[Falcon]

http://www.gamespot.com/news/2005/03/08/news_6119895.html

http://www.gamers-depot.com/interviews/agiea/001.htm

[IceFire]

Interesting...it will need to gain some support if its to work.

I mean we already dole out big sums of cash on:
CPU
VPU/GPU
APU
Now they want to add a physics board?

[Carl]

woot. with this we may finally be able to make realistic liquid.

[Nuke]

what id like to see is a neural net ai unit, now that would be cool, solve the ai problems that has always plagued the game industry. it could be far more dynamic than anything you could program.

[WMCoolmon]

Originally posted by IceFire
Interesting...it will need to gain some support if its to work.

I mean we already dole out big sums of cash on:
CPU
VPU/GPU
APU
Now they want to add a physics board?

Hey, maybe I can run SETI on it, too.

[Galemp]

now THIS is interesting. Initial thoughts:

A.) Can it be applied to non-gaming applications? For instance, can we see much faster and more detailed dynamics and particle systems in 3D animation suites?

B.) Are they drawing on any existing codebases for their software, like Red Faction II or HalfLife 2?

[Fineus]

I'm with Galemp on this one - it's a good idea but it seems to be targeted mainly at gamers (given that they're the ones who usually have to worry about physics in game).

It also rather paves the way for the whole nine yards.. a separate board for lighting and shadow calculation... and goodness knows what else.

[WMCoolmon]

Sound?

[aldo_14]

Originally posted by Nuke
what id like to see is a neural net ai unit, now that would be cool, solve the ai problems that has always plagued the game industry. it could be far more dynamic than anything you could program.

No chance.  It's inherently unpredictable in a bad way, plus it'd be a nightmare to design the scenarios to train the network*.  

This physX thing... I reckon they must have a next gen console contract for it to be financially feasible.  The sony Cell chip is really designed for parallel agent-style processing I think, so this sort of addon wouldn't strike me as the ideal thing for the PS3.  I'd guess the Nintendo Revolution(?) or the XBox2, possibly the latter if they want to make a quick desktop Pc transition (because XB2 games are IMO more likely to be ported to Pc than Nintendo ones).


*I still remember the story of the tank-detecting NN AI; they trained it on a set of photos to find if tanks were present, and also a set of tankless pictures.  It passed the initial tests, but then failed on a set of following pictures - turned out it wasn't detecting the presence of tanks, but the shadows in the training set pictures of tanks.  That's the problem with NN; you have no idea what or why they tune to, it's a black box - and hence unvalidatable - system.  Bad, bad stuff for games...

[Turnsky]

Originally posted by aldo_14

XBox2

it's called the Xbox 360 now i think... makes you wonder what happened to the other 358.. dunnit?

anywho.. all i see out of this is another piece of overpriced hardware, that not many games will support, and will take up space on your mobo, that'd be better used for other things.. like an nVidia SLI or some ****e like that.

the technology would be better used if intergrated into modern videocards..

[phatosealpha]

It seems extraordinairly unlikely we'd see anything like this in the PS3.  The cell processor is massively parralel with multiple DSPs that are essentially the same principle as this thing.  We already know that they're not doing graphics on the cell itself, and games require in order processing for most AI and game logic.  Physics is one of the few places games could genuinely benefit from the cell's parallelism, and I can't imagine they'd toss in an add on card to do it.

To be honest though, most of what I've heard of the cell paints it more as being designed for accelerating ultra-high resolution digital video then it is for gaming or typical computing, so I'm at a bit of a loss as to what Sony is planning for the PS3 anyway.  And if it's not super detailed physics....well, then I can only think they're planning on 'borrowing' PS3 time to sell to customers for large scale computing projects.

That's assuming, of course, that what the cell really is remotely resembles the Sony PR.  Wouldn't be the first time they didn't follow through.  But I suppose that's true of all companies.


I'm pretty sure it will work just fine in 3d rendering apps where you're trying to run physics, if the software is patched/upgraded to be able to use the API.


Adding it onto video cards seems like a likely use, though standalones aren't out of the question if the industry demands it.  I suppose it's also possible it will be licensed and added onto future CPU dies by AMD and Intel, kind of a counter to the Cell.



And then *glances at Vampire:Bloodlines* someone will make a really kick butt game with it, but use it's power solely to create realistically jiggling breasts.

[Clave]

And that would be bad?

[aldo_14]

Originally posted by phatosealpha
It seems extraordinairly unlikely we'd see anything like this in the PS3.  The cell processor is massively parralel with multiple DSPs that are essentially the same principle as this thing.  We already know that they're not doing graphics on the cell itself, and games require in order processing for most AI and game logic.  Physics is one of the few places games could genuinely benefit from the cell's parallelism, and I can't imagine they'd toss in an add on card to do it.

To be honest though, most of what I've heard of the cell paints it more as being designed for accelerating ultra-high resolution digital video then it is for gaming or typical computing, so I'm at a bit of a loss as to what Sony is planning for the PS3 anyway.  And if it's not super detailed physics....well, then I can only think they're planning on 'borrowing' PS3 time to sell to customers for large scale computing projects.

