Does Cell Have Any Other Advantages Over XCPU Other Than FLOPS?

[Alpha_Spartan]

It's going to be more expensive to get the most out of these consoles than it was last gen. So much so that we'll see less exclusives among third parties than we have ever seen before. If you're a dev and you know you're going to put the game on multiple platforms, I doubt you're going to tailor the engine to one specific hardware philosophy and then have it run unacceptably on the others. Alot of devs will take the middle ground.

I can't help but think of the Novodex Physics SDK that apparently runs amazingly on Cell. Why is that? Perhaps the design of the SDK favors Cell's streaming-based architecture. If you're EA and you're concurrently developing a game across all three platforms, I doubt you're going to use Novodex since it runs better on only one hardware implementation. However, if you're Polyphony, that's right up your alley because you only have to worry about the engine running on the PS3.

The most technically impressive games next gen will definately be the exclusives. However, there won't be that many from third party developers.

 

[Shifty Geezer]

so only when used properly with extensive effort and budget will devs get the most out of cell?

The difficulty with which developers can get to grips with Cell is something of an unknown at the moment. Hopefully with the SDK out now we'll get some feedback. However it's a lot easier to get to grips with than EE was (in theory, I've no experience programming either!) in apparently every way (though Faf misses his VUs), plus a lot of the requirements for using Cell efficiently also seem to apply to XeCPU, such as managing data structures to keep memory accesses efficient. This is one of those areas I expect everyone to whinge how hard it is at first, yet once they've got to grips find it as easy to work with as writing C++, because people are great at learning new things.

 

[Titanio]

I define general purpose processing as not geared toward a particular type of processing (i.e. streaming) but well-suited for a wide array of processing styles. A jack-of-all-trades and a master of none.

The problem with this is, it seems to be a very relative thing. Some people like to say Xenon is "general purpose" because MS says so, and tries to use it as a good thing vs Cell, but Xenon does not look very general purpose next to a desktop processor..

And then the ultimate question is - do we need "general purpose"? Who says game code is "balanced"? We need optimisation - to run the most demanding tasks well rather than run anything you might ever want to run on it "ok". Look at what devs say they want to use more CPU power for next gen and Cell starts looking very good IMO.

For example, there will be A LOT of multiplatform games next generation. These games will not necessarily be fine-tuned to one hardware implementation over another. The engines will be hardware agnostic, taking the common denominator into account

Says who? Middleware providers offer specific implementations for each platform. If they didn't, they'd find themselves technically uncompetitive against those that do. We've already heard the likes of Crytek saying they're making "an engine within an engine" just for PS3 and Cell, we've heard the likes of Tim Sweeney who seems to be putting time into optimising for Cell, even if in fairly obvious ways - it'll undoubtedly benefit.

I can't help but think of the Novodex Physics SDK that apparently runs amazingly on Cell. Why is that? Perhaps the design of the SDK favors Cell's streaming-based architecture. If you're EA and you're concurrently developing a game across all three platforms, I doubt you're going to use Novodex since it runs better on only one hardware implementation.

I'll remind you that Novodex started life on PC CPUs..

Just because something runs brilliantly well on one particular chip doesn't mean it's not running as well as it can on others. I know you'd like to think that it can be purely attributed to the software just not being as good as it could be on a particular system, rather than the system itself, but ultimately the hardware is what these guys are up against, and it is in their business interests to getting it work as best as it can on each system. In many cases it'll be the domain rather than a specific implementation that might lend itself more to one system than another also (e.g. "physics", generally, as opposed to "novadex").

It's hard to tell at this stage, but I also think it's likely PS3 will be the lead platform for most next-gen projects. Certain key publishers seem to be already favouring it for 2006 titles (for example, one begins to wonder when EA starts talking about development of a title like next-gen MoH on PS3, while simultaneously putting out current-gen ports on X360 this year).

