NVIDIA A10G vs NVIDIA T4, AI & Machine Learning Comparison
NVIDIA A10G wins 149 of 171 benchmarks, averaging 124.8% faster.
Both cards were measured first-party on our bench, same suite, same test rig.
What the numbers say
The gap is widest in Sana 1.6B, where NVIDIA A10G leads by 1150% (14.13 vs 1.13 images/min); the closest fight is Qwen2.5 1.5B + 0.5B draft (2% apart); VRAM decides part of this one: NVIDIA A10G runs 161 of our 12 AI workloads while the other card runs 119, models that don't fit score zero.
Benchmark results head-to-head
| Benchmark | NVIDIA A10G | NVIDIA T4 | Difference |
|---|---|---|---|
| SmolLM2-135M tok/s | 680.37 | 416.24 | +63% |
| gemma-3-270m tok/s | 597.49 | 362.8 | +65% |
| Qwen1.5-0.5B tok/s | 542.19 | 328.29 | +65% |
| Qwen2.5-0.5B tok/s | 508.78 | 295.47 | +72% |
| Qwen2.5-Coder-0.5B tok/s | 513.36 | 294.96 | +74% |
| Qwen3 0.6B tok/s | 451.38 | 263.47 | +71% |
| Qwen3-0.6B tok/s | 446.74 | 259.91 | +72% |
| Llama 3.2 1B tok/s | 400.6 | 207.71 | +93% |
| gemma-3-1b tok/s | 284.78 | 156.88 | +82% |
| LFM2.5-1.2B tok/s | 429.96 | 221.66 | +94% |
| DeepSeek-R1 Distill 1.5B tok/s | 265.75 | 148.72 | +79% |
| Qwen2-1.5B tok/s | 265.45 | 147.68 | +80% |
| Qwen2.5-1.5B tok/s | 265.7 | 148.14 | +79% |
| Qwen2.5-Coder-1.5B tok/s | 264.79 | 148.51 | +78% |
| Qwen3 1.7B tok/s | 267.08 | 139.92 | +91% |
| Qwen3-1.7B tok/s | 262.49 | 135.21 | +94% |
| MiniCPM5 2B tok/s | 201.2 | 109.06 | +84% |
| gemma-2-2b tok/s | 175.03 | 92.13 | +90% |
| gemma-2-2b-it-abliterated tok/s | 174.56 | 91.91 | +90% |
| LFM2.5 2.6B tok/s | 201.87 | 104.16 | +94% |
| AI21-Jamba-Reasoning-3B tok/s | 166.29 | 86.91 | +91% |
| Granite 4.1 3B tok/s | 147.45 | 79.3 | +86% |
| Hermes-3-Llama-3.2-3B tok/s | 169.23 | 84.99 | +99% |
| Llama 3.2 3B tok/s | 170.47 | 85.94 | +98% |
| Llama-3.2-3B-Instruct-uncensored tok/s | 167.63 | 86.29 | +94% |
| Nanbeige4.2-3B tok/s | 0 | 0 | n/a |
| Qwen2.5-3B tok/s | 168.23 | 86.04 | +96% |
| Qwen2.5-Coder-3B tok/s | 168 | 85.74 | +96% |
| SmolLM3 3B tok/s | 172.15 | 86.08 | +100% |
| SmolLM3-3B tok/s | 171.66 | 86.86 | +98% |
| Phi-4 Mini 3.8B tok/s | 144.42 | 64.15 | +125% |
| Gemma 3 4B tok/s | 124.85 | 65.12 | +92% |
| Nemotron 3 Nano 4B tok/s | 137.33 | 65.33 | +110% |
| Qwen3-4B tok/s | 129.87 | 67.71 | +92% |
| Qwen3-4B-Instruct-2507 tok/s | 129.9 | 67.45 | +93% |
| Qwen3-4B-Instruct-2507 tok/s | 130.83 | 66.89 | +96% |
| Qwen3-4B-Thinking-2507 tok/s | 129.99 | 67 | +94% |
| phi-2 tok/s | 172.45 | 89.6 | +92% |
| Dolphin X1 Trinity Nano 6B tok/s | 171.23 | 114.21 | +50% |
| Phi-3.5-mini tok/s | 145.97 | 72.07 | +103% |
| Phi-4-mini tok/s | 144.27 | 66 | +119% |
| DeepSeek Coder 7B Instruct v1.5 tok/s | 99.14 | 46.77 | +112% |
| DeepSeek-R1 Distill 7B tok/s | 89.79 | 37.58 | +139% |
| Llama-2-7B tok/s | 96.69 | 42.98 | +125% |
