R520 will be able to do Physics nearly as AGEIA do.

[Rakeesh]

I think you're misunderstanding what they were saying. You don't just have one 1800XT and multitask it to graphics and physics, you have a seperate graphics card for graphics, they you plug in another graphics card and use it for physics.

Are you sure about that? I mean that would be pretty dumb on their part. Why would somebody buy a graphics card for use as a physics card when the physics card is cheaper and isn't going to require an additional PCI-x16 slot?

Besides, for all intents and purposes Nvidia could do the exact same thing.

Well, ATI would tell you that their R520 is GREAT at it since it can handle 512 threads.

Hmm...That kinda sounds like asyncronous piping rather than multitasking in the sense that I implied.

THey dont have the power to handle *heavy games* and at the same times physics.

They probably do. It has been well established that GPUs are far more powerful than current processors, and not only that but GPUs already handle many vector functions natively whereas CPU's don't. I guarentee you that a GPU from even the last generation would be able to perform physics calculations multitudes faster than a current generation CPU.

[Subtestube]

I think what NV40 is saying Alpha, and I kind of agree, is that if you've got a game that's already GPU limited on the highest end GPU out there, there won't be processing time for the GPU to do physics as well. Yes, it would be MUCH faster than a CPU for doing many (although not all) physics calculations, but if it's already busy all the time, when are you going to do it? You'll just find yourself even more GPU bound than you were before.

[bkswaney]

I think what NV40 is saying Alpha, and I kind of agree, is that if you've got a game that's already GPU limited on the highest end GPU out there, there won't be processing time for the GPU to do physics as well. Yes, it would be MUCH faster than a CPU for doing many (although not all) physics calculations, but if it's already busy all the time, when are you going to do it? You'll just find yourself even more GPU bound than you were before.


I agree. This is why u need a card just for the physics.

[nutball]

I guarentee you that a GPU from even the last generation would be able to perform physics calculations multitudes faster than a current generation CPU.

That depends what physics you're trying to accelerate. Having bootloads of vector floating-point maths capability isn't the be-all and end-all, there's more to it than that. It's too simplistic to look at headline GFLOPs and draw a conclusion (a mistake people seem to be insisting on making with regard to Cell).

CPUs are much more flexible and cope much better with branchy code than GPUs, for example. Some algorithms are branchy, period. If you've got a branchy algorithm, a CPU is a better choice.

Likewise GPUs don't support random-access write operations (well R520 does, but it's a first, and the operations are uncached so who knows what the performance will be), so if you need these you're stuffed.

There are many other limitations (look at the slides from the presentation by the Stanford Uni guy). GPUs are faster than CPUs for a certain sub-set of calculations, but by no means all of them.

[Subtestube]

Likewise GPUs don't support random-access write operations (well R520 does, but it's a first, and the operations are uncached so who knows what the performance will be), so if you need these you're stuffed.


Nutball: I agree with pretty much everything you said, but do you have a link for the R520 supporting Random Writes? I'm VERY interested in that if it's true, and hadn't seen anything regarding it anywhere. I know the the R500 has a "Memexport" function which lets you read and write to random locations in main memory, but I hadn't seen anything about the R520 supporting it! I mean, you're right, the perf could be atrocious, but I'd be interested to read a little about it.

[nutball]

Here's a link to the slides (http://graphics.stanford.edu/~mhouston/public_talks/R520-mhouston.pdf) I referred to from the Stanford Uni guy. Page 21 onwards talk about general-purpose computing on X1800, and page 23 talks about "scatter", ie. operations like a[i]=x.

[Subtestube]

Wow... veeeery interesting stuff - cheers for that link. I must say, having seen that, I'd be really interested to have a play with an X1800. Random writes would be awesome - they'd make many things much, much, MUCH easier. You're right though - it'll be the performamce that makes the difference! Nevertheless - stuff like that makes GPGPU for physics much easier. I still don't buy it as all that useful to games though - video cards are under enough stress as it is.

[retsam]

i say this physics stuff is useless unless there is a standard API for all graphic cards and then all said cards have to be compliant...

[nutball]

It won't be useless, but it won't advance as quickly as it would (it'll be like the old days of proprietary graphics APIs, eg. Glide). If MS really is coming up with DirectPhysics as rumours suggest... well then there's the "solution". Is there nothing that a bit more control-freakery by Microsoft can't fix?

[retsam]

It won't be useless yes it will.... how many programmers are gonna write code for 1% of the market.... there needs to be a standard to follow that every graphics vendor can follow.... but now the question becomes..how much performance are you willing to give back inorder to do physics...this sounds like the job for a dedicated asic....

[Soylent]

yes it will.... how many programmers are gonna write code for 1% of the market.... there needs to be a standard to follow that every graphics vendor can follow....

Like say "Direct Physics" and included in the direct X sdk? Did you read the post beyond the first sentence?

but now the question becomes..how much performance are you willing to give back inorder to do physics...this sounds like the job for a dedicated asic....

This will take a lot of time. The x1x00 series will be old news when and if there's a direct physics component(which is quite likely, since MS did indeed put out a job add searching for highly qualified GPGPU people to do physics on GPU. Although that could mean xbox 360 or something).

This means that many people will actually have an old x1x00 series card and a newer card. Although a real PPU would consume far less power just reusing an old card may be far cheaper.

[ricercar]

how many programmers are gonna write code for 1% of the market...All the programmers who are passionate about that market. Consider the coders who write for the Macintosh, which has had at maximum heyday, 11% of the market.

[retsam]

Like say "Direct Physics" and included in the direct X sdk? Did you read the post beyond the first sentence?umm did you not read what was said ...direct physics is still in development...

this is what i said ...
there needs to be a standard to follow that every graphics vendor can follow


All the programmers who are passionate about that market. Consider the coders who write for the Macintosh, which has had at maximum heyday, 11% of the market. good point, but the point that i guess i was trying to get at is that alot of programmers are gonna look and hack around with this and wait for an api. i think then they will release products to the public...

[nutball]

yes it will.... how many programmers are gonna write code for 1% of the market.... there needs to be a standard to follow that every graphics vendor can follow....

It only takes two programmers ... Sweeney and Carmack. With so many FPSs being built on top of third-party engines these days, it may not be such a big deal. If those devs are interested in using the latest cool tech to make their engine better than the other guy, it'll receive support in their engines, and from there almost by magic into games derived from their engines.

[Raje]

Are they kidding or something? They are basically saying that just because their GPU's process vectors, they can just arbitrarily use it to do physics calculations. It sounds as though they are trying to imply that you'll have a full blown PPU just by owning an ATI card. Marketing spin at its finest.

Heres an interesting question though, how well are current GPU's designed for multitasking like that?

AEGIA might have a run for it's money if ATI has anything to say about it. One AIB commented today that the idea of a dedicated scalar mathematics processor for game physics could already be replicated on ATI's R520 series silicon, although drivers for such a project only exist in R+D departments (the vendor wouldn't let us have them, we tried). The idea of offloading math to a GPU is not a new idea; many projects exist for Linux for this already. However, the indication we had was that ATI could actually do physics calculations on the card with the graphics processing simultaniously -- the bandwidth is already there. AEGIA's physics processor has already been delayed well into Q2 next year.

Clicky (http://www.anandtech.com/video/showdoc.aspx?i=2574)