 

[flick556]

Having read about cell and 360's cpu, both are very good but cell has the potential for something special. The benchmarks IBM released may be hand picked but the performance is really outstanding. No other pc or console processor can achieve such results even if they did hand select and optimize the tests.

The one ppe alone should be a better processor than either the xbox or ps2 cpus, and this shows just how strong xbox360 is in relation to past generations.

I have to think that devs will do whatever it takes to gain access to all the performance cell has to offer. Because if you’re trying to make the best game possible you want the most power you can possibly get and these spes have tons. The difficulties of paralyzing code, using the spus instruction set, and its preferred memory access patterns appear challenging but not as bad as some of things we have had to deal with in previous architectures.

If middleware is strong for ps3 we may see the majority of devs access the spus potential, otherwise only some devs will be able to afford to put in the effort. As an aside I think its silly the way hundreds of software companies all write the same stuff, middleware needs to become more popular to avoid this type of redundancy. There are still some challenges though in making middleware and software in general easily interoperable. The fact that this power is even available is a good thing making it easy to access would be icing on the cake.

 

[PeterT]

Questionable post, as your hardly contributing to the discussion now are you?

Right, that's why I said that I was sorry for it, and that it's basically a rant.

xbdestroya, most of what you said is also correct. I did not wish to imply that everything not said by an authority is not valuable to the discussion. Nor even that you can't become an authority by dedicatedly persuing a personal interest in hardware for some time. Just that overly broad conclusions from people on console forums in general are seldom based on experience or even informed speculation.

Anyway, I'll now try to actually contribute something to the discussion

I define general purpose processing as not geared toward a particular type of processing (i.e. streaming) but well-suited for a wide array of processing styles. A jack-of-all-trades and a master of none

Nice to finally see a try at defining that. What I'd like to stress in the whole "general purpose" discussion is that for most of the intensive tasks you'd like to do on a CPU in a game there are numerous different algorithms, most of which will be best suited to some particular architecture. Of course it is true that at the moment, the non-parallel versions are very well understood in the game development community. They are in most cases also easier to implement and describe/verify formally than anything featuring concurrent execution. However, as nearly all high-performance gaming platforms are moving to parallelism now, I expect the experience part at least to shift gradually during the coming years.

 

[flick556]

It's going to be more expensive to get the most out of these consoles than it was last gen. So much so that we'll see less exclusives among third parties than we have ever seen before. If you're a dev and you know you're going to put the game on multiple platforms, I doubt you're going to tailor the engine to one specific hardware philosophy and then have it run unacceptably on the others. Alot of devs will take the middle ground.

I can't help but think of the Novodex Physics SDK that apparently runs amazingly on Cell. Why is that? Perhaps the design of the SDK favors Cell's streaming-based architecture. If you're EA and you're concurrently developing a game across all three platforms, I doubt you're going to use Novodex since it runs better on only one hardware implementation. However, if you're Polyphony, that's right up your alley because you only have to worry about the engine running on the PS3.

The most technically impressive games next gen will definately be the exclusives. However, there won't be that many from third party developers.

It may likely be the case that all physics SDKs run better on cell given that physics should be well suited to the way cell is designed. I think Novodex will be highly competitive on both platforms independent of the other. They will be one of the best physics solutions for xbox when compared to other solutions for xbox and the same for ps3. This should be the goal of all cross platform middleware developers.

 

[scooby_dooby]

Shifty & Phil - Yes I realize XeCPU is also untested, however it's a PowerPC based CPU so it's seen real world applications which is more that we can say for CELL.

On the subject of IBM, they are selling the CELL in other markets besides gaming, they are not selling XeCPU to anyone but MS. That's the difference in my mind.

How come IBM hasn't put out benchmarks showing the overwhelming power of the XeCPU? Oh ya, because they aren't trying to SELL it...

 

[ADEX]

I well aware, that all integer code has to be vectorized to run on the SPE's, but just because they have to be vectorized, does not mean it cannot run "general purpose" code. The conversion to vector code can occur during compile time, and it's only if an integer result is needed, that inefficiencies creep in.

SPEs can run scalar code as well as vector. You do not need to vectorise, however it's obviously beneficial if you can.

The peak figures for integer operations in the SPEs are the same as the Floating point ops.

The dual threading of the Xbox 360 cores is seperate register banks to allow fast context switches. That's great, but I can't imagine a 50 percent increase. Can you provide any evidence of this?

They use SMT (Simultaneous Multi Threading) so they can both be in operation at the same time - they don't context switch. They can issue 1 instruction each or one thread can issue 2 instructions. If one thread stalls it gets full use of the instruction units - that's a 100% performance boost since otherwise the CPU would be sitting idle.

People act like XeCPU has no FLOP capabilities, in fact it's very high for a conventional CPU, it's just not as high as the theoretical peaks of the yet untested CELL.

The theoretical peaks of one are not as high as the theoretical peaks of the other. Yes.
Actual performance will depend on how well the developer can make use of the facilities present in each processor. The architectural differences will come into play though, especially the Cell's local stores.

The Cell isn't untested as it isn't a completely new architecture, similar processors have existed before, it bears more than a passing resemblance to the Cray 2 supercomputer.

Since this is a pissing contest (and nothing else) why stop at 8 threads ? Why not make 64 two-register contexts/threads per SPE ? Then the SPEs could provide 448 threads!!!!

That might actually work they'll have less than 4K per thread.

Oh and by the way, the same mechanism can be used on both the PPE and the XCPU, and probably will be (vertical threading anyone?)

They don't have local stores so can't use the same mechanism. You can (and often do) run hundreds of threads on any CPU but it'll be constantly hitting memory.

I can't help but think of the Novodex Physics SDK that apparently runs amazingly on Cell. Why is that?

Again, local stores. Very useful for collision detection, other processors just hit memory.
The only other chip which wont hit memory is the PhysX which happens to be closer to Cell than anything else (at least according to the patent).

There are types of algorithms which SPEs will completely suck on, XCPU will do better on those. The trick is to know what to use on what processor.

As for "general purpose", that's just FUD, they can all run anything, but they'll be better or worse in specific areas.

 

[Alpha_Spartan]

Says who? Middleware providers offer specific implementations for each platform. If they didn't, they'd find themselves technically uncompetitive against those that do. We've already heard the likes of Crytek saying they're making "an engine within an engine" just for PS3 and Cell, we've heard the likes of Tim Sweeney who seems to be putting time into optimising for Cell, even if in fairly obvious ways - it'll undoubtedly benefit.

This is new to me. So what you're saying is that developers like EA have four different versions of Renderware (Renderware for PS3, Renderware for Xbox 360, Renderware for Rev, Renderware for PC)? I always understood SDK's like Renderware to be rather hardware agnostic in their approach. I understood Renderware to be very portable across platforms. I thought Renderware had a lowest common denominator approach and didn't favor one hardware philosophy over another. I could be wrong though.

Just because something runs brilliantly well on one particular chip doesn't mean it's not running as well as it can on others. I know you'd like to think that it can be purely attributed to the software just not being as good as it could be on a particular system, rather than the system itself, but ultimately the hardware is what these guys are up against, and it is in their business interests to getting it work as best as it can on each system. In many cases it'll be the domain rather than a specific implementation that might lend itself more to one system than another also (e.g. "physics", generally, as opposed to "novadex").

You missed my point. I'm saying that the Novodex SDK favors one type of HARDWARE PHILOSOPHY over another. There's no doubt that Novodex runs better on Ageia's PPU. I may be mistaken, but the Cell shares some similarities in hardware philosophy with Ageia's PPU (i.e. streaming-based). I wasn't suggesting that Novodex wasn't running as best as it can on other hardware, but rather that it runs better on the hardware philosophy it was designed around. Shit, you can get any code to run on any processor with a good compiler. But it will ALWAYS run better on the hardware it was designed for. I won't buy into the generalization that Cell can do better PHYSICS than the XCPU. I'd wager that a physics engine designed around the XCPU philosophy would run better on the XCPU than the Cell. It's not like we're comparing two processors that are a couple of generations apart. For illustrative purposes only, a 3.2 GHz single-threaded processor can run multi-threaded code faster than a 50 MHz multi-threaded processor. So no matter what processor you design the code around, a processor that's magnitudes faster will always run it faster. When we're comparing Cell and XCPU we're not talking about one processor that's magnitudes of order faster than the other.

