Hard limit of 64 regs/thread on Fermi? fermi registers limit

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey,

So is there a ‘hard limit’ of 64 registers per thread on the Fermi…I have a program that uses more than 64 regs/thread on the GTX 285 but it’s apparently limited to 64 registers/thread on the Fermi (I’m using the 470…)…anyone know anything about this/is there any documentation available about this? (I see a reference to it here: [url=“http://forums.nvidia.com/index.php?showtopic=167735”]http://forums.nvidia.com/index.php?showtopic=167735[/url] but can’t find anything official…)

Thanks!

2

Yes, this is correct, although on the other hand spilling registers to local memory is a lot faster on Fermi because of the L1 cache.

3

Thanks for the response, any reason for this/might it be changed in future releases/architectures?

4

It certainly might change, although I don’t imagine it would go lower. It’s a constant balancing act between the RAM sizes, registers per thread and the number of threads per multiprocessor.