NVIDIA GeForce GTX 1660 Super
Turing mid-range with GDDR6 — bandwidth jumps to 336 GB/s vs the base 1660. Same 6 GB VRAM ceiling but ~30-45 tok/s on 7B Q4 thanks to the bandwidth bump. Still no Tensor cores. Strong used-market value in 2026 ($140-160).
Affiliate disclosure: as an Amazon Associate and partner of other retailers, we earn from qualifying purchases. The verdict on this page is our editorial opinion; affiliate links never influence what we recommend.
Sub-scores sum to 373 / 1000. Headline = 373 × 0.70 (Estimated-confidence discount) = 261. This is an algorithmic performance-tier score — distinct from, and often lower than, the editorial “Our verdict” below, which weighs value and real-world fit (especially for hardware we haven’t measured yet). How scoring works →
Extrapolated from 336 GB/s bandwidth — 40.3 tok/s estimated. No measured benchmarks yet.
Plain-English: Edge-of-fit for 7B; expect compromises.
Verdicts extrapolated from catalog VRAM + bandwidth + ecosystem flags. Hover any chip for the rationale. Want measured numbers? Submit your own run with runlocalai-bench --submit.
This card is for the operator who needs a cheap inference runner for small models and already has a CUDA stack in place. The 6 GB VRAM ceiling and lack of Tensor cores mean it's strictly a budget workhorse, not a platform for serious local AI.
On 7B Q4 models, the 336 GB/s bandwidth delivers roughly 30-45 tok/s — usable for chat and code completion, but not for real-time streaming. 13B Q4 models (9 GB) won't fit at all, and even 8B Q4 models (~5.5 GB) leave almost no room for context.
What breaks: anything above 6 GB VRAM. No Tensor cores means no FP8 or INT4 acceleration, so the card relies entirely on CUDA cores for compute. Software stack is limited to CUDA-only runtimes; no ROCm or Vulkan support out of the box.
Pass on this card if the workload includes 13B+ models, long-context inference, or any training/fine-tuning. The 6 GB ceiling is a hard wall, and the lack of Tensor cores makes it obsolete for modern quantization formats.
At $150 used, it's a fair price for a disposable inference node, but a used RTX 3060 12 GB for $180 is a far better long-term investment.
›Why this rating
The GTX 1660 Super earns a 4.5 for its niche as a cheap, low-power inference runner for small models. However, the 6 GB VRAM and missing Tensor cores severely limit its usefulness for modern local AI workloads, making it a poor choice for any serious operator.
Overview
Turing mid-range with GDDR6 — bandwidth jumps to 336 GB/s vs the base 1660. Same 6 GB VRAM ceiling but ~30-45 tok/s on 7B Q4 thanks to the bandwidth bump. Still no Tensor cores. Strong used-market value in 2026 ($140-160).
Some links above are affiliate links. We may earn a commission at no extra cost to you. How we make money.
Specs
| VRAM | 6 GB |
| Power draw (peak) | 125 W |
| Released | 2019 |
| MSRP | $229 |
| Backends | CUDA Vulkan |
Models that fit
Open-weight models small enough to run on NVIDIA GeForce GTX 1660 Super with usable context.
Frequently asked
What models can NVIDIA GeForce GTX 1660 Super run?
Does NVIDIA GeForce GTX 1660 Super support CUDA?
How much does NVIDIA GeForce GTX 1660 Super cost?
Where next?
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify hardware specifications.