It's hard to tell at this stage, but I also think it's likely PS3 will be the lead platform for most next-gen projects. Certain key publishers seem to be already favouring it for 2006 titles (for example, one begins to wonder when EA starts talking about development of a title like next-gen MoH on PS3, while simultaneously putting out current-gen ports on X360 this year).

What about Madden, Live, FIFA, etc. Those aren't current gen ports. So things seem to be the opposite of what you make them out to be.

 

[Alpha_Spartan]

It may likely be the case that all physics SDKs run better on cell given that physics should be well suited to the way cell is designed. I think Novodex will be highly competitive on both platforms independent of the other. They will be one of the best physics solutions for xbox when compared to other solutions for xbox and the same for ps3. This should be the goal of all cross platform middleware developers.

There's a very good chance that you're right. Form follows function I suppose, when it comes to physics processing. Streaming seems to be the best implementation for physics processing.

 

[Titanio]

This is new to me. So what you're saying is that developers like EA have four different versions of Renderware (Renderware for PS3, Renderware for Xbox 360, Renderware for Rev, Renderware for PC)? I always understood SDK's like Renderware to be rather hardware agnostic in their approach. I understood Renderware to be very portable across platforms. I thought Renderware had a lowest common denominator approach and didn't favor one hardware philosophy over another. I could be wrong though.

I don't know what Renderware does. But then EA uses it, and not many others anymore. Obviously there must be hardware-specific implementations, however, otherwise the code just won't work. I'm sure you can within the confines of Renderware also get down and talk to the hardware more one-on-one too, and I'm sure it's done.

What about Madden, Live, FIFA, etc. Those aren't current gen ports. So things seem to be the opposite of what you make them out to be.

They're not exactly exemplars of next-generation development. They're basically the same games as on the other systems, but with better graphics/animations, minus some features in some cases. Next year's versions will get a lot more attention - wonder why?

EA are not the only ones. Activision is just throwing X360 ports currently. Same with Ubisoft. It's clear that for a lot of key publishers the real transition to next-gen development has yet to happen. But then you hear about EA beavering away on next-gen MoH in the background for PS3. It's just a little suggestive (and the rumours go that EA likes PS3 a lot, but that's just rumour of course). It's just my guess that PS3 likely will be the lead platform for many publishers, it's not something we can know yet like I said above.

 

[c0_re]

Both CPU's have so much housepower this gen neither will be lacking, not even close. IMO Cell is overkill that will just mkae game development even more difficult and expensive than it already is.

 

[drpepper]

Both CPU's have so much housepower this gen neither will be lacking, not even close. IMO Cell is overkill that will just mkae game development even more difficult and expensive than it already is.

That's never stop the developers before.

 

[Shifty Geezer]

Both CPU's have so much housepower this gen neither will be lacking, not even close. IMO Cell is overkill that will just mkae game development even more difficult and expensive than it already is.

You can't have overkill. Whenever there's power, it's used and devs want more, often using it in ways no-one expected it'd be used. You can't have too much processing power, RAM, shading capacity, bandwidth, or anything else, except where the system has too much of one and not enough of the others in an imbalanced system. Most PC's sold today should be overkill given there uses 10 years ago. You don't need 3 GHz processors to wordprocess and email, but as the technological performance has increased users have taken to doing new things like audio decoding, video editing, digital photoretouching, and 3D gaming, all of which run slower than instantaneously and leave the users wishing it'd go faster and wanting more power.

