NVIDIA H200 vs NVIDIA T4, AI & Machine Learning Comparison
NVIDIA H200 wins 144 of 146 benchmarks, averaging 515.7% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Stable Diffusion XL, where NVIDIA H200 leads by 1475% (37.16 vs 2.36 images/min); the closest fight is Depth Anything V2 Small (96% apart); VRAM decides part of this one: NVIDIA H200 runs 150 of our 12 AI workloads while the other card runs 111, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA H200 | NVIDIA T4 | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 318.84 | 67.71 | +371% |
| Llama 3.1 8B tok/s | 268.31 | 35.01 | +666% |
| Qwen2.5-Coder 14B tok/s | 148.43 | 19.76 | +651% |
| Qwen3 32B tok/s | 76.58 | 0 | n/a |
| Llama 3.3 70B tok/s | 42.66 | 0 | n/a |
| Stable Diffusion XL images/min | 37.16 | 2.36 | +1475% |
| FLUX.1 dev images/min | 9.514 | 0 | n/a |
| FLUX.1 Kontext dev images/min | 4.393 | 0 | n/a |
| Qwen-Image-Edit images/min | 3.76 | 0 | n/a |
| Wan 2.2 5B (720p) frames/s | 1.41 | 0 | n/a |
| DeepSeek-R1 Distill Llama 8B tok/s | 265.25 | 36.36 | +630% |
| DeepSeek-R1 Distill 1.5B tok/s | 542.95 | 148.72 | +265% |
| DeepSeek-R1 Distill 14B tok/s | 146.82 | 19.21 | +664% |
| DeepSeek-R1 Distill 7B tok/s | 267.1 | 37.58 | +611% |
| Gemma 3 12B tok/s | 153.26 | 23.12 | +563% |
| Gemma 3 4B tok/s | 291.76 | 65.12 | +348% |
| Gemma 4 12B tok/s | 151.05 | 24.03 | +529% |
| Llama 3.2 1B tok/s | 875.53 | 207.71 | +322% |
| Llama 3.2 3B tok/s | 424.61 | 85.94 | +394% |
| Mistral 7B v0.3 tok/s | 278.5 | 40.01 | +596% |
| Mistral Small 24B tok/s | 107.84 | 11.33 | +852% |
| Phi-4 14B tok/s | 168.2 | 17.46 | +863% |
| Phi-4 Mini 3.8B tok/s | 392.38 | 64.15 | +512% |
| Qwen2.5-Coder 7B tok/s | 266.36 | 37.54 | +610% |
| Qwen3 0.6B tok/s | 718.67 | 263.47 | +173% |
| Qwen3 1.7B tok/s | 563.52 | 139.92 | +303% |
| Qwen3 14B tok/s | 154.51 | 19.31 | +700% |
| Qwen3 30B A3B tok/s | 292.11 | 0 | n/a |
| Qwen3 8B tok/s | 247.98 | 36.51 | +579% |
| SmolLM3 3B tok/s | 402.24 | 86.08 | +367% |
| Codestral 22B tok/s | 111.34 | 12.59 | +784% |
| DeepSeek-R1 Distill 32B tok/s | 75.8 | 0 | n/a |
| Devstral Small 24B tok/s | 109.05 | 11.4 | +857% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 76.7 | 0 | n/a |
| Dolphin Mistral 24B Venice tok/s | 109.14 | 11.36 | +861% |
| Dolphin X1 8B tok/s | 268.36 | 35.66 | +653% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 267.84 | 35.68 | +651% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 109.06 | 11.32 | +863% |
| Gemma 3 27B tok/s | 84.75 | 0 | n/a |
| Qwen2.5-Coder 32B tok/s | 75.77 | 0 | n/a |
| Qwen3-Coder 30B A3B tok/s | 297.9 | 0 | n/a |
| QwQ 32B tok/s | 75.77 | 0 | n/a |
| StarCoder2 15B tok/s | 136.29 | 15.66 | +770% |
| FLUX.1 Schnell images/min | 60.51 | 0 | n/a |
| BiRefNet images/min | 1457.73 | 183.48 | +694% |
| Depth Anything V2 Large images/min | 1040.81 | 402.34 | +159% |
| Depth Anything V2 Small images/min | 1198.58 | 612.25 | +96% |
| SAM ViT-Base images/min | 1469.34 | 180.73 | +713% |
| SAM ViT-Huge images/min | 325.19 | 34.91 | +832% |
| Swin2SR 4x Upscaler images/min | 25.85 | 10.53 | +145% |
| Qwen2.5 1.5B LoRA train tok/s | 14931.7 | 1434.4 | +941% |
| Qwen2.5 7B LoRA train tok/s | 8855.4 | 0 | n/a |
| SmolLM2 1.7B LoRA train tok/s | 17650.4 | 1561.9 | +1030% |
