8GB · AI Score 2.2/100 · first-party measured on 12 AI workloads
2.2 AI Score Includes estimates
Every number on this page is first-party: NVIDIA GeForce RTX 3070 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) NVIDIA GeForce RTX 3070 Ti delivers about 102.13 tokens/sec. Llama 3.3 70B does not fit. It needs roughly 42GB and this card has 8GB. For image generation, SDXL runs at 1.88 it/s, while FLUX.1-dev won't fit at BF16 (needs ~26GB). 9 of the 12 workloads won't fit on 8GB at the tested precision, Qwen2.5-Coder 14B, Qwen3 32B, Llama 3.3 70B, Z-Image Turbo and others. We publish those as hard gates rather than quietly dropping to a smaller quant.
Hard card to recommend for AI: basically 3070 performance for 300W, the worst tokens-per-watt I measured on Ampere (0.79), and the same 8GB wall gating 9 of 12 workloads. The math just doesn't work. 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 |
|---|---|---|---|
| MiniCPM5 2B | 244.3 tok/s | 174 W64°CQ4_K_M | ✓ Measured |
| LFM2.5 2.6B | 243.25 tok/s | 169 W66°CQ4_K_M | ✓ Measured |
| Granite 4.1 3B | 178.48 tok/s | 198 W61°CQ4_K_M | ✓ Measured |
| Nemotron 3 Nano 4B | 160.46 tok/s | 189 W60°CQ4_K_M | ✓ Measured |
| Qwen3 4B | 154.65 tok/s | 2.9 GB peak196 W72°C0.79 tok/WQ4_K_M | ✓ Measured |
| DeepSeek Coder 7B Instruct v1.5 | 119.13 tok/s | 195 W60°CQ4_K_M | ✓ Measured |
| Llama 3 8B | 105.65 tok/s | 203 W61°CQ4_K_M | ✓ Measured |
| Llama 3.1 8B | 102.13 tok/s | 4.8 GB peak227 W72°C0.45 tok/WQ4_K_M | ✓ Measured |
| Qwen3 8B | 102.4 tok/s | 209 W62°CQ4_K_M | ✓ Measured |
| Nemotron Nano 9B v2 | 76.23 tok/s | 213 W63°CQ4_K_M | ✓ Measured |
| Ornith 1.5 9B | 91.43 tok/s | 209 W63°CQ4_K_M | ✓ Measured |
| Gemma 4 12B | 65.01 tok/s | 216 W63°CQ4_K_M | ✓ Measured |
| Qwen2.5-Coder 14B | ✕ Won't fit needs ~11.5 GB | VRAM-gated at this precision | ✓ Measured |
| Qwen3 14B | ✕ Won't fit needs ~11 GB | VRAM-gated at this precision | Est. |
| 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 | 27.62 images/min | 257 W73°C | ✓ Measured |
| SDXL Turbo | 250.13 images/min | 153 W60°C | ✓ Measured |
| Sana 1.6B | ✕ Won't fit needs ~11 GB | VRAM-gated at this precision | Est. |
| Stable Diffusion XL | 3.76 images/min | 6.5 GB peak230 W72°C16 s/img | ✓ Measured CPU offload |
| Z-Image Turbo | ✕ Won't fit needs ~13 GB | VRAM-gated at this precision | ✓ Measured |
| PixArt-Sigma XL | ✕ Won't fit needs ~14 GB | VRAM-gated at this precision | Est. |
| 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 |
|---|---|---|---|
| LTX-Video (distilled) | ✕ Won't fit needs ~14 GB | VRAM-gated at this precision | ✓ Measured |
| Wan 2.2 5B (720p) | ✕ Won't fit needs ~18 GB | VRAM-gated at this precision | ✓ Measured |
| Architecture | Ampere (GA104) |
| CUDA cores | 6,144 |
| VRAM | 8GB GDDR6X |
| Memory bus | 256-bit |
| Memory bandwidth | 608 GB/s |
| Boost clock | 1,770 MHz |
| TDP | 290 W |
| Process | Samsung 8nm |
| Interface | PCIe 4.0 x16 |
| Release date | 2021-06-10 |
| Launch MSRP | $599 |
NVIDIA GeForce RTX 3070 Ti scores 2.2/100, #78 of 102. It ran 3 of 12; 9 exceeded its 8GB. Every figure here is our own measurement.
100% = this card, AI & Machine Learning headline metric (AI Score). #39 of 61 desktop cards in this vertical.
| GPU | Relative | % | AI Score |
|---|---|---|---|
| NVIDIA GeForce RTX 3080 | 105% | 2.3 | |
| NVIDIA TITAN Xp | 105% | 2.3 | |
| AMD Radeon RX 7700 XT | 100% | 2.2 | |
| NVIDIA GeForce RTX 2080 Ti Founders Edition | 100% | 2.2 | |
| NVIDIA GeForce RTX 3070 Ti | 100% | 2.2 | |
| NVIDIA GeForce RTX 2080 Super | 95% | 2.1 | |
| NVIDIA GeForce RTX 3070 Founders Edition | 95% | 2.1 | |
| NVIDIA GeForce RTX 5060 | 95% | 2.1 | |
| Intel Arc B580 | 95% | 2.1 |
Same card, other workloads: NVIDIA GeForce RTX 3070 Ti Gaming benchmarks
← All AI & Machine Learning GPU rankings
| Transistors | 17,400 million |
| Die size | 392.5 mm² |
| Process node | 8 nm |
| Fabricated by | Samsung |
| Transistor density | 44.3 million per mm² |
Denser than 80% 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.
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), Full codebase review (needs Qwen2.5-Coder 14B).
This card is $599 to buy. The cheapest listed rate on Vast.ai is $0.176/hour, but that is the floor: we budget $0.211/hour, a 20% premium, because idle time, storage and unavailable cheap instances all land on the same bill. At that rate buying wins after 2,836 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 | $154 | 3.9 years |
| 8 hours a day, working on it | 2,920 | $617 | 11.7 months |
| 24/7, always-on agent | 8,760 | $1,850 | 3.9 months |
At steady usage this card pays for itself inside a normal ownership window. 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.