16GB · AI Score 4.7/100 · first-party measured on 12 AI workloads
4.7 AI Score Includes estimates
Every number on this page is first-party: GeForce RTX 5070 Ti was run on our pinned 12-workload AI suite on 2026-07-11, with under 0.5% run-to-run variance. On Llama 3.1 8B (Q4_K_M) GeForce RTX 5070 Ti delivers about 153.23 tokens/sec. Llama 3.3 70B does not fit. It needs roughly 42GB and this card has 16GB. For image generation, SDXL runs at 4.69 it/s, while FLUX.1-dev won't fit at BF16 (needs ~26GB). 6 of the 12 workloads won't fit on 16GB at the tested precision, Qwen3 32B, Llama 3.3 70B, FLUX.1-dev, FLUX.1 Kontext and others. We publish those as hard gates rather than quietly dropping to a smaller quant.
Best efficiency of the Blackwell mid-range at 1.58 tokens/watt, and the 16GB means only 6 of 12 workloads gated. It does work for its numbers though, I logged it at 97% of its 300W rating on the coder model. The Z-Image run was noisy, right at the VRAM edge. Quick note on the setup: all my AI benchmarking was done on rented cloud GPUs, I used all three of Vast.ai, RunPod and Modal depending on which had the card, and they all have their pros and cons. Same pinned harness on every run, and everything here got double-checked before it went up.
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Qwen3 0.6B | 665.68 tok/s | 60 W39°CQ4_K_M | ✓ Measured |
| Llama 3.2 1B | 691.08 tok/s | 78 W45°CQ4_K_M | ✓ Measured |
| MiniCPM5 2B | 337.02 tok/s | 94 W51°CQ4_K_M | ✓ Measured |
| LFM2.5 2.6B | 335.29 tok/s | 100 W53°CQ4_K_M | ✓ Measured |
| Granite 4.1 3B | 229.37 tok/s | 126 W53°CQ4_K_M | ✓ Measured |
| Nemotron 3 Nano 4B | 230.54 tok/s | 134 W60°CQ4_K_M | ✓ Measured |
| Qwen3 4B | 231.69 tok/s | 2.7 GB peak147 W48°C1.58 tok/WQ4_K_M | ✓ Measured |
| DeepSeek Coder 7B Instruct v1.5 | 163.53 tok/s | 163 W59°CQ4_K_M | ✓ Measured |
| Qwen2.5-7B | 161.02 tok/s | 164 W48°CQ4_K_M | ✓ Measured |
| Qwen2.5-Coder 7B | 161.01 tok/s | 162 W51°CQ4_K_M | ✓ Measured |
| Llama 3 8B | 144.39 tok/s | 168 W56°CQ4_K_M | ✓ Measured |
| Llama 3.1 8B | 153.23 tok/s | 4.9 GB peak212 W54°C0.73 tok/WQ4_K_M | ✓ Measured |
| Qwen3 8B | 139.27 tok/s | 166 W56°CQ4_K_M | ✓ Measured |
| Nemotron Nano 9B v2 | 110.37 tok/s | 164 W54°CQ4_K_M | ✓ Measured |
| Ornith 1.5 9B | 124.91 tok/s | 165 W56°CQ4_K_M | ✓ Measured |
| Gemma 4 12B | 88.02 tok/s | 179 W59°CQ4_K_M | ✓ Measured |
| Qwen2.5-Coder 14B | 83.44 tok/s | 8.3 GB peak224 W58°C0.37 tok/WQ4_K_M | ✓ Measured |
| Qwen3 14B | 79.02 tok/s | 197 W58°CQ4_K_M | ✓ Measured |
| gpt-oss-20b | 247.36 tok/s | 106 W48°CQ4_K_M | ✓ Measured |
| Gemma 4 26B A4B | ✕ Won't fit needs ~20 GB | VRAM-gated at this precision | Est. |
| Qwen3 30B A3B | ✕ Won't fit needs ~20 GB | VRAM-gated at this precision | Est. |
| Gemma 4 31B | ✕ Won't fit needs ~22 GB | VRAM-gated at this precision | Est. |
| Qwen3 32B | ✕ Won't fit needs ~23 GB | VRAM-gated at this precision | ✓ Measured |
| Llama 3.3 70B | ✕ Won't fit needs ~46 GB | VRAM-gated at this precision | ✓ Measured |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Stable Diffusion 1.5 | 49.51 images/min | 233 W59°C | ✓ Measured |
| SD Turbo | 644.31 images/min | 62 W55°C | ✓ Measured |
| LCM DreamShaper v7 | 179.8 images/min | 155 W57°C | ✓ Measured |
| SDXL Turbo | 431.35 images/min | 54 W48°C | ✓ Measured |
| Z-Image Turbo | 2.43 images/min | ✓ Measured 3 hosts ±4% | |
| Sana 1.6B | 26.18 images/min | 282 W58°C | ✓ Measured |
| Stable Diffusion XL | 9.38 images/min | 14.4 GB peak273 W66°C6.4 s/img | ✓ Measured |
| Playground v2.5 | 5.88 images/min | 247 W62°C | ✓ Measured |
| PixArt-Sigma XL | 13.34 images/min | 226 W58°C | ✓ Measured |
| FLUX.2 klein 4B | ✕ Won't fit needs ~19 GB | VRAM-gated at this precision | ✓ Measured |
| FLUX.1 dev | ✕ Won't fit needs ~26 GB | VRAM-gated at this precision | ✓ Measured |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| FLUX.1 Kontext dev | ✕ Won't fit needs ~26 GB | VRAM-gated at this precision | ✓ Measured |