| TinyLlama 1.1B LoRA train tok/s | 16222.4 | 1520.1 | +967% |
| Qwen2.5 1.5B served serve tok/s | 7541.6 | 1337.1 | +464% |
| Qwen2.5 7B served serve tok/s | 4394.8 | 0 | n/a |
| SmolLM2 1.7B served serve tok/s | 7107.8 | 1222.5 | +481% |
| TinyLlama 1.1B served serve tok/s | 9137.7 | 1897.8 | +381% |
| Kokoro TTS 82M x realtime | 170.82 | 42.88 | +298% |
| MusicGen Small x realtime | 1.87 | 0.92 | +103% |
| Whisper large-v3 x realtime | 163.78 | 44.36 | +269% |
| AI21-Jamba-Reasoning-3B tok/s | 370.66 | 86.91 | +326% |
| Olmo-3.1-32B-Think tok/s | 78.46 | 0 | n/a |
| Codestral 22B (Q3_K_M) tok/s | 88.03 | 11.3 | +679% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 109.2 | 11.33 | +864% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 267.82 | 35.73 | +650% |
| DeepSeek-Coder-V2-Lite tok/s | 312.1 | 83.49 | +274% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 250.61 | 37.7 | +565% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 42.76 | 0 | n/a |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 119.66 | 16.86 | +610% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 75.71 | 0 | n/a |
| dolphin-2.9-llama3-8b tok/s | 267.96 | 35.1 | +663% |
| Dolphin X1 Trinity Nano 6B tok/s | 259.81 | 114.21 | +127% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 148.74 | 19.92 | +647% |
| gemma-2-2b-it-abliterated tok/s | 397.43 | 91.91 | +332% |
| gemma-2-2b tok/s | 398.48 | 92.13 | +333% |
| gemma-2-9b tok/s | 180.47 | 28.62 | +531% |
| Gemma 3 12B (Q3_K_M) tok/s | 127.04 | 19.79 | +542% |
| gemma-3-1b tok/s | 513.58 | 156.88 | +227% |
| gemma-3-270m tok/s | 967.35 | 362.8 | +167% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 171.45 | 53.83 | +219% |
| GLM-4.7-Flash tok/s | 188.57 | 0 | n/a |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 250.88 | 35.62 | +604% |
| gpt-oss-20b tok/s | 354.23 | 63.63 | +457% |
| Hermes-3-Llama-3.2-3B tok/s | 430.79 | 84.99 | +407% |
| Hermes-4-70B tok/s | 42.77 | 0 | n/a |
| SmolLM3-3B tok/s | 404.27 | 86.86 | +365% |
| Qwen3-Coder-Next-abliterated tok/s | 197.14 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 236.31 | 0 | n/a |
| L3-8B-Stheno-v3.2 tok/s | 267.92 | 36.28 | +638% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 946.53 | 221.66 | +327% |
| LFM2.5-8B-A1B tok/s | 583.96 | 139.03 | +320% |
| Llama-2-7B tok/s | 290.81 | 42.98 | +577% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 429.78 | 86.29 | +398% |
| Llama-3.3-70B-Instruct-abliterated tok/s | 42.75 | 0 | n/a |
| Meta-Llama-3.1-70B tok/s | 42.72 | 0 | n/a |
| Meta-Llama-3.1-8B tok/s | 262.93 | 37.8 | +596% |
| Phi-4-mini tok/s | 395.3 | 66 | +499% |
| Mistral-7B-Instruct-v0.1 tok/s | 282.64 | 40.63 | +596% |
| Mistral-7B-Instruct-v0.2 tok/s | 282.32 | 41.31 | +583% |
| Mistral-7B-Instruct-v0.3 tok/s | 282.39 | 39.87 | +608% |
| Mistral-Nemo-Instruct-2407 tok/s | 181.3 | 22.98 | +689% |
| Mistral Small 24B (Q3_K_M) tok/s | 85.1 | 10.13 | +740% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 181.53 | 23.19 | +683% |
| Nemotron-3-Nano-30B-A3B tok/s | 327.48 | 0 | n/a |
| Hermes-4-14B tok/s | 156.26 | 19.02 | +722% |
| Ornith-1.0-35B tok/s | 208.51 | 0 | n/a |
| Ornith-1.0-9B tok/s | 222.94 | 32.93 | +577% |
| phi-2 tok/s | 347.08 | 89.6 | +287% |
| Phi-3.5-mini tok/s | 348.14 | 72.07 | +383% |
| Phi-4 14B (Q3_K_M) tok/s | 139.97 | 16.04 | +773% |
