Starbreeze take on the Ps3 vs Xbox 360 (the Darkness Int)

Facts? Most of that is inaccurate and at the very least not put into proper context.

OT. It's hard to believe the amount bait on these boards anymore, and yet I still sympathize with how difficult it is not to bite.

 

I actually think the two systems are comparable, and even think the 360 is an awesome system, but just pointing out the main difference in power between the two is CELL. When someone cuts down CELL as hype, that when a few facts should be shared. Feel free to debunk them. My choice in PS3 does not rest on CELL, but the games.

Rockster feel free to point out the lack of proper context. I can see the internal bandwidth lacking proper context, but Xenon does not have an internal bus or does it? It's main method in passing data from processor to processor is L2 cache. I could even be wrong about the Xenon L2 cache bandwidth? Is it running at 3.2 GHz or half that speed?

 

Not interested on going around and around on this. Perhaps it was my misunderstanding, but to me, your interpretation of the data was that Cell was 2-3x Xenon. It's my opinion that comparisons like this are misleading because of the inherent architectural differences between the two processors and are simply a product of the design goals for each.

- total amount of internal memory (2304 KB versus 360's 1024 KB: 2.25 times)
I understand how this was derived but the purpose of cache/local store is to keep the execution units fed. Why only count L2 cache? Isn't L1 and external memory just as important to performance. Consider doubling the cache on a A64 from 512kb to 1024kb, what's the percentage performance increase. How is it possible to extrapolate the performance impact between different architectures.

- number of processors (8 versus 360's 3: 2.7 times )
Pretty obvious, but apples and oranges. 1 apple + 7 oranges vs. 3 slightly different apples. Apples run certain types of code better than oranges, while other bits of code might prefer oranges.

- number of *concurrent* threads (8 versus 360's 3: 2.66 times)
Wrong. I think it's 9 vs. 6. And again, what does this imply?

- GFLOPS rating (204 GFLOPS versus 360's 77 GFLOPS: 2.65 times)
Here's one everyone loves, but doesn't in and of itself mean much. Gives no indication of the types of operations, their relative execution rate, conditionals, penalties, etc. I'll say this, each processor was designed with a different workload mix in mind, and as a result the performance advantage is completely dependant on the application. That's the most important thing to understand I think.

- integer rating (~51.2 billion instructions per second versus 19.2 BIPS: 2.66 times)
- internal bandwidth (204.8 GB/s versus shared 102 GB/sec L2 cache bandwidth: ~2 times)
- external bandwidth (60 GB/sec versus 360's 22 GB/sec: ~3 times)

I don't want to get too redundant. At the end of the day we can pick whatever numbers we want to see. We could include the bandwidth to both L1 (icache + dcache) + L2 cache since Xenon cores have direct access to both if we chose. Or could calculate Cell internal bandwidth by summing SPE bandwidth to local stores + PPE cache bandwidth. And BTW, L2 does only run at half-clock on the frontside, core access to the crossbar on the backside is full speed. Good Xenon read if you're interested. We could say we're likely to be bandwidth bound to main memory anyway and focus on that figure. And BTW again, I don't believe Cell can address the video memory (please feel free to correct me) so that 60GB figure would be inaccurate for external memory bandwidth, or maybe you just wanted to list all the interface speeds. Fine.

DGMW, I think the Cell processor is a neat design that will be extremely fast at processing certain workloads, many of which will be useful for games. But using terms like 2-3x superiority, without specifying the workload and test conditions, for something for which there is no basis for comparison, is just a bit much.

 

Real world performance is a relative term... every developer has its own meaning for it and is largely dependant on the time/effort/skill of the developer... Specs on paper shows the potential, real world performance the results of that developer, though bear in mind the extracted performance of that developer doesn't necessarely have to be in line with what others achieve and therefore can not be used as any fact or evidence.

That's actually not true. VU0 is/was widely used as a Co-Processor to the MIPS core. It hasn't been used in Micromode by many. Do a search.

 

Can you please provide links to the game code benchmarks you are referencing to be able to make this comparison? Thanks.

 

What exactly is "raw performance potential" anyway?

 

Well, first you argue Cell is impressive in peak/paper, but not in reality. Then you agree with me that PC's peak is far underutilized in reality, and then go on to assert that console devs can exploit HW quite well given the fixed target. So wouldn't your conclusion naturally be that devs will be able to exploit CELL pretty well, given it's fixed platform?

Then you'll have to fallback to ideas that coding for a general PC platform (non-fixed) is easier to exploit than a fixed CELL platform due to architectural complexity. Seems to be a bunch of handwaving to me.

 

Im sure the Crysis dev's said that Cell's PPE has better threading than 360's, i'll go and look..

Found it

 

And it also meens that in real world applicatiopns Cell has better Threading peformance

 

Graphicly there may not be much difference between the PS3 and Xbox 360 in the first two generations of games but how about when we're talking physics and animations?
Cell should easily stomp Xenon in physics and animations even in the first two generations of games.

 

Not to nitpick, but while Cell has a potential advantage in the physics department (i keep looking at PS3 as an X360 with a small PPU inside, which is a very flawed view on the thing but makes things easier for me), what we call "animation" has always been, and will be for a long time, largely dependant on the animators who have to tweak the routines.

Games like Jak&Daxter, Kingdom Hearts 2 and SOTC and lots more on all platforms have amazing animation not because of some hardware advantage, but because the animators who worked on those games are freaking GOOD.

I always put "animation" under the "art & contents" category, which is not dependant on the hardware (at this point at least), but more on the people who actually have to animate the characters in the game.

 

You've also got physics-based animation, simulation (and other buzz words which I can't think of right now... )

 

As i said, animation as it's been done lately and will be done in the near future, is largely a human-dependant thing.
Where physics are involved, then i also said Cell has a potential advantage. Be it "physics-based animation" or whatever you want to call it...

 

Animation is up to the developers. Each system has more than enough power for outstanding animation. Not to mention the 360 has more vertex power for the potential of "better" animation which could easily offset any advantage Cell might bring to the table.

 

In this interview didn't Crytek give the perfomance advantage to 360's core, while a threading advantage to PPE?

From what i understood, even if on Xecpu the two threads cant use the same amount of flops (like he hints to be possible on the PPE), it performs better.

 

Not that it makes much difference but Conroe has 3 SSE units per core. On a 3Ghz dual core thats 48 single precision GFLOPs I believe. (im assuming it can do 2 SP for every 1 DP).

 

Replying to your post Rockster:

1) - total amount of internal memory (2304 KB versus 360's 1024 KB: 2.25 times)
On-chip memories is very important to execution, because it's much closer to the execution units, than external memories, and adding L1 cache to the comparison, does not change the ratio all that much.

2) - number of processors (8 versus 360's 3: 2.7 times )
Yes an apples to oranges comparison with CELL having 7 processors that are console centric, and Xenon having more desktop computer like cores.

3) - number of *concurrent* threads (8 versus 360's 3: 2.66 times)
I highlighted concurrent for a reason. Multi-threading on each Xenon core is dual registers, and not dual execution. CELL can execute 8 threads at the same time to Xenon's three. Huge difference there. If you want, you can breakup each SPE register bank four times, for four sets of 32x128-bit registers providing four software threads to each SPE. That would be 30 threads versus 6 on the Xenon.

4) - GFLOPS rating (204 GFLOPS versus 360's 77 GFLOPS: 2.65 times)

Of course you don't like it, as it shows CELL's huge advantage in this area. If Xenon had this advantage, you would be making the same claims as I am. Sony choose a high FLOP design on purpose, because it aids games. Microsoft believes the same, or they could have had cores without VMX units. Not sure how you can claim the GFLOP rating does not mean much. You're reasoning for that?

5) - integer rating (~51.2 billion instructions per second versus 19.2 BIPS: 2.66 times)
You said nothing to counter this, so I guess you don't have a problem with it?

6) - internal bandwidth (204.8 GB/s versus shared 102 GB/sec L2 cache bandwidth: ~2 times)
You mentioned a number of bandwidth comparisons that could have been made. Feel free to provide those comparisons. If you want to use L1 cache fine, but please do weighted averages since those memories are much smaller than L1 and LS memories.

7) - external bandwidth (60 GB/sec versus 360's 22 GB/sec: ~3 times)
I don't see why CELL cannot access video memory, but like you, I'm not 100 percent sure. I did not include CELL's 5 GB/s additional I/O bandwidth by the way.

I never claimed that CELL was 2 to 3 times more powerful overall, but it is certainly 2-3 times more powerful at certain workloads, that are very game specific.

I can say without a doubt, that as a console processor, CELL is superior to the 360's CPU. Not that I think Xenon is crap mind you, I think it is also a great processor.