Does Cell = PPU in Physics Calculations?

[Frank]

So why not just build a cell like device onto a board and call it a PPU? Given the price of the PPU and the required cost of a Cell to go in a console, it would be more than financially viable.

Exactly. I would be very surprised if the PPU isn't more or less a bunch of SIMD FP/vertex units with some general porpose logic for program control.

Bottom line, if Cells more powerful than a PPU then why did Tim say that with a PPU, PC's could have all the physics of PS3 games?

Because he wants to sell his game engine for all platforms?

With an NV2a, can the xbox have all the graphics of a high end PC?

Yes, as long as you limit the graphics to something an XboX with NV2A can run.

 

[pjbliverpool]

I say we just wait and see. It won't be long now.

Yes, as long as you limit the graphics to something an XboX with NV2A can run.

That makes no sense. High end PC's are capable of everything the xbox can do. The reverse is not true.

Tim suggested the PPU is capable of everything the next gen consoles can do. Again, the reverse is not neccerserily true.

Will we be decoding MPEG2 on the PPU? Im guessing no.

 

[randycat99]

So can this be taken that a PPU is a cutdown variant of an SIMD(s) such that it can only accept a specific kind of data and compute it in a specific kind of "physics calculation"? General math is "locked away" (even though that is essentially what it is occurring within), sort of like how a videocard is not typically setup to give you access to its mathematical resources? ...or maybe these are "fixed-function" SIMD(s), such that only a specific few operations are supported?

 

[Frank]

Tim suggested the PPU is capable of everything the next gen consoles can do. Again, the reverse is not neccerserily true.

He isn't saying that one of them will be better than the others, only that they can all do it.

 

[Frank]

So can this be taken that a PPU is a cutdown variant of an SIMD(s) such that it can only accept a specific kind of data and compute it in a specific kind of "physics calculation"? General math is "locked away" (even though that is essentially what it is occurring within), sort of like how a videocard is not typically setup to give you access to its mathematical resources? ...or maybe these are "fixed-function" SIMD(s), such that only a specific few operations are supported?

An ALU already contains a minimal set of math functions, which you can combine to do more complex math. And you can implement just about any function with just add, compare and some flow control.

Anyway, there is too much to physics, especially when done on a scene consisting of objects made up of polygons, that you could limit the functions to just a few very specific ones. Although they might have done some of the most often used ones directly in hardware, like MAD and DP3 on a GPU. But that doesn't mean that you cannot calculate those on another platform, and it would suggest that the clockspeed is brought down to give those complex functions enough time to execute.

 

[randycat99]

Well, that's just it, isn't it? We hear that whatever a PPU is on the hardware level it will be specialized to perform math extremely fast, but then we also hear that to really do physics, it needs to pretty much do all the different kind of math functions as any general purpose CPU. So that really begs the difference of how different a PPU can really be from a general purpose CPU we've known of for some time (or a "smart"-SIMD of some sort). Naturally, we then get staunch refusals that it could resemble a general CPU, because that would then be "slow"...

 

[Frank]

Well, a design like Cell would be a very good choice for physics. Lots of FLOPS for the actual calculations, and flexible enough to do it all.

 

[randycat99]

I would tend to agree, but then I've heard from others who are quite convinced that physics processing is somehow not ideal for Cell's team of SPE's, nor do they express any faith that any 1 of the 3 in-order PPC's in the XB2 could handle speedy physics. So the mind boggles at just what sort of hardware implementation a real PPU could take if the processing demands seemingly transcend that which could be accomplished in the impressive hardware configurations of these 2 consoles or even the coming multicore x86 incarnations in the PC scene.

Perhaps, this thing is simply a refreshened x87 math coprocessor, except now with multiple parallel pipelines (ala GPU style)? Maybe MIMD has some role in this, as well? ...just thinking out loud.

 

[Fox5]

Well, that's just it, isn't it? We hear that whatever a PPU is on the hardware level it will be specialized to perform math extremely fast, but then we also hear that to really do physics, it needs to pretty much do all the different kind of math functions as any general purpose CPU. So that really begs the difference of how different a PPU can really be from a general purpose CPU we've known of for some time (or a "smart"-SIMD of some sort). Naturally, we then get staunch refusals that it could resemble a general CPU, because that would then be "slow"...

Supposendly the PPU doesn't do actual physics calculations, but rather just approximations that are good enough for games, so it won't be doing the same math as cell.(or at least not cell as a workstation)

 

[randycat99]

...interesting, so some function approximations and maybe some table-lookups, perhaps.

 

[London Geezer]

Im sorry but i have to ask...

Who is going to notice the difference between a processor that performs, say, 35M "physics calculations" per second and another that does 45M?

What the hell is a "Physics caluclation" anyway? That sounds as uselessly generalised as saying "graphics calculation"... :?

All i see is that beauty of a game that is Heavenly Sword and really, this whole PPU thing get as redundant as the Realtime REYES craze.

A bit like Crazy Frog (by the way, thanks to all the f'king sad low-life bastards who made it a N.1 record )

 

[_xxx_]

A bit like Crazy Frog (by the way, thanks to all the f'king sad low-life bastards who made it a N.1 record )

We had "Sweety The Chicken" on No.1 for months last year. That one was even worse, it drove me completely nuts...

Thank MTV for that. They should burn in hell.

 

[Nite_Hawk]

Im sorry but i have to ask...

Who is going to notice the difference between a processor that performs, say, 35M "physics calculations" per second and another that does 45M?

