Strong New Trend: GPGPU - Parallel Computing

Started by Theo Gottwald, January 08, 2016, 08:13:17 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Theo Gottwald

Last week i submittet my results to the Luxmark.info -Site.
http://www.luxmark.info/

And the result was:
1208 -> 12 Core -CPU (all Core srunning ...)
vs.
27908 -> HD 7990 GPU -> If i add another one i go up to 55820 ... thats 46x the 12 Core CPU.

Means, for many of todays Computing Tasks, The CPU is an outdated device.
ASM is not more the fastest way to go.
Even Multi-Threading on 12 Cores delivers alltogether just the "Measurement tolerance" of a modern GPU architecture (my GPU is from 2013).

PARALLEL COMPUTING.

OpenCL outruns even Parallel SIMD Code on my 12 Core CPU bei 30x in the current Luxmark-Test.

OpenCL makes another point: Its vendor independent and does not only run on GPU's from one vendor, but also for example on many other devices.
OpenCL runs on GPU's, CPU's, ASICs, DSP's ... on anything that can act as a Parallel processing device.

Many HPC-Application have to accept the fact that even most modern CPU's will just be able to reach the "measurement tolerance" of a modern GPU and OpenCL.

Thats why more and more Applications are calculated by the GPU "leaving the  CPU" as a helper device.

Thats a very strong Trend not only in Computer Graphics, but of course in Games, Physical Simulation (Gas-Simulations, Weather and Rigid Objects Simulations),of course in Raytracing, Encryption and much more...).

Unseen from the public, Open CL is already part of car-electronic, as well as on any handheld device.
Here are some introductions.





We have OpenCL Headers for PB from Jose and Peter Schreiber in Jose's Forum since ~ 10 Years.
While most often I think people will not use OpenCL directly but use "Libraries".

Who wants take a look ...

OpenCL Part I


Open CL Part II


OpenCL Part III


Another one from 2014: