36 | 37.16 | -56% |
| Z-Image Turbo images/min | 9.975 | 23.175 | -57% |
| FLUX.1 dev images/min | 4.264 | 9.514 | -55% |
| FLUX.1 Kontext dev images/min | 1.993 | 4.393 | -55% |
| Qwen-Image-Edit images/min | 1.64 | 3.76 | -56% |
| LTX-Video (distilled) frames/s | 8.95 | 17.87 | -50% |
| Wan 2.2 5B (720p) frames/s | 0.66 | 1.41 | -53% |
| DeepSeek-R1 Distill Llama 8B tok/s | 159.37 | 265.25 | -40% |
| DeepSeek-R1 Distill 1.5B tok/s | 327.66 | 542.95 | -40% |
| DeepSeek-R1 Distill 14B tok/s | 87 | 146.82 | -41% |
| DeepSeek-R1 Distill 7B tok/s | 162.96 | 267.1 | -39% |
| Gemma 3 12B tok/s | 91.45 | 153.26 | -40% |
| Gemma 3 4B tok/s | 176.94 | 291.76 | -39% |
| Gemma 4 12B tok/s | 92.43 | 151.05 | -39% |
| Llama 3.2 1B tok/s | 525.32 | 875.53 | -40% |
| Llama 3.2 3B tok/s | 264.25 | 424.61 | -38% |
| Mistral 7B v0.3 tok/s | 171.92 | 278.5 | -38% |
| Mistral Small 24B tok/s | 61.38 | 107.84 | -43% |
| Phi-4 14B tok/s | 95.22 | 168.2 | -43% |
| Phi-4 Mini 3.8B tok/s | 239.17 | 392.38 | -39% |
| Qwen2.5-Coder 7B tok/s | 164.6 | 266.36 | -38% |
| Qwen3 0.6B tok/s | 428.39 | 718.67 | -40% |
| Qwen3 1.7B tok/s | 343.58 | 563.52 | -39% |
| Qwen3 14B tok/s | 90.64 | 154.51 | -41% |
| Qwen3 30B A3B tok/s | 177.32 | 292.11 | -39% |
| Qwen3 8B tok/s | 149.96 | 247.98 | -40% |
| SmolLM3 3B tok/s | 249.76 | 402.24 | -38% |
| TRELLIS Image-to-3D assets/hour | 488.7 | 818.8 | -40% |
| Codestral 22B tok/s | 63.26 | 111.34 | -43% |
| DeepSeek-R1 Distill 32B tok/s | 43.11 | 75.8 | -43% |
| Devstral Small 24B tok/s | 61.63 | 109.05 | -43% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 43.47 | 76.7 | -43% |
| Dolphin Mistral 24B Venice tok/s | 62.72 | 109.14 | -43% |
| Dolphin X1 8B tok/s | 158.62 | 268.36 | -41% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 158.63 | 267.84 | -41% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 61.17 | 109.06 | -44% |
| Gemma 3 27B tok/s | 48.43 | 84.75 | -43% |
| Qwen2.5-Coder 32B tok/s | 43.36 | 75.77 | -43% |
| Qwen3-Coder 30B A3B tok/s | 182.26 | 297.9 | -39% |
| QwQ 32B tok/s | 43.27 | 75.77 | -43% |
| StarCoder2 15B tok/s | 79.22 | 136.29 | -42% |
| FLUX.1 Schnell images/min | 28.39 | 60.51 | -53% |
| Z-Image Turbo (1024px) images/min | 18.65 | 40.74 | -54% |
| Krea 2 Turbo images/min | 3.35 | 7.25 | -54% |
| TRELLIS.2 Image-to-3D (1536³ max quality) assets/hour | 32.7 | 65.8 | -50% |
| BiRefNet images/min | 916.67 | 1457.73 | -37% |
| Depth Anything V2 Large images/min | 876.29 | 1040.81 | -16% |
| Depth Anything V2 Small images/min | 925.15 | 1198.58 | -23% |
| SAM ViT-Base images/min | 744.79 | 1469.34 | -49% |
| SAM ViT-Huge images/min | 145.7 | 325.19 | -55% |
| Swin2SR 4x Upscaler images/min | 28.72 | 25.85 | +11% |