What the hell is a "Physics caluclation" anyway? That sounds as uselessly generalised as saying "graphics calculation"... :?

All i see is that beauty of a game that is Heavenly Sword and really, this whole PPU thing get as redundant as the Realtime REYES craze.

A bit like Crazy Frog (by the way, thanks to all the f'king sad low-life bastards who made it a N.1 record )

I actually thought the realtime REYES craze was much more interesting.

As far as the PPU goes, it's all up in the air until we actually get something concrete. Even specs change between announcement and release and we don't even have those yet. I'll take a wild guess and say that I think it'll be faster than a PC alone would be for doing certain kinds of math, and that it'll probably be slower than a Cell at doing it.

Nite_Hawk

 

[Frank]

I would tend to agree, but then I've heard from others who are quite convinced that physics processing is somehow not ideal for Cell's team of SPE's, nor do they express any faith that any 1 of the 3 in-order PPC's in the XB2 could handle speedy physics.

Cell cannot access all of the memory at random, the 3-core PPC processor has much less FLOPS and has to synchronize everything all the time.

So the mind boggles at just what sort of hardware implementation a real PPU could take if the processing demands seemingly transcend that which could be accomplished in the impressive hardware configurations of these 2 consoles or even the coming multicore x86 incarnations in the PC scene.

Perhaps, this thing is simply a refreshened x87 math coprocessor, except now with multiple parallel pipelines (ala GPU style)? Maybe MIMD has some role in this, as well? ...just thinking out loud.

Most likely, it's a set of parallel pipelines that can all do 4 flops in SIMD, or 1 vertex calculation. With a small general purpose core for program management and flow control.

Like a 1+4 core Cell, but with cache memory instead of large register files.

 

[_phil_]

not sure if already posted somewhere.

http://personal.inet.fi/atk/kjh2348fs/ageia_physx.html

- Manufactured by TSMC (130nm process)
- 125 million transistors
- 182 mm² die size
- 20 watts power consumption
- 128MB GDDR3
- PCI only (PCIe cards expected further in the future)
- ASUS first board manufacturer
- May be integrated in graphics cards in the future
- Only 1 model at launch
- The cards bandwith exceeds that which can be put through a PCI bus
- The PCI and PCIe cards will be seperate
- Samples in Q3 2005
- Expected to become available in Q4 2005 (December)
- Price roughly $249 to $299

PCI ,that sucks
. :?

 

[Ragemare]

I think Sweeny is correct.

I don't see why a dedicated piece of hardware shouldn't be able to compete with a general purpose CPU that, as JVD said, is doing other things as well. While information on how exactly the chip works isn't publicly availabe, we can deduce from what Ageia and it's partners have said that it is reasonably powerfull and can compete with consoles. Although we have to make the assumption that Ageia and co. arn't downright lieing.

As to how it works, I think the most popular theory right now is that all the required physical data for a scene i.e all the data that the Novedex(sp?) lib requires to function, is first uploaded to the card. The chip then makes the appropiate calculations and when finnished sends the new positional data back to the CPU.

The chip itself will be set up to deal with the type of maths calculations that PC physics require and will be able to deal with high data transfere from one part of the pipeline to another. I gather that there are certain patterns to physics processing that a specialised piece of circuitry would be able to cater to. Which means, if done right, a PPU could be much faster than a general purpose CPU, even if that CPU has a high FP processing abuility.

 

[gurgi]

To me, the cool thing about a ppu is that it is dedicated physics hardware. In a PC, you still have the problem of attach rate (who actually will have one in thier PC?), which is why I was hoping one of the consoles would have a ppu. If devs wanted the most out of a machine, they'd use the ppu, and all of the ppu for physics. As a consumer, this would have put me in the driver's seat a bit, as I would jump to buy a console with dedicated physics hardware. In a way, I'd have some say towards game design, instead of some publisher that wants to shit out shovelware.

PS3 and Xbox360 have beastly cpu set-ups, but it is silly on one hand to say that ageia has a great physics api (which is what sweeny uses no?), and on the other hand to assume that a PPU from the same company is worthless (using sweeny's comments no less).

Anyway, it's kinda moot unless the PPU is adopted by PC gamers. Hopefully, console devs will really utilize resources for some great physics. This would help demand for a PPU in the PC game hardware market too.

 

[Frank]

Ok. What would a dedicated PPU do?

You have all your game objects, consisting of bones, surrounded by polygons and with anchor points. You send all those over to the PPU. You tell the PPU some other parameters of all the objects, like their weight, if they can break / deform, the vector they're moving at, if they are part of larger objects, if they behave like fluids (three dimensional mesh) and the amount of inner points to use, etc.

The PPU calculates the bounding volumes, does collision detection, calculates interactions and new vectors, and repeats that a few times. When ready, it calculates the new positions and returns the updated position data and all the updated (deformed, broken and new) objects and meshes as well. The CPU updates the objects and meshes, calculates the scene to display, skins what's needed and starts sending the data and shader commands in the right sequence to the GPU.

Right?

 

[London Geezer]

I'm just a bit worried at how much bandwidth this constant data-swapping will take. Not like a PC has a lot to spare afterall.

 

[PC-Engine]

I'm just a bit worried at how much bandwidth this constant data-swapping will take. Not like a PC has a lot to spare afterall.

Well the Ageia guys said bandwidth is not an issue because the card already has it's own memory and it doesn't need a lot of communication bandwidth with the CPU or GPU.