What are the differences between CUDA compute capabilities?

Question

What does compute capability 2.0 add over 1.3, 2.1 over 2.0, and 3.0 over 2.1?

Tigerish

The relevant section of the wikipedia article has a reasonably good overview. – Cromagnon

I voted to reopen. I think the question is on topic. – Vaporimeter

The question is absolutely on topic. Voting to reopen again. – Artichoke

Now cuda compute capability 6.1 (as well as 6.0, 5.2) is out, and I would like to learn what's new. Voting to reopen the question to enable new answers and editing. The question is absolutely on topic and affects programming a lot: some CUDA frameworks set limits on the minimal compute capability. – Philipines

Answer 1

The Compute Capabilities designate different architectures. In general, newer architectures run both CUDA programs and graphics faster than previous architectures. Note, though, that a high end card in a previous generation may be faster than a lower end card in the generation after.

From the CUDA C Programming Guide (v6.0):

Vaporimeter

Thanks, that helps. It's a shame they're splitting the compute and graphics lines, as I suspect that will make the compute lines way more expensive since they'll not benefit from the cost advantages that come with being for the mass graphics market. – Tigerish

given this, does it make any sense to compile for a lower compute capability (eg 1.1 instead of 3.0), even if your card supports 3.0 - will it run faster? – Disadvantage

@user1191840: Why would it run faster if you compile for an older architecture? – Vaporimeter

no idea why, but I thought that's what I observed, yet to go back and test it properly though. I thought it could be possible just because nvidia is trying to limit compute performance on later (geforce) cards, and this could be linked to compute version. – Disadvantage

Kindly refer to the official reference of Compute Capability for more updated information. – Extravagancy