Dr Donald Kinghorn (HPC and Scientific Computing)

NVIDIA GPU Starter DevKit with OpenACC

Written on June 2, 2014 by Dr Donald Kinghorn

One of the things we were showing off at the 2014 GTC meeting was the Peak Mini with what would become the NVIDIA GPU Starter Dev Kit. This is something we have been talking with NVIDIA about since SC13 last November. It is now formally released. This is a nice system bundle with a Tesla K20 and the PGI Workstation compilers with OpenACC support. The bundle is priced to bring the system cost down to under $5000. It’s a pretty good deal since the K20 and PGI compilers alone would normally cost more than that. 

The main idea is to get good quality production ready hardware and software into the hands of developers who may have hesitated at in the past on trying their hand with CUDA programming. OpenACC with the PGI compilers gives these developers a directive (pragma) based GPU programming methodology very similar to OpenMP. I think OpenACC has serious potential to bring GPU acceleration to a lot more projects. 

What do you get?

  • A Peak Mini with a Intel Haswell CPU
  • NVIDIA Tesla K20
  • PGI Workstation Accelerator with OpenACC plus 1 year support subscription
  • CUDA 6  Toolkit
  • PGI CUDA Fortran
  • CentOS 6.5 developer setup
  • Learning resources

Everything pre-installed, configured and tested!

Why is this interesting?

The main reason this is nifty is that with the PGI compilers with OpenACC you can write code like the following and it will execute on the GPU… simple directives, nice! (yes, that’s fortran, I like it!)  This probably an overly simple code snippet but you get the idea.

!$acc kernels loop
do i=1,n
    vecY(i) = alpha * vecX(i) + vecY(i)

The DevKit lowers the barrier to experimentation with OpenACC on Tesla. I’ve been interested in this for some time but have had a lot of projects on my plate. I’m hoping to get to spend some quality time with the PGI compilers now. I’ll keep you posted on how it goes.

Happy computing --dbk

Tags: Nvidia, Tesla, HPC