Need someone to break down info on stream processors

[poiZin]

So I've read wikipedia's explanation of stream processing, but I am not at that level of understanding yet.. If someone could kindly break that down into something that is easier understood, it would be greatly appreciated.

For the most part I understand how a CPU works, maybe compare stream processor to a CPU as well?

Thanks in advance!

 

[poiZin]

Ok so.. I think i figured it out. To anyone else who might be curious about this, here you go. (Please correct me if I am wrong)

A stream processor only needs to perform specific operations (such as vertex operations, geometry operations, and pixel operations working with floating points), where as a CPU is designed to perform a series of extensive operations and arithmetic simultaneously.

Don't quote me on this but I believe a single stream processor can perform more operations per clock cycle than a CPU.

The reason that so many stream processors are required to get the job done (ie the newer cards like the 4870 has 800 stream processors) as opposed to 1 CPU on a system is because some software (mostly intense graphical games like Crysis Warhead) requires the GPU to perform tons of these operations at once.


Now really quick, as i said earlier the Radeon 4870 advertises 800 stream processors, while the GTX 280 advertises 240. I was a bit confused about this too, but what I think is Nvidia's stream processors are a bit more powerful than ATI's. I also think that ATI may be advertising each processor as if they were 5 (logical units) because each stream processor can perform 5 operations per clock cycle. So in ATI's language, the GTX 280 would equal 1200 (240*5)


So using CPUs on a video card would be a total waste. There is the reason why the price of 1 CPU is around the same price as a video card containing hundreds of stream processors.

AGAIN- I am not saying this is correct, this is how I perceive it. If you know this is wrong, please enlighten us with your reply.

 

[Martimus]

Not quite, but here is an article that will help explain the ATI streaming processor. Anandtech Article on the 4850 and 4870.
Here is one for the nVidia streaming processor. Anandtech article on the GT200.

 

[poiZin]

AHHH.. Very interesting! I thank you for that link, that made it a lot clearer.

Do you have any idea if I am correct about the difference between a CPU and stream processor?

 

[Martimus]

You are pretty much correct about the difference. A CPU uses the X86 instuction set, while the stream processor has its own instruction set. Far more simplistic. (Obviously, each core is only a fraction the size of a logical Intel or AMD core.)

 

[Nathelion]

Originally posted by: poiZin
Ok so.. I think i figured it out. To anyone else who might be curious about this, here you go. (Please correct me if I am wrong)

A stream processor only needs to perform specific operations (such as vertex operations, geometry operations, and pixel operations working with floating points), where as a CPU is designed to perform a series of extensive operations and arithmetic simultaneously.

Don't quote me on this but I believe a single stream processor can perform more operations per clock cycle than a CPU.

The reason that so many stream processors are required to get the job done (ie the newer cards like the 4870 has 800 stream processors) as opposed to 1 CPU on a system is because some software (mostly intense graphical games like Crysis Warhead) requires the GPU to perform tons of these operations at once.


Now really quick, as i said earlier the Radeon 4870 advertises 800 stream processors, while the GTX 280 advertises 240. I was a bit confused about this too, but what I think is Nvidia's stream processors are a bit more powerful than ATI's. I also think that ATI may be advertising each processor as if they were 5 (logical units) because each stream processor can perform 5 operations per clock cycle. So in ATI's language, the GTX 280 would equal 1200 (240*5)


So using CPUs on a video card would be a total waste. There is the reason why the price of 1 CPU is around the same price as a video card containing hundreds of stream processors.

AGAIN- I am not saying this is correct, this is how I perceive it. If you know this is wrong, please enlighten us with your reply.

Stream processing is a SIMD (single instruction, multiple data) programming paradigm, nothing more, nothing less. The entire idea is that the stream processor accepts a large number of data elements and a small program (a "kernel", at least in ATI terminology) that needs to be run on all the data elements. The key part is that the execution of the kernel on one data element can't be dependent on any other data element, so the data can be processed in any order or in parallel. The stream processor then runs the kernel on all the data elements (transparently to the programmer, aka. the scheduling is entirely done by the stream processor or intermediate software stack).

 

[BEL6772]

Another analogy that might help:

A CPU is like a swiss army knife, and a graphics processor is like a fillet knife.

If you have to clean and fillet a fish, there's no better tool than a fillet knife ... that's what it was made to do.

Sure you could use that fillet knife to dig a cork out of a bottle, turn screws, etc. But it doesn't do those things very efficiently.

A Swiss army knife has tools built in especially for pulling corks, turning screws, etc. You could even use it to clean and fillet a fish, but it wouldn't do that as easily as your fillet knife.