NVIDIA Officially Launches CUDA 9.0

[phoronix]

Phoronix: NVIDIA Officially Launches CUDA 9.0

NVIDIA has kicked off their GTC world tour and in beginning that, they have launched the CUDA 9.0 toolkit...

http://www.phoronix.com/scan.php?page=news_item&px=CUDA-9.0-Launches

[polarathene]

Will there be any benchmarks done to see how much of a perf difference CUDA apps have with 8 vs 9? Is there likely to be much perf gain for proprietary apps that use CUDA?

[bug77]

Will there be any benchmarks done to see how much of a perf difference CUDA apps have with 8 vs 9? Is there likely to be much perf gain for proprietary apps that use CUDA?

I don't think that's how it works. A newer version of an API usually brings new functions for speeding things up/simplifying implementations. Until apps are updated to actually use the newer API, you won't see much difference.

[sdack]

I don't think that's how it works. A newer version of an API usually brings new functions for speeding things up/simplifying implementations. Until apps are updated to actually use the newer API, you won't see much difference.

CUDA isn't just an API, but it's a whole bunch of libraries plus compiler support and more. So there will certainly be gains and losses to be found in all of this. But I agree with you in that it first requires software, which makes use of it before one can measure any differences. Often is CUDA used to implement custom algorithms on a GPU and as long as these algorithms themselves haven't changed will there be little difference.

Previous versions of CUDA used gcc 4.9 and the latest version allows the use of gcc 6 and possibly even gcc 7. So perhaps there is a chance that a few custom algorithms might see an improvement. Chances are though most of them have been hand-tweaked and are simple enough not to make a difference in the choice of compiler.

My guess is that differences will most likely be found with software, which uses Nvidia's libraries delivered with CUDA. So can ffmpeg make use of CUDA/NPP to scale images on a GPU during transcoding (h/w decode -> h/w scale -> hw/encode). Darktable also uses parts of CUDA, but I currently don't know which libraries exactly it's using. So in theory would one have to benchmark these applications with PTS to see if it makes a difference. Nvidia did promise speed ups in CUDA/NPP, but I haven't noticed anything major yet.

[kenjitamura]

What's all that dam hammering? Oh right. Nvidia putting another nail in that OpenCL coffin.

[defaultUser]

Will there be any benchmarks done to see how much of a perf difference CUDA apps have with 8 vs 9? Is there likely to be much perf gain for proprietary apps that use CUDA?

The gains depend on the applications and also is important to notice that most of the major CUDA versions are launched to coincide with the introduction of a new architecture, in this case the Volta. Due to the tensor cores on this architecture the performance gains for neural nets can be greater than ten times the performance when compared with the P100. Other very important features includes support for C++14, cooperative groups and new Nvidia libraries like the NCCL 2.0. And since the the Pascal architecture the Nvidia is pushing the towards better integration between the different memory spaces of GPU and CPU. When the the inclusion of the HMM patches on the kernel even an array allocated with malloc will be available to use on the GPU.

[speculatrix]

Fedora 25 support. lets hope this marks the point where they keep up to date more.
I resisted the urge to be snarky.

[GdeR]

Fedora 25 support. lets hope this marks the point where they keep up to date more.
I resisted the urge to be snarky.

That has been my main concern for years. Fedora 25 is reaching its EOL really fast..