NVIDIA H200 vs NVIDIA L4, AI & Machine Learning Comparison
NVIDIA H200 wins 148 of 150 benchmarks, averaging 350% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Wan 2.2 5B (720p), where NVIDIA H200 leads by 840% (1.41 vs 0.15 frames/s); the closest fight is Depth Anything V2 Small (31% apart); VRAM decides part of this one: NVIDIA H200 runs 150 of our 12 AI workloads while the other card runs 138, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA H200 | NVIDIA L4 | Difference |
|---|---|---|---|
| Qwen3 4B tok/s | 318.84 | 84 | +280% |
| Llama 3.1 8B tok/s | 268.31 | 50.45 | +432% |
| Qwen2.5-Coder 14B tok/s | 148.43 | 27.46 | +441% |
| Qwen3 32B tok/s | 76.58 | 12.42 | +517% |
| Llama 3.3 70B tok/s | 42.66 | 0 | n/a |
| Stable Diffusion XL images/min | 37.16 | 5.18 | +617% |
| Z-Image Turbo images/min | 23.175 | 2.475 | +836% |
| 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 | 0.15 | +840% |
| DeepSeek-R1 Distill Llama 8B tok/s | 265.25 | 50.39 | +426% |
| DeepSeek-R1 Distill 1.5B tok/s | 542.95 | 191.38 | +184% |
| DeepSeek-R1 Distill 14B tok/s | 146.82 | 27.39 | +436% |
| DeepSeek-R1 Distill 7B tok/s | 267.1 | 53.27 | +401% |
| Gemma 3 12B tok/s | 153.26 | 31.01 | +394% |
| Gemma 3 4B tok/s | 291.76 | 82.03 | +256% |
| Gemma 4 12B tok/s | 151.05 | 31.49 | +380% |
| Llama 3.2 1B tok/s | 875.53 | 259.48 | +237% |
| Llama 3.2 3B tok/s | 424.61 | 107.1 | +296% |
| Mistral 7B v0.3 tok/s | 278.5 | 54.02 | +416% |
| Mistral Small 24B tok/s | 107.84 | 17.33 | +522% |
| Phi-4 14B tok/s | 168.2 | 27.24 | +517% |
| Phi-4 Mini 3.8B tok/s | 392.38 | 88.48 | +343% |
| Qwen2.5-Coder 7B tok/s | 266.36 | 53.26 | +400% |
| Qwen3 0.6B tok/s | 718.67 | 357.79 | +101% |
| Qwen3 1.7B tok/s | 563.52 | 177.5 | +217% |
| Qwen3 14B tok/s | 154.51 | 27.61 | +460% |
| Qwen3 30B A3B tok/s | 292.11 | 96.15 | +204% |
| Qwen3 8B tok/s | 247.98 | 48.95 | +407% |
| SmolLM3 3B tok/s | 402.24 | 112.07 | +259% |
| TRELLIS Image-to-3D assets/hour | 818.8 | 284.8 | +188% |
| Codestral 22B tok/s | 111.34 | 18.14 | +514% |
| DeepSeek-R1 Distill 32B tok/s | 75.8 | 12.28 | +517% |
| Devstral Small 24B tok/s | 109.05 | 17.34 | +529% |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 76.7 | 11.92 | +543% |
| Dolphin Mistral 24B Venice tok/s | 109.14 | 17.34 | +529% |
| Dolphin X1 8B tok/s | 268.36 | 50.12 | +435% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 267.84 | 50.24 | +433% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 109.06 | 17.32 | +530% |
| Gemma 3 27B tok/s | 84.75 | 14.18 | +498% |
| Qwen2.5-Coder 32B tok/s | 75.77 | 12.29 | +517% |
| Qwen3-Coder 30B A3B tok/s | 297.9 | 99.14 | +200% |
| QwQ 32B tok/s | 75.77 | 12.28 | +517% |
| StarCoder2 15B tok/s | 136.29 | 24.4 | +459% |
| FLUX.1 Schnell images/min | 60.51 | 0 | n/a |
| Z-Image Turbo (1024px) images/min | 40.74 | 4.6 | +786% |
| Krea 2 Turbo images/min | 7.25 | 0 | n/a |
| BiRefNet images/min | 1457.73 | 305.81 | +377% |
| Depth Anything V2 Large images/min | 1040.81 | 770.37 | +35% |
| Depth Anything V2 Small images/min | 1198.58 | 916.66 | +31% |
| SAM ViT-Base images/min | 1469.34 | 315.89 | +365% |
| SAM ViT-Huge images/min | 325.19 | 70.46 | +362% |
| Swin2SR 4x Upscaler images/min | 25.85 | 18.23 | +42% |
| Qwen2.5 1.5B LoRA train tok/s | 14931.7 | 3720.9 | +301% |
| Qwen2.5 7B LoRA train tok/s | 8855.4 | 1063.9 | +732% |
| SmolLM2 1.7B LoRA train tok/s | 17650.4 | 3240.2 | +445% |
| TinyLlama 1.1B LoRA train tok/s | 16222.4 | 4756 | +241% |
| Qwen2.5 1.5B served serve tok/s | 7541.6 | 1827 | +313% |
| Qwen2.5 7B served serve tok/s | 4394.8 | 469.9 | +835% |
| SmolLM2 1.7B served serve tok/s | 7107.8 | 1524.9 | +366% |
| TinyLlama 1.1B served serve tok/s | 9137.7 | 2584.3 | +254% |
| Kokoro TTS 82M x realtime | 170.82 | 97.35 | +75% |
| MusicGen Small x realtime | 1.87 | 1.01 | +85% |
| Whisper large-v3 x realtime | 163.78 | 70.11 | +134% |
| AI21-Jamba-Reasoning-3B tok/s | 370.66 | 107.59 | +245% |
| Olmo-3.1-32B-Think tok/s | 78.46 | 12.27 | +539% |
| Codestral 22B (Q3_K_M) tok/s | 88.03 | 17.96 | +390% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 109.2 | 17.2 | +535% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 267.82 | 48.82 | +449% |
| DeepSeek-Coder-V2-Lite tok/s | 312.1 | 110.11 | +183% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 250.61 | 49.07 | +411% |
| DeepSeek-R1-Distill-Llama-70B tok/s | 42.76 | 0 | n/a |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 119.66 | 27.87 | +329% |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 75.71 | 12.26 | +518% |
| dolphin-2.9-llama3-8b tok/s | 267.96 | 50.04 | +435% |
| Dolphin X1 Trinity Nano 6B tok/s | 259.81 | 166.6 | +56% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 148.74 | 27.3 | +445% |
| gemma-2-2b-it-abliterated tok/s | 397.43 | 113 | +252% |
| gemma-2-2b tok/s | 398.48 | 114.17 | +249% |
| gemma-2-9b tok/s | 180.47 | 33.8 | +434% |
| Gemma 3 12B (Q3_K_M) tok/s | 127.04 | 32.29 | +293% |
| gemma-3-1b tok/s | 513.58 | 216.27 | +137% |
| gemma-3-270m tok/s | 967.35 | 493.31 | +96% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 171.45 | 70.87 | +142% |
| GLM-4.7-Flash tok/s | 188.57 | 75.03 | +151% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 250.88 | 48.72 | +415% |
| gpt-oss-20b tok/s | 354.23 | 91.78 | +286% |
| Hermes-3-Llama-3.2-3B tok/s | 430.79 | 106.14 | +306% |
| Hermes-4-70B tok/s | 42.77 | 0 | n/a |
| SmolLM3-3B tok/s | 404.27 | 110.77 | +265% |
| Qwen3-Coder-Next-abliterated tok/s | 197.14 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 236.31 | 79.01 | +199% |
| L3-8B-Stheno-v3.2 tok/s | 267.92 | 50.07 | +435% |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| LFM2.5-1.2B tok/s | 946.53 | 278.53 | +240% |
| LFM2.5-8B-A1B tok/s | 583.96 | 170.2 | +243% |
| Llama-2-7B tok/s | 290.81 | 56.06 | +419% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 429.78 | 101.16 | +325% |
| 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 | 50.36 | +422% |
| Phi-4-mini tok/s | 395.3 | 87.21 | +353% |
| Mistral-7B-Instruct-v0.1 tok/s | 282.64 | 53.6 | +427% |
| Mistral-7B-Instruct-v0.2 tok/s | 282.32 | 53.31 | +430% |
| Mistral-7B-Instruct-v0.3 tok/s | 282.39 | 53.66 | +426% |
| Mistral-Nemo-Instruct-2407 tok/s | 181.3 | 33.01 | +449% |
| Mistral Small 24B (Q3_K_M) tok/s | 85.1 | 16.62 | +412% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| NemoMix-Unleashed-12B tok/s | 181.53 | 33.02 | +450% |
| Nemotron-3-Nano-30B-A3B tok/s | 327.48 | 0 | n/a |
| Hermes-4-14B tok/s | 156.26 | 27.48 | +469% |
| Ornith-1.0-35B tok/s | 208.51 | 70.31 | +197% |
| Ornith-1.0-9B tok/s | 222.94 | 43.35 | +414% |
| phi-2 tok/s | 347.08 | 114.99 | +202% |
| Phi-3.5-mini tok/s | 348.14 | 90.56 | +284% |
| Phi-4 14B (Q3_K_M) tok/s | 139.97 | 27.86 | +402% |
| Qwen-AgentWorld-35B-A3B tok/s | 225.31 | 70.23 | +221% |
| Qwen3-0.6B tok/s | 724.67 | 356.93 | +103% |
| Qwen3-1.7B tok/s | 568.85 | 176.08 | +223% |
| Qwen3-14B tok/s | 156.3 | 27.48 | +469% |
| Qwen3-30B-A3B tok/s | 293.47 | 95.02 | +209% |
| Qwen3-4B-Instruct-2507 tok/s | 318.67 | 83.03 | +284% |
| Qwen3-8B tok/s | 251.04 | 48.34 | +419% |
| 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 | 431.41 | +98% |
| Qwen2-1.5B tok/s | 546.47 | 189.18 | +189% |
| Qwen2.5-0.5B tok/s | 914.58 | 399.4 | +129% |
| Qwen2.5-1.5B tok/s | 549.36 | 191.39 | +187% |
| Uncensored tok/s | 148.7 | 27.29 | +445% |
| Qwen2.5-14B tok/s | 148.76 | 27.37 | +444% |
| Qwen2.5-32B tok/s | 75.76 | 12.24 | +519% |
| Qwen2.5-3B tok/s | 400.48 | 110.84 | +261% |
| Qwen2.5-72B tok/s | 43.4 | 0 | n/a |
| Qwen2.5-7B tok/s | 265.15 | 53.32 | +397% |
| Qwen2.5-Coder-0.5B tok/s | 920.36 | 397.1 | +132% |
| Qwen2.5-Coder-1.5B tok/s | 545.54 | 184.6 | +196% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 148.48 | 27.29 | +444% |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 58.87 | 12.05 | +389% |
| Qwen2.5-Coder-3B tok/s | 401.11 | 109.42 | +267% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 270.92 | 53.02 | +411% |
| Qwen3 30B A3B (Q3_K_M) tok/s | 246.35 | 94.03 | +162% |
| Qwen3-4B-Instruct-2507 tok/s | 318.27 | 83.13 | +283% |
| Qwen3-4B-Thinking-2507 tok/s | 318.94 | 83.12 | +284% |
| 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 | 652.15 | +39% |
| Cydonia-24B-v4.3 tok/s | 109.2 | 17.3 | +531% |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA H200 is faster on 8 of 8; NVIDIA L4 on 0.
| Workflow | NVIDIA H200 | NVIDIA L4 | Difference | Cost per run |
|---|---|---|---|---|
| 50-image depth pass | 6 s | 17 s | NVIDIA H200 2.75x faster | $0.006 vs $0.002 |
| 30-minute podcast pass | 46 s | 2 min | NVIDIA H200 2.63x faster | $0.046 vs $0.015 |
| 500-image masking run | 1.7 min | 7.5 min | NVIDIA H200 4.47x faster | $0.100 vs $0.055 |
| 20-asset 3D game kit | 2.4 min | 8.3 min | NVIDIA H200 3.45x faster | $0.145 vs $0.061 |
| 24-frame storyboard | 3.9 min | 12.2 min | NVIDIA H200 3.12x faster | $0.234 vs $0.089 |
| Full codebase review | 6.7 min | 36.4 min | NVIDIA H200 5.41x faster | $0.403 vs $0.267 |
| 200-product catalogue cutout | 8 min | 12.1 min | NVIDIA H200 1.51x faster | $0.479 vs $0.089 |
| 10 short social clips | 10 min | 60.3 min | NVIDIA H200 6.03x faster | $0.598 vs $0.442 |
Renting by the hour, NVIDIA L4 finishes 8 of 8 cheaper. The quicker card is not automatically the cheaper way to get the work done.
Cost to rent
| Card | Per hour |
|---|---|
| NVIDIA H200 | $3.590 |
| NVIDIA L4 | $0.440 |
NVIDIA L4 is 8.16x cheaper per hour. Cheapest on-demand rate we see across RunPod and Vast.
Specifications compared
| NVIDIA H200 | NVIDIA L4 | |
|---|---|---|
| VRAM | 141GB | 24GB |
| Architecture | Hopper | Ada Lovelace |
| Memory bandwidth | 4800 GB/s | 300 GB/s |
| Boost clock | 1,980 MHz | 2,040 MHz |
| TDP | 700 W | 72 W |
| Launch MSRP | $31,000 | $2,500 |
| Release | 2024-03-18 | 2023-03-21 |
FAQ
- Which is better for ai & machine learning: NVIDIA H200 or NVIDIA L4?
- NVIDIA H200 performs better for ai & machine learning, winning 148 of 150 benchmarks in our suite with an average 350% advantage.
- What are the main hardware differences between NVIDIA H200 and NVIDIA L4?
- NVIDIA H200 has 141GB VRAM and a 700W TDP, while NVIDIA L4 has 24GB VRAM and a 72W TDP.
- Does VRAM matter more than speed between NVIDIA H200 and NVIDIA L4?
- For AI, yes, NVIDIA H200 fits 22 more of our 12 workloads than NVIDIA L4. 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 L4?
- Wan 2.2 5B (720p): NVIDIA H200 leads by roughly 126% (1.41 vs 0.15 frames/s) in our testing.
NVIDIA H200 full review · NVIDIA L4 full review · All AI & Machine Learning rankings