| Qwen-AgentWorld-35B-A3B tok/s | 225.31 | 0 | n/a |
| Qwen3-0.6B tok/s | 724.67 | 259.91 | +179% |
| Qwen3-1.7B tok/s | 568.85 | 135.21 | +321% |
| Qwen3-14B tok/s | 156.3 | 19.44 | +704% |
| Qwen3-30B-A3B tok/s | 293.47 | 0 | n/a |
| Qwen3-4B-Instruct-2507 tok/s | 318.67 | 66.89 | +376% |
| Qwen3-8B tok/s | 251.04 | 37.37 | +572% |
| Qwen3-Coder-Next tok/s | 195.21 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 194.82 | 0 | n/a |
| Qwen1.5-0.5B tok/s | 855.22 | 328.29 | +161% |
| Qwen2-1.5B tok/s | 546.47 | 147.68 | +270% |
| Qwen2.5-0.5B tok/s | 914.58 | 295.47 | +210% |
| Qwen2.5-1.5B tok/s | 549.36 | 148.14 | +271% |
| Uncensored tok/s | 148.7 | 18.94 | +685% |
| Qwen2.5-14B tok/s | 148.76 | 19.84 | +650% |
| Qwen2.5-32B tok/s | 75.76 | 0 | n/a |
| Qwen2.5-3B tok/s | 400.48 | 86.04 | +365% |
| Qwen2.5-72B tok/s | 43.4 | 0 | n/a |
| Qwen2.5-7B tok/s | 265.15 | 39.01 | +580% |
| Qwen2.5-Coder-0.5B tok/s | 920.36 | 294.96 | +212% |
| Qwen2.5-Coder-1.5B tok/s | 545.54 | 148.51 | +267% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 148.48 | 18.86 | +687% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 58.87 | 0 | n/a |
| Qwen2.5-Coder-3B tok/s | 401.11 | 85.74 | +368% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 270.92 | 37.31 | +626% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 246.35 | 61.82 | +298% |
| Qwen3-4B-Instruct-2507 tok/s | 318.27 | 67.45 | +372% |
| Qwen3-4B-Thinking-2507 tok/s | 318.94 | 67 | +376% |
| Qwen3-Coder-Next tok/s | 196.65 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 201.12 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 192.52 | 0 | n/a |
| SmolLM2-135M tok/s | 905.38 | 416.24 | +118% |
| Cydonia-24B-v4.3 tok/s | 109.2 | 11.21 | +874% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA H200 is faster on 4 of 4; NVIDIA T4 on 0.
| Workflow | NVIDIA H200 | NVIDIA T4 | Difference |
|---|---|---|---|
| 50-image depth pass | 6 s | 13 s | NVIDIA H200 2.15x faster |
| 500-image masking run | 1.7 min | 14.4 min | NVIDIA H200 8.64x faster |
| Full codebase review | 6.7 min | 50.7 min | NVIDIA H200 7.53x faster |
| 200-product catalogue cutout | 8 min | 20.3 min | NVIDIA H200 2.54x faster |
Specifications compared
| NVIDIA H200 | NVIDIA T4 | |
|---|---|---|
| VRAM | 141GB | 16GB |
| Architecture | Hopper | Turing |
| Memory bandwidth | 4800 GB/s | 320 GB/s |
| Boost clock | 1,980 MHz | 1,590 MHz |
| TDP | 700 W | 70 W |
| Launch MSRP | $31,000 | $2,299 |
| Release | 2024-03-18 | 2018-09-13 |
FAQ
- Which is better for ai & machine learning: NVIDIA H200 or NVIDIA T4?
- NVIDIA H200 performs better for ai & machine learning, winning 144 of 146 benchmarks in our suite with an average 515.7% advantage.
- What are the main hardware differences between NVIDIA H200 and NVIDIA T4?
- NVIDIA H200 has 141GB VRAM and a 700W TDP, while NVIDIA T4 has 16GB VRAM and a 70W TDP.
- Does VRAM matter more than speed between NVIDIA H200 and NVIDIA T4?
- For AI, yes, NVIDIA H200 fits 42 more of our 12 workloads than NVIDIA T4. 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 H200 and NVIDIA T4?
- Stable Diffusion XL: NVIDIA H200 leads by roughly 1475% (37.16 vs 2.36 images/min) in our testing.
NVIDIA H200 full review · NVIDIA T4 full review · All AI & Machine Learning rankings