| Qwen2.5 1.5B LoRA train tok/s | 8252.8 | 14931.7 | -45% |
| Qwen2.5 7B LoRA train tok/s | 3900.6 | 8855.4 | -56% |
| SmolLM2 1.7B LoRA train tok/s | 9750.8 | 17650.4 | -45% |
| TinyLlama 1.1B LoRA train tok/s | 9553.3 | 16222.4 | -41% |
| Qwen2.5 1.5B served serve tok/s | 4564.7 | 7541.6 | -39% |
| Qwen2.5 7B served serve tok/s | 2187.2 | 4394.8 | -50% |
| SmolLM2 1.7B served serve tok/s | 4351.4 | 7107.8 | -39% |
| TinyLlama 1.1B served serve tok/s | 5701.5 | 9137.7 | -38% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 22.99 | 42.76 | -46% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 43.48 | 75.71 | -43% |
| Kokoro TTS 82M x realtime | 110.17 | 170.82 | -36% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 22.92 | 42.75 | -46% |
| Meta-Llama-3.1-70B tok/s | 22.92 | 42.72 | -46% |
| MusicGen Small x realtime | 0.74 | 1.87 | -60% |
| Nemotron-3-Nano-30B-A3B tok/s | 204.06 | 327.48 | -38% |
| Whisper large-v3 x realtime | 73.83 | 163.78 | -55% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 31.99 | 58.87 | -46% |
| AI21-Jamba-Reasoning-3B tok/s | 239.23 | 370.66 | -35% |
| Olmo-3.1-32B-Think tok/s | 46.49 | 78.46 | -41% |
| Codestral 22B (Q3_K_M) tok/s | 49.52 | 88.03 | -44% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 65.77 | 109.2 | -40% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 164.46 | 267.82 | -39% |
| DeepSeek-Coder-V2-Lite tok/s | 213.91 | 312.1 | -31% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 152.76 | 250.61 | -39% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 67.13 | 119.66 | -44% |
| dolphin-2.9-llama3-8b tok/s | 164.09 | 267.96 | -39% |
| Dolphin X1 Trinity Nano 6B tok/s | 158.41 | 259.81 | -39% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 90.35 | 148.74 | -39% |
| gemma-2-2b-it-abliterated tok/s | 244.41 | 397.43 | -39% |
| gemma-2-2b tok/s | 244.52 | 398.48 | -39% |
| gemma-2-9b tok/s | 108.68 | 180.47 | -40% |
| Gemma 3 12B (Q3_K_M) tok/s | 71.88 | 127.04 | -43% |
| gemma-3-1b tok/s | 324.08 | 513.58 | -37% |
| gemma-3-270m tok/s | 617.58 | 967.35 | -36% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 112.75 | 171.45 | -34% |
| GLM-4.7-Flash tok/s | 124.08 | 188.57 | -34% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 152.29 | 250.88 | -39% |
| gpt-oss-20b tok/s | 212.63 | 354.23 | -40% |
| Hermes-3-Llama-3.2-3B tok/s | 269.13 | 430.79 | -38% |
| Hermes-4-70B tok/s | 24.47 | 42.77 | -43% |
| SmolLM3-3B tok/s | 254.6 | 404.27 | -37% |
| Qwen3-Coder-Next-abliterated tok/s | 119.56 | 197.14 | -39% |
| KAT-Coder-V2.5-Dev tok/s | 146.38 | 236.31 | -38% |
| L3-8B-Stheno-v3.2 tok/s | 164.51 | 267.92 | -39% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 576.95 | 946.53 | -39% |
