Does anyone have any experience with the xcelerit API? Supposedly simplifies GPU programming. From reading the literature on their site (xcelerit sdk) it isn’t clear what it adds to the CUDA + thrust API.
Had a look and tried a few things with it. It seems quite capable and goes beyond the scope of CUDA and the like. Code written in a C-like language can execute on all sorts of processors, including but not limited to GPUs. It also supports running the application in a distributed system (cluster or something), using MPI under the hood. I guess the key difference is the programming model they use. It is based on dataflow, and that’s how you write your application. This enables them to automatically parallelise a lot of stuff from generic and sequential C code. A source-to-source compiler takes this code and has back-ends for various platforms (NVIDA, AMD/ATI, CPUs, clusters).
The performance they are getting is quite impressive. But don’t get overexcited - it generates CUDA/OpenCL source code or something. So it can’t be faster than carefully written manual code. The advantage is that you don’t have to do that and still get great performance.
Thanks, that’s very helpful.
Thanks, that’s very helpful.
Steve,
I do business development work at Xcelerit and if you (or anyone else on this forum) is interested in trying out the API, we
can set you up a virtual machine instance on one of our partner’s cloud-computing facilities that has the SDK and some sample code
pre-installed. It is a good way to experiment to see what speedups you can get on GPU hardware without having to battle with CUDA programming. Let me know if you are interested - or just click on www.xcelerit.com/xcelerit-in-the-cloud/
All the best,
Donal