Examples for parallel frameworks

Raytracer

To test the performance of individual frameworks and hardware, I have used a simple raytracer as a visual benchmark.

A raytracer is well-suited, because

it is easy to parallelize, because every pixel can be processed independently
but the workload for every pixel may be different. If the ray hits an objects, the reflected ray gets followed as the next step.

Comparison

The following table contains the results.

Language	Technique	CPU/GPU	Cores/SMs/Threads	Avg FPS/s	Min FPS/s	Max FPS/s
C++	Sequential	CPU	1 Core = 1 Thread	--.--	--.--	--.--
C++	POSIX Threads (pthreads)	CPU	4 Cores * 2 Hyperthreads = 8 Threads	--.--	--.--	--.--
C++	OpenMP	CPU	4 Cores * 2 Hyperthreads = 8 Threads	--.--	--.--	--.--
C++	NVIDIA CUDA	GPU	16 SM = 512 Cores = 512 Threads	--.--	--.--	--.--
C++	OpenCL	GPU	16 SM = 512 Cores = 512 Threads	--.--	--.--	--.--
C++	Cell-Processor	CPU	6 SPUs	--.--	--.--	--.--
Haskell	Sequential	CPU	1	--.--	--.--	--.--
Haskell	`par`	CPU	4 Cores * 2 Hyperthreads = 8 Threads	--.--	--.--	--.--
Java	Sequential	CPU	1	--.--	--.--	--.--
Java 7	`ForkJoinPool`	CPU	4 Cores * 2 Hyperthreads = 8 Threads	--.--	--.--	--.--

Do you have questions? Send me an email!

Copyright © 2007-2012 Jörn Dinkla. All rights reserved.

Last modified: