Dedicated physics card or dedicated physics core?

[jshuck3]

Just curious what you guys think....do you think the market will tend toward dedicated physics cards or dedicated physics cores? I know one developer has already shown a physics core with Core 2 Quad, but I'm curious if anyone knows if other developers will be heading down that path.

 

[Gamingphreek]

I assume you mean General Purpose Cores on CPU's. It is true that CPU's and GPU's are becoming more and more general purpose, but I doubt that "Physics Cores" will ever be used. It is possible that Video Cards will incorporate a general purpose core of some sort that is dedicated to Physics Processing, but I don't see a CPU having that.

-Kevin

 

[Peter]

AMD want specialized coprocessors connected straight to interprocessor Hypertransport bus. It even has been done already for a couple of specialized applications (not games).

For anything that needs to be calculated before a frame is rendered, having a short and fast path to the RAM (where the data reside) is good. Hypertransport, being an open standard, lets anyone do such a thing. Who knows, maybe next year we'll upgrade AMD 4x4 boards from twin dualcore CPUs to one quad plus one physics coprocessor?

Being out on an I/O branch like PCIE is much less good - but it's the easiest and obvious way to abuse existing chips (like GPUs) as physics helpers.

 

[jshuck3]

I'm thinking like this. I have a quad core chip. The game developer says "physics, use core 3" and all that core does is do physics. I know some company demoed it recently where they had one of the cores on intel's quad stuff doing physics only. So I'm not talking a dedicated physics core, but more so a game is smart enough to see the machine has 4 cores available and just dedicates one of the cores to physics calculations when I'm playing their game.

 

[SunnyD]

Physics co-processing is a much more specialized piece of work than a CPU is suited for, just as graphics is much more specialized and the GPU is optimized for it. Think of it this way... physics uses something akin to SSE, as does graphics. The idea is to push a calculation on multiple objects in parallel... AS MANY of these objects at the same time. Be it pixels (pixel shaders), textures (TMU's), vertices (Vertex Shaders), etc., or physics "units".

While SSE allows several calculations to occur in parallel, video cards and PPU's are highly parallel devices, allowing usually 8 to now 100's of operations to done at once. CPU's are really meant to push one instruction on one piece of data at a time.

With that said, I highlly doubt that we'll see dedicating a core to physics. It's just not as capable of doing the job.

 

[40sTheme]

I believe physics cards will dominate in physics processing, mostly because of what SunnyD said, but developers are slowly becoming more accustomed to physics cards, and once large-scale development suited for heavy physics processing begins, PPUs will be a very important piece of hardware.

 

[kopema]

Really, it's just a question of packaging. You could put it anywhere you want: CPU, motherboard, separate card, etc. It's more of a marketing issue than a technical one.

A separate physics card will always cost a certain amount of money to build and advertising to sell. But if you can package a physics chip with something else, the production and marketing costs drop off quite a bit.

As I see it, lots of people want CPU's who don't have any use for physics processors. But pretty much everyone who buys a 3D accelerator card will also want to be able to do physics processing.

So maybe graphics card makers could go ahead and incorporate some physics processing capability. It would take some space and generate some heat, but I think it would be a good selling point. If, for example, nVidia had a decent physics chip on their latest generation of cards and Radeon didn't, I'd shell out a few extra bucks (or trade a new shading hue I can't see anyway) to take some load off my CPU, speed up my games and save me the time, trouble and slot a separate physics card would take.

Just think what that would do to video card reviewers' benchmarks. Some game framerates might be DOUBLED. I think that could help sell a card or two.

From a game designer's point of view, if one of the two major video cards added this, developers would start incorporating the algorithms within months instead of years. And if both major GPU producers did it, developers would start going nuts with 3D acceleration practically overnight.

 

[TanisHalfElven]

well the 8800 series already has onboard physics processing. in fact i think i read in one of the reviews that physics processing is part of the directx 10 specification (i can't prove this or remember where i read so don't jump on my throat if its not right)