| Qwen-Image-Edit | ✕ Won't fit needs ~42 GB | VRAM-gated at this precision | ✓ Measured |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Stable Video Diffusion | 1.4 clips/min | 282 W68°C | ✓ Measured |
| LTX-Video (image to video) | 1.83 clips/min | 141 W55°C | ✓ Measured 2 hosts ±11% · CPU offload |
| Wan 2.2 TI2V-5B (image to video) | 0.77 clips/min | 182 W68°C | ✓ Measured 2 hosts ±4% · CPU offload |
| Stable Video Diffusion XT | 0.77 clips/min | 253 W69°C | ✓ Measured 2 hosts ±0% · CPU offload |
| CogVideoX-5B I2V | 0.21 clips/min | 224 W65°C | ✓ Measured 2 hosts ±1% · CPU offload |
| Workload | Result | Telemetry | Data |
|---|---|---|---|
| Wan 2.1 1.3B | 0.37 frames/s | 195 W63°C129.5 s/clip | ✓ Measured 2 hosts ±2% · CPU offload |
| CogVideoX-2B | 0.29 frames/s | 220 W65°C164.3 s/clip | ✓ Measured 2 hosts ±2% · CPU offload |
| LTX-Video (distilled) | 2.8 frames/s | 126 W51°C43.9 s/clip | ✓ Measured 3 hosts ±15% · CPU offload |
| Wan 2.2 5B (720p) | ✕ Won't fit needs ~18 GB | VRAM-gated at this precision | ✓ Measured |
| Architecture | Blackwell |
| CUDA cores | 8,960 |
| VRAM | 16GB GDDR7 |
| Memory bus | 256-bit |
| Memory bandwidth | 896 GB/s |
| Boost clock | 2,452 MHz |
| TDP | 300 W |
| Process | 4nm |
| Interface | PCIe 5.0 x16 |
| Release date | 2025-02-20 |
| Launch MSRP | $749 |
GeForce RTX 5070 Ti scores 4.7/100, #44 of 102. It ran 6 of 12; 6 exceeded its 16GB. Every figure here is our own measurement.
100% = this card, AI & Machine Learning headline metric (AI Score). #12 of 61 desktop cards in this vertical.
| GPU | Relative | % | AI Score |
|---|---|---|---|
| NVIDIA GeForce RTX 3090 | 164% | 7.7 | |
| GeForce RTX 5080 | 111% | 5.2 | |
| NVIDIA GeForce RTX 4080 | 102% | 4.8 | |
| GeForce RTX 4080 Super | 100% | 4.7 | |
| GeForce RTX 5070 Ti | 100% | 4.7 | |
| NVIDIA RTX 4000 (Ada Generation) | 94% | 4.4 | |
| NVIDIA GeForce RTX 4070 Ti Super | 91% | 4.3 | |
| AMD Radeon RX 7900 XTX | 79% | 3.7 | |
| GeForce RTX 5060 Ti | 79% | 3.7 |
Same card, other workloads: GeForce RTX 5070 Ti Gaming benchmarks
← All AI & Machine Learning GPU rankings
| Transistors | 45,600 million |
| Die size | 378 mm² |
| Process node | 4 nm |
| Fabricated by | TSMC |
| Transistor density | 120.6 million per mm² |
Denser than 92% of the 746 cards we have silicon data for. Density is the clearest measure of what a process node bought: a card that gained it without growing the die got its speed from the fab rather than the architecture.
Silicon figures from Wikipedia (CC BY-SA 4.0). Benchmarks on this page are our own. Compare every chip.
Whole-job timings, composed from our measured per-model results on this card.
| Workflow | Time | Energy | Basis |
|---|---|---|---|
| Full codebase review | 12 min | 44.78 Wh | measured |
| Animate a batch of images | 26.1 min | 79.16 Wh | measured |
Can't run: Product photo shoot (needs FLUX.1 Kontext dev), Photo restoration batch (needs FLUX.1 Kontext dev), Restore and enlarge photos (needs FLUX.1 Kontext dev), Character sheet, 12 poses (needs FLUX.1 dev), Short social clips (needs Qwen3 32B), 60-second AI short film (needs Qwen3 32B), 24-frame storyboard (needs Qwen3 32B), 6-panel comic page (needs Qwen3 32B), Long-form article batch (needs Llama 3.3 70B).
This card is $749 to buy. The cheapest listed rate on Vast.ai is $0.149/hour, but that is the floor: we budget $0.179/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 4,189 GPU-hours. Below it you are paying for idle silicon.
| How you would use it | GPU-hours a year | Rental cost a year | Time to break even |
|---|---|---|---|
| 2 hours a day, hobby | 730 | $131 | 5.7 years |
| 8 hours a day, working on it | 2,920 | $522 | 1.4 years |
| 24/7, always-on agent | 8,760 | $1,566 | 5.7 months |
At hobby usage this card is very unlikely to pay for itself before it is superseded. Rent it. Rental figures include a 20% premium over the cheapest listed rate. Ignores electricity, resale and the fact that a rented card can be a newer one tomorrow.
Cheapest of the RunPod and Vast on-demand rates we see, sampled daily. Spot and interruptible pricing runs lower.