NVIDIA A100 40GB SXM4 vs NVIDIA H200, AI & Machine Learning Comparison
NVIDIA H200 wins 147 of 149 benchmarks, averaging 77.7% faster.
NVIDIA A100 40GB SXM4's numbers are anchored estimates calibrated against our measured cards, pending first-party measurement. Treat small gaps as ties.
What the numbers say
The gap is widest in TinyLlama 1.1B served, where NVIDIA H200 leads by 184% (3221.3 vs 9137.7 serve tok/s); the closest fight is Kokoro TTS 82M (20% apart); VRAM decides part of this one: NVIDIA H200 runs 150 of our 12 AI workloads while the other card runs 135, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA A100 40GB SXM4 | NVIDIA H200 | Difference |
|---|---|---|---|
| Qwen3-4B tok/s | 193.14 | 318.84 | -39% |
| Llama-3.1-8B tok/s | 157.26 | 268.31 | -41% |
| Qwen2.5-Coder-14B tok/s | 86.2 | 148.43 | -42% |
| Qwen3-32B tok/s | 43.8 | 76.58 | -43% |
| Llama-3.3-70B tok/s | 0 | 42.66 | n/a |
| Stable Diffusion XL images/min | 18.576 | 37.16 | -50% |
| 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 | 0 | 3.76 | n/a |
| 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 | 157.16 | 265.25 | -41% |
| DeepSeek-R1 Distill 1.5B tok/s | 320.78 | 542.95 | -41% |
| DeepSeek-R1 Distill 14B tok/s | 86.39 | 146.82 | -41% |
| DeepSeek-R1 Distill 7B tok/s | 159.09 | 267.1 | -40% |
| Gemma 3 12B tok/s | 88.97 | 153.26 | -42% |
| Gemma 3 4B tok/s | 174.14 | 291.76 | -40% |
| Gemma 4 12B tok/s | 87.14 | 151.05 | -42% |
| Llama 3.2 1B tok/s | 532.08 | 875.53 | -39% |
| Llama 3.2 3B tok/s | 261.24 | 424.61 | -38% |
| Mistral 7B v0.3 tok/s | 165.9 | 278.5 | -40% |
| Mistral Small 24B tok/s | 62.42 | 107.84 | -42% |
| Phi-4 14B tok/s | 98.65 | 168.2 | -41% |
| Phi-4 Mini 3.8B tok/s | 236.71 | 392.38 | -40% |
| Qwen2.5-Coder 7B tok/s | 159.9 | 266.36 | -40% |
| Qwen3 0.6B tok/s | 417.88 | 718.67 | -42% |
| Qwen3 1.7B tok/s | 340.76 | 563.52 | -40% |
| Qwen3 14B tok/s | 90.01 | 154.51 | -42% |
| Qwen3 30B A3B tok/s | 169.1 | 292.11 | -42% |
| Qwen3 8B tok/s | 146.21 | 247.98 | -41% |
| SmolLM3 3B tok/s | 246.25 | 402.24 | -39% |
| Codestral 22B tok/s | 63.74 | 111.34 | -43% |
| DeepSeek-R1 Distill 32B tok/s | 43.13 | 75.8 | -43% |
| Devstral Small 24B tok/s | 62.4 | 109.05 | -43% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 42.97 | 76.7 | -44% |
| Dolphin Mistral 24B Venice tok/s | 62.46 | 109.14 | -43% |
| Dolphin X1 8B tok/s | 156.3 | 268.36 | -42% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 156.55 | 267.84 | -42% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 62.4 | 109.06 | -43% |
| Gemma 3 27B tok/s | 47.47 | 84.75 | -44% |
| Qwen2.5-Coder 32B tok/s | 43.14 | 75.77 | -43% |
| Qwen3-Coder 30B A3B tok/s | 173.47 | 297.9 | -42% |
| QwQ 32B tok/s | 43.02 | 75.77 | -43% |
| StarCoder2 15B tok/s | 80.25 | 136.29 | -41% |
| FLUX.1 Schnell images/min | 28.44 | 60.51 | -53% |
| Z-Image Turbo (1024px) images/min | 18.7 | 40.74 | -54% |