| Mistral 7B v0.3 tok/s | 92.82 | 40.01 | +132% |
| Mistral-7B-Instruct-v0.1 tok/s | 92.45 | 40.63 | +128% |
| Mistral-7B-Instruct-v0.2 tok/s | 92.74 | 41.31 | +124% |
| Mistral-7B-Instruct-v0.3 tok/s | 92.45 | 39.87 | +132% |
| Qwen2.5-7B tok/s | 89.68 | 39.01 | +130% |
| Qwen2.5-Coder 7B tok/s | 90.05 | 37.54 | +140% |
| Qwen2.5-Coder-7B-Instruct-abliterated tok/s | 91.41 | 37.31 | +145% |
| DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensored tok/s | 87.84 | 35.73 | +146% |
| DeepSeek-R1 Distill Llama 8B tok/s | 86.83 | 36.36 | +139% |
| DeepSeek-R1-0528-Qwen3-8B tok/s | 83.22 | 37.7 | +121% |
| Dolphin 3.0 Llama 3.1 8B tok/s | 86.54 | 35.68 | +143% |
| Dolphin X1 8B tok/s | 86.61 | 35.66 | +143% |
| Josiefied-Qwen3-8B-abliterated-v1 tok/s | 83.26 | 35.62 | +134% |
| L3-8B-Stheno-v3.2 tok/s | 86.8 | 36.28 | +139% |
| LFM2.5-8B-A1B tok/s | 268.04 | 139.03 | +93% |
| Llama 3 8B tok/s | 87.2 | 41.34 | +111% |
| Llama-3.1-8B tok/s | 86.62 | 35.01 | +147% |
| Meta-Llama-3.1-8B tok/s | 86.69 | 37.8 | +129% |
| Qwen3 8B tok/s | 84.09 | 36.51 | +130% |
| Qwen3-8B tok/s | 83.04 | 37.37 | +122% |
| dolphin-2.9-llama3-8b tok/s | 87.98 | 35.1 | +151% |
| Nemotron Nano 9B v2 tok/s | 64.8 | 31.73 | +104% |
| Ornith 1.5 9B tok/s | 75.38 | 34.7 | +117% |
| Ornith-1.0-9B tok/s | 73.78 | 32.93 | +124% |
| gemma-2-9b tok/s | 57.37 | 28.62 | +100% |
| Gemma 3 12B tok/s | 52.07 | 23.12 | +125% |
| Gemma 3 12B (Q3_K_M) tok/s | 46.29 | 19.79 | +134% |
| Gemma 4 12B tok/s | 52.67 | 24.03 | +119% |
| NemoMix-Unleashed-12B tok/s | 57.52 | 23.19 | +148% |
| DeepSeek-R1 Distill 14B tok/s | 46.96 | 19.21 | +144% |
| DeepSeek-R1 Distill 14B (Q3_K_M) tok/s | 41.72 | 16.86 | +147% |
| EVA-Qwen2.5-14B-v0.2 tok/s | 47 | 19.92 | +136% |
| Hermes-4-14B tok/s | 47.8 | 19.02 | +151% |
| Phi-4 14B tok/s | 48.18 | 17.46 | +176% |
| Phi-4 14B (Q3_K_M) tok/s | 43.56 | 16.04 | +172% |
| Qwen2.5-14B tok/s | 47.02 | 19.84 | +137% |
| Qwen2.5-Coder-14B tok/s | 47.07 | 19.76 | +138% |
| Qwen2.5-Coder-14B-Instruct-abliterated tok/s | 46.91 | 18.86 | +149% |
| Qwen3 14B tok/s | 48.12 | 19.31 | +149% |
| Qwen3-14B tok/s | 47.8 | 19.44 | +146% |
| Uncensored tok/s | 47.45 | 18.94 | +151% |
| StarCoder2 15B tok/s | 42.19 | 15.66 | +169% |
| Mistral-Nemo-Instruct-2407 tok/s | 57 | 22.98 | +148% |
| gpt-oss-20b tok/s | 144.61 | 63.63 | +127% |
| Codestral 22B tok/s | 32.09 | 12.59 | +155% |
| Codestral 22B (Q3_K_M) tok/s | 28.01 | 11.3 | +148% |
| GLM-4.7-Flash-REAP-23B-A3B tok/s | 97.48 | 53.83 | +81% |
| DeepSeek-Coder-V2-Lite tok/s | 160.13 | 83.49 | +92% |
| Cydonia-24B-v4.3 tok/s | 31.04 | 11.21 | +177% |
| Devstral Small 24B tok/s | 30.98 | 11.4 | +172% |
| Dolphin 3.0 R1 Mistral 24B tok/s | 30.9 | 11.32 | +173% |
| Dolphin Mistral 24B Venice tok/s | 30.95 | 11.36 | +172% |
