CUDA 2.0: impressions so far? The good and the bad of 2.0

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,

I’m still happily working with CUDA 1.1 (XP, VS2005); I’m now using regular 175.21 WHQL drivers on my G92 and all goes well.

I’m curious about CUDA 2.0: which benefits would I get, considering my environment (XP/G92) and the fact that I mainly work with CUDAFFT?

Are there performance improvements on G92 hardware?
Is CUDAFFT better/faster?

I’ve seen that some people suffers with problems with the supplied 177.84 driver, by the way (performance drops, higher CPU usage etc.).

So, well, it’s sort of a poll I guess. :) Please tell me about your experiences going from 1.1 to 2.0 final.

Thanks,

Fernando

2

1.1 → 2.0 final was painless for me. Zero code changes and zero performance delta across all platforms.

The only small problem I have with the 177.67 drivers (on linux) is that they down-clock one of the GPUs in the Tesla D870 into a power saving mode. Hopefully this will be fixed in the next driver release.

3

This is Windows user opinion.

2.0 Pros:

  1. Better integration with MSVC++ IDE (double click on the error or warning message in the output window highlights the place in code).
  2. Nvcc supports exceptions now.
  3. Visual profiler is useful and clear tool.
  4. I did not notice performance problems with 2.0.

2.0 Cons:

  1. Secondary cards watchdog problem is not solved (but promised to be solved), this is very annoying.
  2. Drivers look very crude. In addition to watchdog and performance problems there are always some issues with hardware support. For example, my primary card is 6600 (monitor is connected to it), 2.0 final driver does not recognize it (so it is necessary to install newer drivers over old ones). Such problems are not novel, beta drivers had same defects.