That's assuming, of course, that what the cell really is remotely resembles the Sony PR.  Wouldn't be the first time they didn't follow through.  But I suppose that's true of all companies.


I'm pretty sure it will work just fine in 3d rendering apps where you're trying to run physics, if the software is patched/upgraded to be able to use the API.

 

Cell strikes me as almost a hardware implementation of an agent system, whereas this strikes me as being a single-unit adjunct to existing systems (working in parallel to other CPU/GPU etc hardware but not in parallel with other physEx hardware).

I think I agree with you on the idea of the Cell being licensable for distributed supercomputing....it make sense IMO.

 IIRC sony talked of Cell programs being able to 'seek' free remotely located processors and run upon them - but I have no idea how they would build in a sensible global control for that sort of thing, which is pretty essential if you want to load-balance.    Sony do seem to have a thing for parallelisation, it seems; I believe the PS2s architecture has multiple pipelines and PUs, which made development a pain in the tits (for load balancing and avoiding deadlocks waiting on other tasks).

There's lots of examples of supercomputers built from thousands of multiple 'normal' Pc chips; if Sony can achieve the same thing with distributed PS3 cells (including latency problems?), with even a fraction of the consoles sold, then they could build a truly massive supercomputer.... I'm not sure if anyone has ever made a hundred-thousand CPU supercomputer.

Oh, and interesting Cell articles
http://www.theregister.co.uk/2005/02/01/cell_analysis_part_one/
http://www.theregister.co.uk/2005/02/03/cell_analysis_part_two/

[pyro-manic]

Interesting. Though it's probably going to be horrendously expensive...

[Gloriano]

Very interesting indeed, but i do think it will fail because if price is too high. Not lot peoples aren't going interested in it and they need lot support

[aldo_14]

AFAIK the price hasn't been set, for the PS3 anyways.  I'd imagine they'll set the price at a level sufficient to undercut the opposition and get market share.

[Bobboau]

I've been expecting something like this for a while, sence the introduction of the GPU, it seemed only natural to export more and more processing tasks to dedicated hardware, eventualy the cpu will only be used for extreemly high level flow controle, ai, vector math, probly a seperate geometry processor (outside the GPU or PPU, something capable of boolean subtraction for instance) will probly become standard in the not too distant future, eventualy they may all get re-intergrated back together into a game card that has all of the elements of the game handled on a single pece of hardware with seperate processors dedicated to diferent aspects of the game, and all game data will be held on memory on the card, someday you may even be able to by additional memory modules for your graphics/game card
.

[aldo_14]

I'd think that the games programmers would be somewhat concerned with that sort of change, though; they'd either have to resort to programming for many more hardware combinations, or use a performance sapping abstraction layer.  I'm not surprised by the movement - we've aready seen it for sound and graphics, which are themselves pretty much game-exclusive - but I can't help but wonder if this is all that much of step forward; surely there's going to be issues with your bus bandwidth and management if you keep shifting processing onto dedicated cards?

[Bobboau]

but you get paralel processing, while the PPU is multiplying a half dozen matricies the game can already be sending the  polygon data off to the graphics card, the PPU can send it's result to the graphics card when it gets done, and by this time the program has sent another matrix stack to the PPU.

abstraction layers don't sap performence unless built stupidly, FS's abstraction layer is just as fast as calling the API specific functions directly. an interface based abstraction is slightly slower, but the cost is negligable.

I do see muti-threading and concurency issues comeing to the fore front soon though.

[aldo_14]

Originally posted by Bobboau
but you get paralel processing, while the PPU is multiplying a half dozen matricies the game can already be sending the  polygon data off to the graphics card, the PPU can send it's result to the graphics card when it gets done, and by this time the program has sent another matrix stack to the PPU.

That depends on the amount of data being sent; it's unlikely at the moment to be enough to cause a problem (with the exception of gfx cards; the XBox and GC both have the GPU rather than CPU responsible for the data bus because of that IIRC), but I can't help but wonder how easy it would be to maintain control if you keep parallelizing that.

And also the concurrency issue; what if the physics code is modifying the exact same co-ordinate or geometry values being manipulated by the GPU or CPU?  How easy is it to maintain global coherence over shared data (and does that actually matter in real terms or not?)

Originally posted by Bobboau
abstraction layers don't sap performence unless built stupidly, FS's abstraction layer is just as fast as calling the API specific functions directly.

Ah, but i'm talking about high-level hardware abstraction here; specifically, that would hide the hardware access details in order to allow cheap (in cost) programming - and also to ensure you don't need to have a dedicated physics processor when running it, for example.  

There might also be a side issue of contention between the parallel processing - firstly if you have multiple pipelines you need to balance, and also if you have data consistency within them (also maybe if you need to share data between them).....basically if you need to synch complex processing and data management tasks between multiple parallel processes/threads, you'd either have to implement code for handling the hardware combinations, or consider an abstraction layer.  

And as you move from specific code->abstraction you lose performance control, because more of your hardware level optimization is associated with the abstraction code.

The main thing, though, is the increase of hardware combinations - how will developers handle it, and what will the cost in terms of dev time?  That'll be key to adoption on PCs, IMO.