| BiRefNet images/min | 611.56 | 1457.73 | -58% |
| Depth Anything V2 Large images/min | 536.74 | 1040.81 | -48% |
| Depth Anything V2 Small images/min | 613.42 | 1198.58 | -49% |
| SAM ViT-Base images/min | 696.77 | 1469.34 | -53% |
| SAM ViT-Huge images/min | 146.81 | 325.19 | -55% |
| Swin2SR 4x Upscaler images/min | 17.25 | 25.85 | -33% |
| Kokoro TTS 82M x realtime | 142.03 | 170.82 | -17% |
| MusicGen Small x realtime | 1.08 | 1.87 | -42% |
| Whisper large-v3 x realtime | 102.47 | 163.78 | -37% |
| Dolphin X1 Trinity Nano 6B tok/s | 153.27 | 259.81 | -41% |
| gpt-oss-20b tok/s | 205.1 | 354.23 | -42% |
| Olmo-3.1-32B-Think tok/s | 43.67 | 78.46 | -44% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 62.56 | 109.2 | -43% |
| DeepSeek-Coder-V2-Lite tok/s | 205.42 | 312.1 | -34% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 145.84 | 250.61 | -42% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 86.48 | 148.74 | -42% |
| Qwen2.5 1.5B LoRA train tok/s | 6349.7 | 14931.7 | -57% |
| Qwen2.5 7B LoRA train tok/s | 3670.8 | 8855.4 | -59% |
| SmolLM2 1.7B LoRA train tok/s | 6938.8 | 17650.4 | -61% |
| TinyLlama 1.1B LoRA train tok/s | 7471 | 16222.4 | -54% |
| gemma-2-2b-it-abliterated tok/s | 232.27 | 397.43 | -42% |
| gemma-2-2b tok/s | 234.68 | 398.48 | -41% |
| gemma-2-9b tok/s | 101.97 | 180.47 | -43% |
| gemma-3-1b tok/s | 310.69 | 513.58 | -40% |
| gemma-3-270m tok/s | 574.24 | 967.35 | -41% |
| GLM-4.7-Flash tok/s | 119.1 | 188.57 | -37% |
| SmolLM3-3B tok/s | 247.66 | 404.27 | -39% |
| KAT-Coder-V2.5-Dev tok/s | 137.31 | 236.31 | -42% |
| LFM2.5-1.2B tok/s | 572.69 | 946.53 | -39% |
| Llama-2-7B tok/s | 172.53 | 290.81 | -41% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 260.83 | 429.78 | -39% |
| Meta-Llama-3.1-8B tok/s | 156.68 | 262.93 | -40% |
| Phi-4-mini tok/s | 237.38 | 395.3 | -40% |
| Mistral-7B-Instruct-v0.1 tok/s | 166.39 | 282.64 | -41% |
| Mistral-7B-Instruct-v0.2 tok/s | 166.8 | 282.32 | -41% |
| Mistral-7B-Instruct-v0.3 tok/s | 167.06 | 282.39 | -41% |
| Ornith-1.0-35B tok/s | 131.77 | 208.51 | -37% |
| Ornith-1.0-9B tok/s | 126.79 | 222.94 | -43% |
| Phi-3.5-mini tok/s | 222.53 | 348.14 | -36% |
| Qwen-AgentWorld-35B-A3B tok/s | 131.16 | 225.31 | -42% |
| Qwen3-0.6B tok/s | 408.39 | 724.67 | -44% |
| Qwen3-1.7B tok/s | 341.85 | 568.85 | -40% |
| Qwen3-14B tok/s | 90.07 | 156.3 | -42% |
| Qwen3-8B tok/s | 145.96 | 251.04 | -42% |
| Qwen2.5-0.5B tok/s | 517.09 | 914.58 | -43% |
| Qwen2.5-1.5B tok/s | 315.59 | 549.36 | -43% |
| Qwen2.5-14B tok/s | 86.07 | 148.76 | -42% |
| Qwen2.5-32B tok/s | 43.23 | 75.76 | -43% |
| Qwen2.5-3B tok/s | 240.48 | 400.48 | -40% |
| Qwen2.5-7B tok/s | 159.81 | 265.15 | -40% |
| Qwen2.5-Coder-1.5B tok/s | 322.96 | 545.54 | -41% |
| Qwen2.5-Coder-3B tok/s | 241.51 | 401.11 | -40% |
| AI21-Jamba-Reasoning-3B tok/s | 231.15 | 370.66 | -38% |
