Do CUDA cores make a difference in speed?

I have a GTX 560 Ti (384 CUDA cores).

If I upgrade to a GTX 660 Ti (1344 CUDE cores), I understand it will make a difference since it is a more advanced video card, but should I be looking at CUDA cores as a comparison?

How can I plan for a nicely optimized system for Cycles renders?

  1. The more CUDA cores and the higher their clocks, the faster the card.

A very good indicator is to check the wikipedia pages on the Nvidia card generations and their FMA GFLOPs. It doesn’t tell you how fast it’ll render, but how fast the cards are relative to each other.

The speed is also not really comparable between different generations if they have another architecture.

  1. The more memory the card has, the better for your rendering.

All the geometry and material information (including textures) has to go to the video card memory before starting the render.
While a scene is not too large, the textures can soon limit you.

You simply check the GFLOPS, the Blender Render Benchmark thread.
Then check how many Watts/GFLOP the card consumes.
How many Euro/GFLOP it costs.

Then choose the “best” card for your budget.

To be specific, the 660 will actually be slower, because the architecture doesn’t fair as well in Cycles. Core count is more important than clock speed, but it’s only comparable within the same architecture. In my opinion, an upgrade to anything less than a GTX780 isn’t worth it.