Xtrema
06-16-2006, 09:33 AM
Very good article.
http://arstechnica.com/articles/culture/mattlee.ars/
http://media.arstechnica.com/articles/paedia/hardware/crossplatform.media/ps3-arch.gif
What do you think of the PS3's Cell architecture? How much harder is it for developers to write asymmetric multiprocessing code that will run on the PPE and the SPE's? How difficult will it be to port code that is tuned for the Cell over to the 360 while maintaining good performance? Do you think Sony made the architecture complex deliberately to make porting that way harder, or were they just trying to get maximum performance out of their system?
I don't think the Cell is as well designed for game development as Sony would have you believe. Some aspects of the SPEs, such as the lack of branch prediction, make them particularly unsuited to running most game code, which contains a lot of branches. They appear to be designed more for serialized streaming math code, more common in video codecs and audio processing, the traditional domain of digital signal processing chips. The memory architecture of the SPEs, specifically their lack of automatic cache coherency in favor of DMA transactions, seems like a lot of overhead is needed to feed work units to the SPEs and copy the results back to system memory.
The PPE appears to be essentially identical to one of the Xbox 360 cores, except without the VMX128 enhancements and with half the cache. However, a much greater assortment of work has to be crammed into this single core—all of the game loop, all of the rendering commands, and the system allegedly takes over some time as well. Only the second and third cores on Xbox 360 use a small timeslice to provide cool stuff like the Guide, music playback, Dolby Digital encoding, and more things that we can add in the future to all games, past and present.
I think porting from Xbox 360 to PS3 will be reasonably difficult, since the Xbox 360 has a lot more general purpose processing power that can be flexibly reallocated, and all of the Xbox 360 CPU cores have equal access to all memory. The asymmetric nature of the Cell could easily lead to situations where the game has too little of one type of processing power and too much of another. And the content might suffer as well, since you'll never see a PS3 title with more than 256MB of textures at any given time, due to the split graphics and system memory banks. When we announced 512MB of unified memory on Xbox 360, I think all of our game developers (and the artists too) did a little happy dance. It's easier to use and gives developers much more flexibility in how they allocate memory for various resources.
In terms of performance, I think that the PS3 and the Xbox 360 will essentially be a wash. We ran the numbers a while back and the two systems come up surprisingly close in theoretical peak performance, despite the one year difference in release dates. However, I know for a fact that we have a great advantage in software and services—our development environment and tools are years ahead of the competition, and this will ensure that Xbox 360 game developers can easily realize all of this performance and make superior games. Xbox 360 is a great system to develop on, a real pleasure—and I believe our developers agree.
http://arstechnica.com/articles/culture/mattlee.ars/
http://media.arstechnica.com/articles/paedia/hardware/crossplatform.media/ps3-arch.gif
What do you think of the PS3's Cell architecture? How much harder is it for developers to write asymmetric multiprocessing code that will run on the PPE and the SPE's? How difficult will it be to port code that is tuned for the Cell over to the 360 while maintaining good performance? Do you think Sony made the architecture complex deliberately to make porting that way harder, or were they just trying to get maximum performance out of their system?
I don't think the Cell is as well designed for game development as Sony would have you believe. Some aspects of the SPEs, such as the lack of branch prediction, make them particularly unsuited to running most game code, which contains a lot of branches. They appear to be designed more for serialized streaming math code, more common in video codecs and audio processing, the traditional domain of digital signal processing chips. The memory architecture of the SPEs, specifically their lack of automatic cache coherency in favor of DMA transactions, seems like a lot of overhead is needed to feed work units to the SPEs and copy the results back to system memory.
The PPE appears to be essentially identical to one of the Xbox 360 cores, except without the VMX128 enhancements and with half the cache. However, a much greater assortment of work has to be crammed into this single core—all of the game loop, all of the rendering commands, and the system allegedly takes over some time as well. Only the second and third cores on Xbox 360 use a small timeslice to provide cool stuff like the Guide, music playback, Dolby Digital encoding, and more things that we can add in the future to all games, past and present.
I think porting from Xbox 360 to PS3 will be reasonably difficult, since the Xbox 360 has a lot more general purpose processing power that can be flexibly reallocated, and all of the Xbox 360 CPU cores have equal access to all memory. The asymmetric nature of the Cell could easily lead to situations where the game has too little of one type of processing power and too much of another. And the content might suffer as well, since you'll never see a PS3 title with more than 256MB of textures at any given time, due to the split graphics and system memory banks. When we announced 512MB of unified memory on Xbox 360, I think all of our game developers (and the artists too) did a little happy dance. It's easier to use and gives developers much more flexibility in how they allocate memory for various resources.
In terms of performance, I think that the PS3 and the Xbox 360 will essentially be a wash. We ran the numbers a while back and the two systems come up surprisingly close in theoretical peak performance, despite the one year difference in release dates. However, I know for a fact that we have a great advantage in software and services—our development environment and tools are years ahead of the competition, and this will ensure that Xbox 360 game developers can easily realize all of this performance and make superior games. Xbox 360 is a great system to develop on, a real pleasure—and I believe our developers agree.