| Codestral 22B (Q3_K_M) tok/s | 48.37 | 88.03 | -45% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 156.62 | 267.82 | -42% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 65.33 | 119.66 | -45% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 43.07 | 75.71 | -43% |
| dolphin-2.9-llama3-8b tok/s | 153.34 | 267.96 | -43% |
| Gemma 3 12B (Q3_K_M) tok/s | 69.7 | 127.04 | -45% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 109.1 | 171.45 | -36% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 146.8 | 250.88 | -41% |
| Hermes-3-Llama-3.2-3B tok/s | 261 | 430.79 | -39% |
| L3-8B-Stheno-v3.2 tok/s | 156.55 | 267.92 | -42% |
| LFM2.5-8B-A1B tok/s | 351.47 | 583.96 | -40% |
| Mistral-Nemo-Instruct-2407 tok/s | 106.33 | 181.3 | -41% |
| Mistral Small 24B (Q3_K_M) tok/s | 45.34 | 85.1 | -47% |
| NemoMix-Unleashed-12B tok/s | 105.05 | 181.53 | -42% |
| Nemotron-3-Nano-30B-A3B tok/s | 192.44 | 327.48 | -41% |
| Hermes-4-14B tok/s | 89.97 | 156.26 | -42% |
| phi-2 tok/s | 228.1 | 347.08 | -34% |
| Phi-4 14B (Q3_K_M) tok/s | 80.71 | 139.97 | -42% |
| Qwen3-30B-A3B tok/s | 167.1 | 293.47 | -43% |
| Qwen3-4B-Instruct-2507 tok/s | 192.89 | 318.67 | -39% |
| Qwen1.5-0.5B tok/s | 497.73 | 855.22 | -42% |
| Qwen2-1.5B tok/s | 322.09 | 546.47 | -41% |
| Uncensored tok/s | 85.82 | 148.7 | -42% |
| Qwen2.5-Coder-0.5B tok/s | 542.36 | 920.36 | -41% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 86.34 | 148.48 | -42% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 31.04 | 58.87 | -47% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 159.76 | 270.92 | -41% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 142.78 | 246.35 | -42% |
| Qwen3-4B-Instruct-2507 tok/s | 193.39 | 318.27 | -39% |
| Qwen3-4B-Thinking-2507 tok/s | 189.89 | 318.94 | -40% |
| SmolLM2-135M tok/s | 544.06 | 905.38 | -40% |
| Cydonia-24B-v4.3 tok/s | 62.41 | 109.2 | -43% |
| Qwen2.5 1.5B served serve tok/s | 4111.1 | 7541.6 | -45% |
| Qwen2.5 7B served serve tok/s | 1769.6 | 4394.8 | -60% |
| SmolLM2 1.7B served serve tok/s | 3865.2 | 7107.8 | -46% |
| TinyLlama 1.1B served serve tok/s | 3221.3 | 9137.7 | -65% |
| Qwen3-Coder-Next-abliterated tok/s | 0 | 197.14 | n/a |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| Qwen3-Coder-Next tok/s | 0 | 195.21 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 194.82 | n/a |
| Qwen2.5-72B tok/s | 0 | 43.4 | n/a |
| Qwen3-Coder-Next tok/s | 0 | 196.65 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 201.12 | n/a |
| Qwen3-Next-80B-A3B tok/s | 0 | 192.52 | n/a |
| DeepSeek-R1-Distill-Llama-70B tok/s | 0 | 42.76 | n/a |
| Hermes-4-70B tok/s | 0 | 42.77 | n/a |
| Llama-3.3-70B-Instruct-abliterated tok/s | 0 | 42.75 | n/a |
| Meta-Llama-3.1-70B tok/s | 0 | 42.72 | n/a |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA A100 40GB SXM4 is faster on 1 of 15; NVIDIA H200 on 14.
