arstechnica part 2 : inside the xbox 360
- Thread starter jvd
- Start date
The claim that physics, AI, or game logic can't be multithreaded are debatable. Novodex is already multithreaded for example. And whether game logic or AI can be multithreaded is highly dependent on actual algorithm used.
There is also a difference between pure concurrency, and partial concurrency. There are many algorithms that fit inbetween, where large pieces can run truly concurrent, but some shared data must be atomically updated and thus synchronization is forced. But where and how this synchronization of a shared database (geometry, etc) happens is also highly dependent on implementation, datastructure, and other requirements. For example, sometimes all you need is atomicity, not consistency (different threads get different values when they read the same sequence of variables), depending on how tolerant of errors you are.
So I think the view that something is either inherently serial or inherently parallel is too simplistic.
There is also a difference between pure concurrency, and partial concurrency. There are many algorithms that fit inbetween, where large pieces can run truly concurrent, but some shared data must be atomically updated and thus synchronization is forced. But where and how this synchronization of a shared database (geometry, etc) happens is also highly dependent on implementation, datastructure, and other requirements. For example, sometimes all you need is atomicity, not consistency (different threads get different values when they read the same sequence of variables), depending on how tolerant of errors you are.
So I think the view that something is either inherently serial or inherently parallel is too simplistic.
Well, I felt that it deals as much of the in order aspect as any and I am not much of a thread starter. But no worries, seems both sides are going to have their hands full with thread scheduling and on the compiling front, as i mentioned on another board it's shades of Itanium of a sort. Lets hope they work out better than that monster.
No one vindicated, because people keep making unsubtantiated claims that physics is inherently serial.
If physics is inherently serial, than the PPU is DEAD, because the whole point of a PPU is slapping a huge number of SIMD units on die and using parallelism either ILP or TLP. But if you believe ARS, then Physics can't be parallelized, and therefore, the PPU can't possibly work.
But physics isn't inherently serial, in fact, most of it is inherently parallelizable. That's why the government keeps buying supercomputers like ASCI White and Purple to run weapons simulations. Or why Japan built the Earth Simulator to simulate weather. Or why Boeing and NASA use supercomputers to do computational fluid dynamics.
If physics is inherently serial, than the PPU is DEAD, because the whole point of a PPU is slapping a huge number of SIMD units on die and using parallelism either ILP or TLP. But if you believe ARS, then Physics can't be parallelized, and therefore, the PPU can't possibly work.
But physics isn't inherently serial, in fact, most of it is inherently parallelizable. That's why the government keeps buying supercomputers like ASCI White and Purple to run weapons simulations. Or why Japan built the Earth Simulator to simulate weather. Or why Boeing and NASA use supercomputers to do computational fluid dynamics.
DemoCoder said:
This makes me wonder if you actually looked at the Novodex SDK?
Novodex Physics SDK Documentation said:
As far as I can tell, all that the "multithreading" support in Novodex does is pull the physics code out of the application thread, and into its own (single) thread.
Incidentally, this setup is ideal for the way xenon works.
Hannibal seems to believe the XeCPU PPC cores and PPE in CELL are pretty similar less some minor tweaks. It is interesting that this rumor from late 2004 seems to be true. Even more interesting because this should help 3rd party cross platform titles.
I wonder if this means we will see, at least early one, some games that ignore the "special sauces" in both designs and just use a single core, and only get ambitious down the road?
I wonder if this means we will see, at least early one, some games that ignore the "special sauces" in both designs and just use a single core, and only get ambitious down the road?
DemoCoder, are all physics algorithms equally paralellizable?
The examples you give of paralellized physics (explosion and fluid modeling) are grid-based simulations. They aren't very useful for games, because even with Cell the largest real-time simulation will be too small/coarse to make an interesting game world.
Isn't game physics (polygon soup collision detection and reaction) more of a data base search and update problem? While there's some spatial coherency you can take advantage of, isn't most of the time in these algorithms spent searching data structures?
(I have Eberley's Game Physics book on order, but haven't read it yet, so maybe I'm off base here...)
The examples you give of paralellized physics (explosion and fluid modeling) are grid-based simulations. They aren't very useful for games, because even with Cell the largest real-time simulation will be too small/coarse to make an interesting game world.
Isn't game physics (polygon soup collision detection and reaction) more of a data base search and update problem? While there's some spatial coherency you can take advantage of, isn't most of the time in these algorithms spent searching data structures?
(I have Eberley's Game Physics book on order, but haven't read it yet, so maybe I'm off base here...)
Oddly, I find myself agreeing with the posters here more than with Hannibal. I had expected to agree more with Hannibal since he's supposed to be the expert here.
He does have a point that the X360 won't outrun the Xbox in every field equally. But he seems to dwell on that point. It's almost as if he thinks IBM and MS are unaware of the kind of code devs will be running. IMO, we're entering the era of unconsidered code running bad across the board, so none of this sounds out of place.
I think he rather underplays the significance of a compiler with profiling for a fixed system. That combined with Microsoft's ATG should be enough to make decent performance attainable.
Did it bug anyone else that he more or less assumed the X360's CPU cores were identical to the CELL's PPE? I think there's a lot of similarity there, but I don't think it's accurate to take a bunch of CELL charts and stats and say, "This is how the X360 is."
He does have a point that the X360 won't outrun the Xbox in every field equally. But he seems to dwell on that point. It's almost as if he thinks IBM and MS are unaware of the kind of code devs will be running. IMO, we're entering the era of unconsidered code running bad across the board, so none of this sounds out of place.
I think he rather underplays the significance of a compiler with profiling for a fixed system. That combined with Microsoft's ATG should be enough to make decent performance attainable.
Did it bug anyone else that he more or less assumed the X360's CPU cores were identical to the CELL's PPE? I think there's a lot of similarity there, but I don't think it's accurate to take a bunch of CELL charts and stats and say, "This is how the X360 is."
I think they used at least one SPE, what appeared to be volumetric explosion effects were clearly seen in the beginning. You'd hope they'd figured out some thread parellism there to not bog the main core down, even if it was a non AI, non gameplay prescripted realtime render.
aaaaa00 said:Acert93 said:
Seems to be what Epic did to get UE3 running on the CELL so quickly.
Yeah, reading his quote here seems to hint at that direction:
http://www.beyond3d.com/forum/viewtopic.php?t=23585
Obviously it is more important for 3rd parties to get quality games on as many platforms as possible QUICKLY and not have to mess around with a lot of porting and rewriting of engines (unless they got big money for doing such).
Both with PPC cores and GPUs from PC IHVs that are using common APIs/derivatives in DX and OpenGL, I think this will be a very nice environment for porting. Obviously it would be nice for games to exploit the features on both platforms, but it seems with the cut throat business it can be hard finding the time, money, and resources to do that. I wonder if both companies using similar cores was intentional? I wonder if Nintendo will be using a similar CPU?
aaaaa00 said:
If that's what they can get with just 2 hardware threads, what could they do with the other 7 fairly humming along with some measure of utility? Not to mention that the SPE's I understand are best optimized under microthreading. Being ignorant how does that differ from normal threaded operation?