| Dolphin-Mistral-24B-Venice-Edition tok/s | 31 | 11.33 | +174% |
| Mistral Small 24B tok/s | 30.94 | 11.33 | +173% |
| Mistral Small 24B (Q3_K_M) tok/s | 26.8 | 10.13 | +165% |
| Gemma 4 26B A4B tok/s | 101.91 | 0 | n/a |
| Gemma 3 27B tok/s | 24.8 | 0 | n/a |
| Qwen3.6 27B tok/s | 24.86 | 0 | n/a |
| Qwen3.8 27B tok/s | 24.51 | 0 | n/a |
| Nemotron 3.5 Lightning 30B A3B tok/s | 0 | 0 | n/a |
| Nemotron-3-Nano-30B-A3B tok/s | 0 | 0 | n/a |
| Qwen3 30B A3B tok/s | 140.12 | 0 | n/a |
| Qwen3 30B A3B (Q3_K_M) tok/s | 120.38 | 61.82 | +95% |
| Qwen3 30B A3B Instruct 2507 tok/s | 147.33 | 0 | n/a |
| Qwen3-30B-A3B tok/s | 140.14 | 0 | n/a |
| Qwen3-Coder 30B A3B tok/s | 142.77 | 0 | n/a |
| Gemma 4 31B tok/s | 22.67 | 0 | n/a |
| DeepSeek-R1 Distill 32B tok/s | 21.85 | 0 | n/a |
| DeepSeek-R1-Distill-Qwen-32B-abliterated tok/s | 22.14 | 0 | n/a |
| Olmo-3.1-32B-Think tok/s | 22.13 | 0 | n/a |
| QwQ 32B tok/s | 21.87 | 0 | n/a |
| Qwen2.5-32B tok/s | 21.98 | 0 | n/a |
| Qwen2.5-Coder 32B tok/s | 21.92 | 0 | n/a |
| Qwen2.5-Coder 32B (Q3_K_M) tok/s | 18.78 | 0 | n/a |
| Qwen3-32B tok/s | 22.04 | 0 | n/a |
| Dolphin 2.9.1 Yi 1.5 34B tok/s | 21.12 | 0 | n/a |
| Ornith 1.5 35B A3B tok/s | 123.77 | 0 | n/a |
| Ornith-1.0-35B tok/s | 104.9 | 0 | n/a |
| Qwen-AgentWorld-35B-A3B tok/s | 104.36 | 0 | n/a |
| Qwen3.6 35B A3B tok/s | 117.92 | 0 | n/a |
| GLM-4.7-Flash tok/s | 103.24 | 0 | n/a |
| Laguna-XS-2.1 tok/s | 0 | 0 | n/a |
| KAT-Coder-V2.5-Dev tok/s | 113.6 | 0 | n/a |
| DeepSeek-R1-Distill-Llama-70B tok/s | 0 | 0 | n/a |
| Hermes-4-70B tok/s | 0 | 0 | n/a |
| Llama 3.3 70B tok/s | 0 | 0 | n/a |
| Llama-3.3-70B-Instruct-abliterated tok/s | 0 | 0 | n/a |
| Meta-Llama-3.1-70B tok/s | 0 | 0 | n/a |
| Qwen2.5-72B tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 0 | n/a |
| Qwen3-Next-80B-A3B-Thinking tok/s | 0 | 0 | n/a |
| Qwen3-Coder-Next tok/s | 0 | 0 | n/a |
| Qwen3-Coder-Next tok/s | 0 | 0 | n/a |
| Qwen3-Coder-Next-abliterated tok/s | 0 | 0 | n/a |
| gpt-oss-120b tok/s | 0 | 0 | n/a |
| BiRefNet images/min | 432.33 | 183.48 | +136% |
| Swin2SR 4x Upscaler images/min | 22.31 | 10.53 | +112% |
| Sana 1.6B images/min | 14.13 | 1.13 | +1150% |
| Stable Diffusion XL images/min | 6.38 | 2.36 | +170% |
| PixArt-Sigma XL images/min | 9.93 | 0 | n/a |
| Stable Diffusion 3.5 Medium images/min | 4.11 | 0 | n/a |
| AuraFlow v0.3 images/min | 1.39 | 0 | n/a |
| FLUX.1 Kontext dev images/min | 0 | 0 | n/a |
| FLUX.1 dev images/min | 0 | 0 | n/a |
| Stable Diffusion 3.5 Large images/min | 0 | 0 | n/a |
| FLUX.1 Schnell images/min | 0 | 0 | n/a |
| Qwen-Image-Edit images/min | 0 | 0 | n/a |
| Wan 2.2 5B (720p) frames/s | 0.18 | 0 | n/a |
| Kokoro TTS 82M x realtime | 101.11 | 42.88 | +136% |
| MusicGen Small x realtime | 1.2 | 0.92 | +30% |
| Whisper large-v3 x realtime | 86.07 | 44.36 | +94% |
