Kentsfield as an alternative to Cell in PS3

I think you got your wires crossed there. The statement was that havok 4.5 runs 5-10x faster than a typical scene on the PS3 (compared to pre 4.5 on PS3), which according to a slide at the presentation said was comparable to a tri-core PC.

*edit* sorry my bad, I read it wrong. As you were.

However I would still take this with a grain of salt. The comment "typical game scene" is not what was shown with 4.0 when the comparable to PC CPU statement was made.
Especially with physics things rarely work out so well

 

First of all; havoc alone isnt exactly the best way to measure of how well a cpu is for gaming. I dont know what the best way to measure it is, but i know its more to it than a physics engine.

What 3 cored CPU did they compare it to?

Further, Havoc, has spend alot of time optimizing their physics engine to run on the PS3, im positive that they have barely done anything to optimize it for a quad-core Conroe extreme

 

We have 3 core OOO processors?

 

First that "benchmark" with the CELL vs the tri-core CPU is BS. Becouse they dont name the "tri-core" CPU type (there is no Intel or AMD tri-core at all for PC) second they dont say the "tri-cores" MHZ speed and third no talk about optimizations or benchmark conditions = BS benchmark.

But lets pretend the "tri-core" CPU is actually the speed derived from 3 CPUs. Then what type of CPU and what MHZ speed where used, but they dont say it ehhh?
It could be a Celeron 3.2GHZ, it ccould be a P3 1GHZ, it could be another CPU but whe wont know and thats becouse they used weak CPUs from PC otherwise they would have stated what CPU and what MHZ. And I dont believe they used Kentsfield CPU for the benchmark becouse then they would have stated it to brag and claim victory becouse Kentsfield is the flag CPU for PC and I dont think Kentsfield was out to be buyable at that time where the benchmark took place!

 

I only know one CPU with 3 cores...

 

Same here, the Xenon, thats why im curious.

 

Discussion of the Havok results was done in the Havok thread. IIRC concensus was a quad-core CPU with 3 of the cores used for Havok, leaving one core to run the rest of the game.

For the OP, no-one can answer that. It's all theoretical at the moment. A lot depends on what algorithms you use and how effective they are, and for Cell we don't yet know what it can do in that regard. The only surefire points to mention are a Kentsfield wasn't available for the consoles and it costs a lot more than the CBE. Even if the Kentsfield performs better in all gaming situations, the economics means it would never find it's way into a current-gen console.

 

I would take Kentsfield over cell any day...

Just add up die size. If nothing else, it should give you a good idea which chip is more powerful...

I would take any decent desktop dual core over what is in the consoles..

 

Search the forum for the many discussions already on this! At the moment it's mostly hypothetical. You have people saying some tasks won't map well, and you have people showing new implementations of techniques that aren't supposed to map well to SPEs and yet show considerable speed-up. How much proves problematic to restructure onto SPEs is an unknown, and won't be known for a couple of years. It's probably nowhere near as bad as some parties would suggest though. GPGPU is showing similar clever thinking that's getting unexpected applications onto fast SIMD algorithms.

 

Console has limited space and little heat tolerance.

How power efficient are Cell and Kentsfield ? I did some googling and found unofficial estimates for 90nm Cell ranging from 50-80W (8 SPU, 4.6Ghz @ 1.3V). What kind of change should I expect when it goes to 65nm ? The 6Ghz Cell is supposed to be shown in ISSC on 11th Feb too.

According to Wikipedia, initial samples of 65nm Kentsfield consumed 110-130W @ 2.66Ghz, but is expected to go down to 80W.

 

Which is a stupid assumption because the code wasnt optimized for the Quad-core, while it has been for the cell.

The architecture itself, if used properly is probably better suited for 3d stuff, yes.

High memory bandwidth?? What exactly are you refering to, this is something that depends on how you connect the chip, and what kind of memory you have, not something about the cell itself.

Unless your refering to how fast the cache is, in which case i believe the Quad-core not only has several times more cache, but its also faster..

Higher speed interface with the GPU, is again depending on how you connect it, in a traditional PC environment compared to the cell, it is faster.

Considering that the Cell doesnt support dynamic branching, AI is going to be much better on a kentfield, the quad-core also has much local memory, which also helps. (8mb L2 cache vs 512kb)

 

Which is itself an assumption. If anything Cell would initially be at a disadvantage because its code would be derived from that same PC base, which isn't going to help it at all. That may have been more than offset now with 4.5 and the PS3 optimisations therein, but in 4.0, when that test was made, things didn't look like they were too optimised.

 

Guess he's refering to the direct CELL->RSX bandwidth, which is several times faster than any current PCI-E implementation something you just don't have on a classic PC architecture (yet).

Things like crowd-AI (which are no less demanding) are perfectly suitable even for SPEs in case you adopt your algorithms, see Heavenly Sword.

The Quad Core does not have any local memory but transparent cache. CELL has 2MB "real" local memory (LS).

 

Errr... Isn't the 3 core CPU they are referring to Xenon, and they can't say so because of an NDA with Microsoft requiring that they do not disclose benchmarks of Microsoft of Microsoft products? Microsoft has such clauses written into their operating system EULAs so you can only publish benchmark results that Microsoft permits, I would be surprised if they didn't have one for Xenon. Nothing else really makes sense. No other 3 core cpu exists and Xenon is a 3 core CPU that Havok has worked on.

 

You must mean dynamic branch prediction. Obviously all cores in the Cell support dynamic branching.

And there are a lot more factors to the equation for AI performance than just whether or not a CPU supports branch prediction.

 

Oh God, please, please, please don't do this. Looking at some narrowly defined data point and trying to draw wide-ranging conclusions from it is absolutely ludicrous. In fact, unless you have a great understanding of this "benchmark's" instruction mix compared to Cell's microarchitecture, you shouldn't really draw any conclusions from it. You might as well look at the 0-60 of a car running on sand and conclude that it's the best handling car on the road today.

 

Sure. When Havok wrote their physics engine, they didn't optimize for any platforms at all. All this time the PC has been running a really naff and slow physics system. It's only because Havok have optimized for Cell that's Cell's got a 10x speed up. The moment they get their fingers out and optimize Havok for Intel, that'll get a 20x speed up!

What's with this idea that Havok don't optimize their engine for any platforms and this Cell improvement in 4.5 is something out of the ordinary? You can be sure that when designing Havok for multicore PCs, the engineers put a lot of effort into making it fast. This idea that they release award winning software built of unoptimized code, for years and years, is ridiculous. Probably an offshoot of the 'developers are lazy' armchair coders mentality.

 

I didnt say that.

What i said was that it has not been optimized at all for a quad-core PC architecture. There is a big difference in optimizing for the PC architecture in general, where 99,99% have either dual core or single core cpu's, and optimizing for a quad-core cpu.