| LFM2.5-8B-A1B tok/s | 369.01 | 583.96 | -37% |
| Llama-2-7B tok/s | 180.09 | 290.81 | -38% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 269.13 | 429.78 | -37% |
| Meta-Llama-3.1-8B tok/s | 164.24 | 262.93 | -38% |
| Phi-4-mini tok/s | 243.38 | 395.3 | -38% |
| Mistral-7B-Instruct-v0.1 tok/s | 173.87 | 282.64 | -38% |
| Mistral-7B-Instruct-v0.2 tok/s | 173.8 | 282.32 | -38% |
| Mistral-7B-Instruct-v0.3 tok/s | 173.27 | 282.39 | -39% |
| Mistral-Nemo-Instruct-2407 tok/s | 111.07 | 181.3 | -39% |
| Mistral Small 24B (Q3_K_M) tok/s | 46.63 | 85.1 | -45% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 110.99 | 181.53 | -39% |
| Hermes-4-14B tok/s | 94.48 | 156.26 | -40% |
| Ornith-1.0-35B tok/s | 138.8 | 208.51 | -33% |
| Ornith-1.0-9B tok/s | 134.67 | 222.94 | -40% |
| phi-2 tok/s | 235.55 | 347.08 | -32% |
| Phi-3.5-mini tok/s | 229.46 | 348.14 | -34% |
| Phi-4 14B (Q3_K_M) tok/s | 82.68 | 139.97 | -41% |
| Qwen-AgentWorld-35B-A3B tok/s | 139.69 | 225.31 | -38% |
| Qwen3-0.6B tok/s | 432.66 | 724.67 | -40% |
| Qwen3-1.7B tok/s | 350.4 | 568.85 | -38% |
| Qwen3-14B tok/s | 94.35 | 156.3 | -40% |
| Qwen3-30B-A3B tok/s | 178.14 | 293.47 | -39% |
| Qwen3-4B-Instruct-2507 tok/s | 199.39 | 318.67 | -37% |
| Qwen3-8B tok/s | 152.8 | 251.04 | -39% |
| Qwen3-Coder-Next tok/s | 118.62 | 195.21 | -39% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 118.93 | 194.82 | -39% |
| Qwen1.5-0.5B tok/s | 540.32 | 855.22 | -37% |
| Qwen2-1.5B tok/s | 333.79 | 546.47 | -39% |
| Qwen2.5-0.5B tok/s | 571.49 | 914.58 | -38% |
| Qwen2.5-1.5B tok/s | 335.29 | 549.36 | -39% |
| Uncensored tok/s | 89.43 | 148.7 | -40% |
| Qwen2.5-14B tok/s | 90.46 | 148.76 | -39% |
| Qwen2.5-32B tok/s | 45.6 | 75.76 | -40% |
| Qwen2.5-3B tok/s | 248.47 | 400.48 | -38% |
| Qwen2.5-72B tok/s | 23.95 | 43.4 | -45% |
| Qwen2.5-7B tok/s | 166.92 | 265.15 | -37% |
| Qwen2.5-Coder-0.5B tok/s | 562.59 | 920.36 | -39% |
| Qwen2.5-Coder-1.5B tok/s | 333.87 | 545.54 | -39% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 90.28 | 148.48 | -39% |
| Qwen2.5-Coder-3B tok/s | 248.7 | 401.11 | -38% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 166.84 | 270.92 | -38% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 146.72 | 246.35 | -40% |
| Qwen3-4B-Instruct-2507 tok/s | 199.65 | 318.27 | -37% |
| Qwen3-4B-Thinking-2507 tok/s | 199.71 | 318.94 | -37% |
| Qwen3-Coder-Next tok/s | 120.71 | 196.65 | -39% |
| Qwen3-Next-80B-A3B-Thinking tok/s | 120.98 | 201.12 | -40% |
| Qwen3-Next-80B-A3B tok/s | 117.44 | 192.52 | -39% |
| SmolLM2-135M tok/s | 554.44 | 905.38 | -39% |
| Cydonia-24B-v4.3 tok/s | 65.82 | 109.2 | -40% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA A100 80GB SXM4 is faster on 3 of 17; NVIDIA H200 on 14.
| Workflow | NVIDIA A100 80GB SXM4 | NVIDIA H200 | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 5 s | 6 s | NVIDIA A100 80GB SXM4 1.24x faster | $0.002 vs $0.006 |
| 30-minute podcast pass | 67 s | 46 s | NVIDIA H200 1.47x faster | $0.026 vs $0.046 |
| 500-image masking run | 3.5 min | 1.7 min | NVIDIA H200 2.09x faster | $0.081 vs $0.100 |
| 24-frame storyboard | 3.7 min | 3.9 min | NVIDIA A100 80GB SXM4 1.06x faster | $0.085 vs $0.234 |
| 20-asset 3D game kit | 4 min | 2.4 min | NVIDIA H200 1.65x faster | $0.092 vs $0.145 |
| 60-second AI short film | 4.7 min | 3.9 min | NVIDIA H200 1.20x faster | $0.108 vs $0.233 |
| 60-second AI short film, narrated | 4.9 min | 4.1 min | NVIDIA H200 1.21x faster | $0.114 vs $0.243 |
| 6-panel comic page | 6 min | 3.2 min | NVIDIA H200 1.86x faster | $0.139 vs $0.192 |
| 200-product catalogue cutout | 7.3 min | 8 min | NVIDIA A100 80GB SXM4 1.10x faster | $0.169 vs $0.479 |
| Character sheet, 12 poses | 7.6 min | 3.9 min | NVIDIA H200 1.94x faster | $0.175 vs $0.234 |
| Full codebase review | 11.2 min | 6.7 min | NVIDIA H200 1.66x faster | $0.259 vs $0.403 |
| 10 short social clips | 15.1 min | 10 min | NVIDIA H200 1.51x faster | $0.350 vs $0.598 |
| 20 long-form articles | 19.1 min | 10.9 min | NVIDIA H200 1.75x faster | $0.443 vs $0.655 |
| 40-product photo shoot | 23.5 min | 11.2 min | NVIDIA H200 2.10x faster | $0.543 vs $0.668 |
| 40-product shoot, start to finish | 25 min | 12.9 min | NVIDIA H200 1.94x faster | $0.579 vs $0.770 |
| 100-photo restoration batch | 50.8 min | 23.3 min | NVIDIA H200 2.18x faster | $1.177 vs $1.396 |
| 100-photo restore and enlarge | 54.3 min | 27.2 min | NVIDIA H200 1.99x faster | $1.258 vs $1.629 |
Renting by the hour, NVIDIA A100 80GB SXM4 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 A100 80GB SXM4 | $1.390 |
| NVIDIA H200 | $3.590 |
NVIDIA A100 80GB SXM4 is 2.58x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA A100 80GB SXM4 | NVIDIA H200 | |
|---|---|---|
| VRAM | 80GB | 141GB |
| Architecture | Ampere | Hopper |
| Memory bandwidth | 2039 GB/s | 4800 GB/s |
| Boost clock | 1,410 MHz | 1,980 MHz |
| TDP | 400 W | 700 W |
| Launch MSRP | $17,000 | $31,000 |
| Release | 2020-11-16 | 2024-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA A100 80GB SXM4 or NVIDIA H200?
- NVIDIA H200 performs better for ai & machine learning, winning 149 of 152 benchmarks in our suite with an average 70.8% advantage.
- What are the main hardware differences between NVIDIA A100 80GB SXM4 and NVIDIA H200?
- NVIDIA A100 80GB SXM4 has 80GB VRAM and a 400W TDP, while NVIDIA H200 has 141GB VRAM and a 700W TDP.
- Where is the biggest performance difference between NVIDIA A100 80GB SXM4 and NVIDIA H200?
- MusicGen Small: NVIDIA H200 leads by roughly 113% (0.74 vs 1.87 x realtime) in our testing.
NVIDIA A100 80GB SXM4 full review · NVIDIA H200 full review · All AI & Machine Learning rankings