 

[expletive]

You can't have overkill. Whenever there's power, it's used and devs want more, often using it in ways no-one expected it'd be used. You can't have too much processing power, RAM, shading capacity, bandwidth, or anything else, except where the system has too much of one and not enough of the others in an imbalanced system. Most PC's sold today should be overkill given there uses 10 years ago. You don't need 3 GHz processors to wordprocess and email, but as the technological performance has increased users have taken to doing new things like audio decoding, video editing, digital photoretouching, and 3D gaming, all of which run slower than instantaneously and leave the users wishing it'd go faster and wanting more power.

I think the argument may be that both of these systems will eventually be limited by their (albeit very powerful) GPUs. Being able to produce tons of FLOPS, on either CPU, eventually leads to the laws of diminishing returns in terms of software 'output' at some point. And by output i'm using the primary feature we all seem to be focused on, graphics.

I know the counterpoint is physics, AI, etc. When I play the same game side by side on these two systems and any of those things actually noticeably changes the game experience, then i'll be happy to factor them into my thinking.

 

[expletive]

That's never stop the developers before.

How can you be so sure? I'll bet that there have been numerous instances this past gen where coding difficulty left a feature or enhancement 'on the cutting room floor'. This would be particularly true, and will be more so this gen, when there is a dollar amount attached to that difficulty.

 

[one]

This is new to me. So what you're saying is that developers like EA have four different versions of Renderware (Renderware for PS3, Renderware for Xbox 360, Renderware for Rev, Renderware for PC)? I always understood SDK's like Renderware to be rather hardware agnostic in their approach. I understood Renderware to be very portable across platforms.

Only API and high-level code are portable. Middlewares have platform-specific low-level modules in lower parts of the stack. It's the purpose of middleware to offload platform-specific optimization onto middleware developers instead of game developers.
http://www.intrinsic.com/vvisionsTechnology.php?s=products

 

[flick556]

I think the argument may be that both of these systems will eventually be limited by their (albeit very powerful) GPUs. Being able to produce tons of FLOPS, on either CPU, eventually leads to the laws of diminishing returns in terms of software 'output' at some point. And by output i'm using the primary feature we all seem to be focused on, graphics.

I know the counterpoint is physics, AI, etc. When I play the same game side by side on these two systems and any of those things actually noticeably changes the game experience, then i'll be happy to factor them into my thinking.

I don't think the same game side by side can have different physics or AI, not because of hardware or software but because of game design. These things need to be integrated into the gameplay. So the only comparisons we will likely be able to make is to compare different games on each system. If the difference is large enough between the systems we will notice.

 

[Titanio]

I think the argument may be that both of these systems will eventually be limited by their (albeit very powerful) GPUs. Being able to produce tons of FLOPS, on either CPU, eventually leads to the laws of diminishing returns in terms of software 'output' at some point. And by output i'm using the primary feature we all seem to be focused on, graphics.

I know the counterpoint is physics, AI, etc. When I play the same game side by side on these two systems and any of those things actually noticeably changes the game experience, then i'll be happy to factor them into my thinking.

The things you mention - physics etc. - are not limited by the GPU. The GPU only really places limits on the amount of rendering data the CPU can provide - the amount of data it can take in - but not on the complexity or sophistication of the operations performed on that data by the CPU. And things like physics etc. certainly can affect how a game looks, if not how it also plays.

 

[expletive]

The things you mention - physics etc. - are not limited by the GPU. The GPU only really places limits on the amount of rendering data the CPU can provide - the amount of data it can take in - but not on the complexity or sophistication of the operations performed on that data by the CPU. And things like physics etc. certainly can affect how a game looks, if not how it also plays.

Sorry i wasnt clear, the reason i brought up physics, AI was to point out examples of things that would NOT be bound by the GPU and instances where the 'more always=better' for CPU power might be valid. For me though, i odnt really believe we'll see large enough differences in those things from system to system (in the upcomign console gen) for a mjaority of the people to notice or care.