| Workflow | NVIDIA A100 40GB SXM4 | NVIDIA H200 | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 10 s | 6 s | NVIDIA H200 1.56x faster | $0.003 vs $0.006 |
| 30-minute podcast pass | 59 s | 46 s | NVIDIA H200 1.29x faster | $0.016 vs $0.046 |
| 24-frame storyboard | 3 min | 3.9 min | NVIDIA A100 40GB SXM4 1.28x faster | $0.051 vs $0.234 |
| 500-image masking run | 3.5 min | 1.7 min | NVIDIA H200 2.09x faster | $0.058 vs $0.100 |
| 60-second AI short film | 4 min | 3.9 min | NVIDIA H200 1.02x faster | $0.066 vs $0.233 |
| 60-second AI short film, narrated | 4.2 min | 4.1 min | NVIDIA H200 1.03x faster | $0.070 vs $0.243 |
| 6-panel comic page | 4.8 min | 3.2 min | NVIDIA H200 1.49x faster | $0.080 vs $0.192 |
| Character sheet, 12 poses | 6.3 min | 3.9 min | NVIDIA H200 1.60x faster | $0.104 vs $0.234 |
| Full codebase review | 11.7 min | 6.7 min | NVIDIA H200 1.73x faster | $0.195 vs $0.403 |
| 200-product catalogue cutout | 12.2 min | 8 min | NVIDIA H200 1.52x faster | $0.203 vs $0.479 |
| 10 short social clips | 13.7 min | 10 min | NVIDIA H200 1.37x faster | $0.228 vs $0.598 |
| 40-product photo shoot | 22.2 min | 11.2 min | NVIDIA H200 1.99x faster | $0.370 vs $0.668 |
| 40-product shoot, start to finish | 24.8 min | 12.9 min | NVIDIA H200 1.93x faster | $0.414 vs $0.770 |
| 100-photo restoration batch | 50.2 min | 23.3 min | NVIDIA H200 2.15x faster | $0.836 vs $1.396 |
| 100-photo restore and enlarge | 56 min | 27.2 min | NVIDIA H200 2.06x faster | $0.934 vs $1.629 |
Renting by the hour, NVIDIA A100 40GB SXM4 finishes 15 of 15 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA A100 40GB SXM4 | $1.000 |
| NVIDIA H200 | $3.590 |
NVIDIA A100 40GB SXM4 is 3.59x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA A100 40GB SXM4 | NVIDIA H200 | |
|---|---|---|
| VRAM | 40GB | 141GB |
| Architecture | Ampere | Hopper |
| Memory bandwidth | 1555 GB/s | 4800 GB/s |
| Boost clock | 1,410 MHz | 1,980 MHz |
| TDP | 400 W | 700 W |
| Launch MSRP | $12,000 | $31,000 |
| Release | 2020-05-14 | 2024-03-18 |
FAQ
- Which is better for ai & machine learning: NVIDIA A100 40GB SXM4 or NVIDIA H200?
- NVIDIA H200 performs better for ai & machine learning, winning 147 of 149 benchmarks in our suite with an average 77.7% advantage.
- What are the main hardware differences between NVIDIA A100 40GB SXM4 and NVIDIA H200?
- NVIDIA A100 40GB SXM4 has 40GB VRAM and a 400W TDP, while NVIDIA H200 has 141GB VRAM and a 700W TDP.
- Does VRAM matter more than speed between NVIDIA A100 40GB SXM4 and NVIDIA H200?
- For AI, yes, NVIDIA H200 fits 13 more of our 12 workloads than NVIDIA A100 40GB SXM4. 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 A100 40GB SXM4 and NVIDIA H200?
- TinyLlama 1.1B served: NVIDIA H200 leads by roughly 184% (3221.3 vs 9137.7 serve tok/s) in our testing.
NVIDIA A100 40GB SXM4 full review · NVIDIA H200 full review · All AI & Machine Learning rankings