Which is more advanced, GPU's or CPU's? (Gamewise)

[jonmullen]

I am a doctoral student in Computer Science working with embedded systems. Lots of robotics and networked sensors. As a hobby though I have been working on writing optimized crypto libraries for the PS3, which is what brought me to do most of my research in to Generalized Computer on Graphic Processing Units of GPGPU. If your interested Sh is real interesting project, it has all but been abandoned though for its commercial implementation RapidMind Not to spew more marking hype but since they went commercial they shut out developers and academics so marking is all I can really show you on their current efforts, but they show huge performance advantages in GPU hardware over top of the line CPU in Financial and Media Processing applications, both of which provide many opportunities for parallel optimization. What is neat is that their meta language compiler works at run time on top of C++ and optimizes for your hardware. You have a good ol' P3 it will run have a nice Quad-Core Xeon it will make use of all 4 cores no problem, have a dual Cell with two SLI new graphic cards it will spit the work among your four G5 cores on the Cells, it will also feed work to the 16 SPE between the two and even have some work for the two graphics card to do. Not to mention the operations that are done on the graphic cards are converted to OpenGL code first, and it still beats the pants off a top end CPU. Granted that kind of performance increase can't be seen across the board, but none the less they are some quite impressive numbers.

 

[BrownTown]

Yeah, I thought you were probably comming at this from a CS point of view, so my question is how much of the hardware do you actually understand here? I mean for example I am an EE and know hardware, and I do know C++ and JAVA too (along with VB but that hardly counts ), but I've never done anything multithreaded, most difficult C++ program I've ever made was a single threaded raytracer an the most difficult Java program was a raycaster.

 

[degibson]

Easy guys... lets not get tripped up about terminology here. Whether GPUs sport bona fide threads or not, nobody can argue that the amount of parallelism, even reasonably-programmable parallelism, on modern GPUs is astounding. However, no GPU is yet ready to run arbitrary C code... different strokes as it were.

 

[jonmullen]

I will be the first to say it, I am not one really for hardware design, but I do know quite a lot about how the hardware processes instructions most for optimization purposes. Like in the cell you have to know which PPC instructions are micro-encoded rather than hardware encoded thus forcing the processor to flush both pipes before it has to go fetch the instruction from memory, or that if I hint towards a taken branch at least like 10 cycles before the branch I can reduce my branch penalty to effective one cycle and stuff like that.

The main thing I just really want to get out there is the idea that the hardware is doing great, but we are going to run into an issue of getting programmers to be able to code parallel programs, as that seems to be where the hardware is going, and despite their age they are not well suited to or scalable enough to be the programming model for utilizing multi codes, but rather trying to think of how you can do the same thing in a parallel nature, or at least in a way a good compile can vectorize it.

 

[degibson]

Originally posted by: jonmullen
...
The main thing I just really want to get out there is the idea that the hardware is doing great, but we are going to run into an issue of getting programmers to be able to code parallel programs...?

In my opinion, yes. This will be a major challenge for next-generation software.

 

[RobertL]

Originally posted by: jonmullen
I am a doctoral student in Computer Science working with embedded systems. Lots of robotics and networked sensors. As a hobby though I have been working on writing optimized crypto libraries for the PS3, which is what brought me to do most of my research in to Generalized Computer on Graphic Processing Units of GPGPU. If your interested Sh is real interesting project, it has all but been abandoned though for its commercial implementation RapidMind Not to spew more marking hype but since they went commercial they shut out developers and academics so marking is all I can really show you on their current efforts, but they show huge performance advantages in GPU hardware over top of the line CPU in Financial and Media Processing applications, both of which provide many opportunities for parallel optimization. What is neat is that their meta language compiler works at run time on top of C++ and optimizes for your hardware. You have a good ol' P3 it will run have a nice Quad-Core Xeon it will make use of all 4 cores no problem, have a dual Cell with two SLI new graphic cards it will spit the work among your four G5 cores on the Cells, it will also feed work to the 16 SPE between the two and even have some work for the two graphics card to do. Not to mention the operations that are done on the graphic cards are converted to OpenGL code first, and it still beats the pants off a top end CPU. Granted that kind of performance increase can't be seen across the board, but none the less they are some quite impressive numbers.

Can you show any concrete evidence an GPU can calculate non graphical tasks faster than a CPU? Or are you talking out of your arse?

 

[BrownTown]

Originally posted by: RobertL
Can you show any concrete evidence an GPU can calculate non graphical tasks faster than a CPU? Or are you talking out of your arse?

A GPU can calculate SOME non-graphical tasks faster than a CPU (like 10 times faster), but they are all tasks which are very similar to graphics in that they require tons of very simply and predictible operations on a dataset. A GPU is more or less a very complicated DSP chip and any sort of signal processing will do very good on it. However very branchy code will run very slow because a GPU is not designed for it. A GPU can run SOME things very fast, a CPU can run EVERYTHING pretty fast.

 

[degibson]

Originally posted by: RobertL

Can you show any concrete evidence an GPU can calculate non graphical tasks faster than a CPU? Or are you talking out of your arse?

There are some links to case studies on www.gpgpu.org, I believe. I have seen performance comparison on numerical codes that favor GPU-based computation over reasonable CPU baselines.