| Depth Anything V2 Small images/min | 941.18 | 612.25 | +54% |
| Florence-2 Base images/min | 210.79 | 148.61 | +42% |
| SAM ViT-Base images/min | 398.72 | 180.73 | +121% |
| Depth Anything V2 Large images/min | 740.74 | 402.34 | +84% |
| Florence-2 Large images/min | 117.94 | 83.43 | +41% |
| SAM ViT-Huge images/min | 84.22 | 34.91 | +141% |
| TinyLlama 1.1B LoRA train tok/s | 5499.6 | 1520.1 | +262% |
| TinyLlama 1.1B served serve tok/s | 3558.2 | 1897.8 | +87% |
| Qwen2.5 1.5B LoRA train tok/s | 4214.4 | 1434.4 | +194% |
| Qwen2.5 1.5B served serve tok/s | 2702.5 | 1337.1 | +102% |
| SmolLM2 1.7B LoRA train tok/s | 3668.3 | 1561.9 | +135% |
| SmolLM2 1.7B served serve tok/s | 2242.7 | 1222.5 | +83% |
| Qwen2.5 1.5B + 0.5B draft x vs solo | 0.919 | 0.9 | +2% |
| Qwen2.5 7B LoRA train tok/s | 1239 | 0 | n/a |
| Qwen2.5 7B served serve tok/s | 858.7 | 0 | n/a |
Whole-job comparison
How long each card takes to finish a complete pipeline, not just one model. NVIDIA A10G is faster on 7 of 7; NVIDIA T4 on 0.
| Workflow | NVIDIA A10G | NVIDIA T4 | Difference |
|---|---|---|---|
| Depth pass on a batch | 7 s | 13 s | NVIDIA A10G 2.02x faster |
| Voiceovers from scripts | 28 s | 55 s | NVIDIA A10G 1.96x faster |
| Transcribe and subtitle videos | 3.6 min | 6.9 min | NVIDIA A10G 1.93x faster |
| Masking run | 6 min | 14.4 min | NVIDIA A10G 2.40x faster |
| Caption a training dataset | 8.5 min | 12.1 min | NVIDIA A10G 1.42x faster |
| Product catalogue cutout | 9.6 min | 20.3 min | NVIDIA A10G 2.12x faster |
| Full codebase review | 21.3 min | 50.7 min | NVIDIA A10G 2.38x faster |
Specifications compared
| NVIDIA A10G | NVIDIA T4 | |
|---|---|---|
| VRAM | 24GB | 16GB |
| Transistors | 28,300M | 13,600M |
| Die size | 628.4 mm² | 545 mm² |
| Process node | 8 nm | 12 nm |
| Transistor density | 45 M/mm² | 25 M/mm² |
| Architecture | Ampere | Turing |
| Memory bandwidth | 600 GB/s | 320 GB/s |
| Boost clock | 1,710 MHz | 1,590 MHz |
| TDP | 150 W | 70 W |
| Launch MSRP | $2,800 | $2,299 |
| Release | 2021-11-01 | 2018-09-13 |
FAQ
- Which is better for ai & machine learning: NVIDIA A10G or NVIDIA T4?
- NVIDIA A10G performs better for ai & machine learning, winning 149 of 171 benchmarks in our suite with an average 124.8% advantage.
- What are the main hardware differences between NVIDIA A10G and NVIDIA T4?
- NVIDIA A10G has 24GB VRAM and a 150W TDP, while NVIDIA T4 has 16GB VRAM and a 70W TDP.
- Does VRAM matter more than speed between NVIDIA A10G and NVIDIA T4?
- For AI, yes, NVIDIA A10G fits 28 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 A10G and NVIDIA T4?
- Sana 1.6B: NVIDIA A10G leads by roughly 1150% (14.13 vs 1.13 images/min) in our testing.
NVIDIA A10G full review · NVIDIA T4 full review · All AI & Machine Learning rankings