Oh look, the usual suspects showed up to turn this into a console war thread. Guys, the 'war' is over. All three consoles are here and have good games. Anyone peddling the myth that the PS3 doesn't have good games is living in a world of their own making.

@Sarah, Cell BE is very fast on single precision and is still faster than any desktop chip for Double precision. IIRC each Cell at about 3.2GHz hits in the region of 250GFLOPS single precision and 25GFLOPS double precision. Yes, the SPEs are optimized for SP work. However, apart from the most recent quad core (and multiple execution units per core) x86 processors I'm not aware of any existing x86 that comes close to Cell floating point performance. There is a revision of the Cell that further optimizes floating point math vastly improving the performance in double precision work.

@the AC who continues the blather about the PS3 suffering because Cell wasn't designed for gaming. You sir are a fool. Cell BE was designed primarily for the game console market. It was a custom design job by Sony/Toshiba/IBM as a consortium, specifically for the PS3. It so happens that the high performance architecture they created is also suited to HPC. The RSX is not the GeForce 7800, it's a derivation from that architecture. There is additional hardware onboard as well as design revisions specifically targeted at the needs of a game console architecture.

You're utterly wrong to state that something with two or four generic cores would be better suited to gaming. It would be easier on day 1 to use, but it would most certainly not be better suited.

@Iain

Sony specifically wanted to segregate the memory of Cell and the memory of the RSX. Cell can write through RSX to the GPU memory far faster than the 16Mbps read performance. There aren't many occasions when you are going to want the CPU to read directly from the GPUs memory. The only time when this becomes an issue is when you try to make the PS3 become something it's not. How many times have we ever heard of PC applications where the CPU is reading the GPU's memory directly?

I attended a public talk by a Playstation 3 (PS3) games developer in which he said that future IBM Cell processors would contain hundreds of Synergistic Processing Elements (SPEs). He did this in order to emphasise the importance of using as many SPEs as possible when writing very high performance games.

It is a reasonable assumption that the PS4 will contain hundreds of SPEs. In PS3 Linux only 6 SPEs are available to the user. If I assume that the PS4's Cell Processor will have 200 SPEs then this would imply that the PS4 will perform scientific calculations around 30 times faster than the PS3.

Note that I am assuming that the PS4's Cell processor will be clocked at the same frequency as that in the PS3 to ensure that PS3 games will run on the PS4.

Talk to IBM about their Cell based blades....they *don't* come with a 256MB memory limit.

Absolutely agree with your comment about typical desktop Linux, but I also have to point out to everyone that Linux was once touted as having an incredibly light memory footprint, what has happened there then? Since when did Linux need more than 256MB to manage a simple desktop for users? Ah, you'll be worried about the frame buffer size? 2MP at 4 bytes per pixel is still only 8MB. Even triple buffered that's 24MB. It's not like you're going to want to run texture intensive games under Linux on your PS3. Honestly, people seem to forget that 2GB per desktop is an obscene amount of RAM for a simple desktop. Only in Microsoft's world does it require 1GB to run the OS and a second GB to ensure you have sufficient RAM to run your standard office applications. Windows plus Office used to run very well on systems with 4MB of system RAM and maybe 1MB of video memory. Have we lost the plot that much that we can't run a word processor in less than a quarter gig of RAM?

Sheesh!

I think linking 16 PS3s is a neat trick in this application (maybe we can stick some PS3s in our next budget ;-) ), but several commentators are missing an important point.

The key is the fact that his code runs EXTREMELY well in parallel, with little communication overhead, otherwise the Gigabit ethernet would completely kill the performance. I would be very interested in the speed-up he achieves with respect to running on a single PS3. I work on various parallel computing problems, and would LOVE to see 16 cell processors configured in shared memory formation (with crossbar switch or some other fast interconnect) then my code might really run neatly. As it is I will stick with multiple dual core opterons (Barcelona, where are you?), or the Nehalem type Xeons.

*IF* there is a PS4, and *if* it's based on a new and improved Cell then I would think that the version used will have a 'classic Cell' mode that it can switch to, and the system will downclock if needed.

If there is a PS4 and backwards compatibility is still a feature on the table. Then look for Sony to have substantially upgraded the Cell to allow the box to do real time ray tracing, and the GPU will become a secondary issue. Heck, they might even get away with 4 slightly uprated Cells working together with an RSX for handling the screen. I know that they are talking about a 10 year timescale, but the investment needed to go to yet another custom architecture is astronomical, and there are some serious payoffs for developing the Cell further over the coming 5 years before decisions have to be made.

forgiving the X86 only centric readers commenting so far its clear that they dont understand the PPC/Altivec or the other vector units on the CELL

its true the PPC/Altivec/CELL can stand some optimisations when running PPC linux ,as the old Mac PPC Coders obviously didnt want to undemine the OSx Altivec optimisations inside that OS.

it strikes me reading the linked page to the cluster page http://gravity.phy.umassd.edu/ps3.html that infact it does appear that he used a generic PPC linux, and didnt bother to look at optimisations outside the compiler.

apparently not even considered using the PPC coders choice, 'PPC Gentoo' were all your current PPC Altivec multimedia optimised code is first produced by the likes of lu_zero and the altivec guys at Power dev.

http://www.powerdeveloper.org/forums/viewtopic.php?t=1082

take a look at the old school PPC guys code thread here http://www.powerdeveloper.org/forums/viewtopic.php?t=1426&postdays=0&postorder=asc&start=0

for a practical tryed and tested code base.

http://www.powerdeveloper.org/forums/viewtopic.php?t=1426&postdays=0&postorder=asc&start=0

http://www.powerdeveloper.org/forums/viewtopic.php?t=1494

its clear that the PPC linux DOES NOT currently use any Altivec vector optimisations (as the x86 linux with its limited MMX vector unit does).

if you read some of the Altivec threads found at powerdeveloper, you will find some answers/Numbers, and the lads are always looking for feed back on the likes of the freevec optimised codebase http://bbrv.blogspot.com/2008/02/freevec-updated.html ,and helping the new user/PPC programmer better understand the PPC/CELL and their vector optimisations you might find informative and practical.

look at this chart in the thread above for an indication of a generic memcopy/network speedup for instance.

-----------

gunnar:

Quick update:

I have looked at the Linux Kernel code a bit.

Its not difficult to improve the performance on PPC.

The Linux Kernel has a copy function which is used to cope between kernel and user space.

As this function copies a lot of data its performance has direct influence on network or filessystem performance.

Improving the speed of it was actuelly easy as you can see:

http://www2.greyhound-data.com/gunnar/glibc/throughput_970.gif

http://www2.greyhound-data.com/gunnar/glibc/throughput_cell.gif

Especially on the Cell improving this function

does result in feelable performance improvements of the total Linux system.

I'll do some more testing and then publish the patch soon.

Cheers

Gunnar

-----------------------------

Posted: Wed Nov 14, 2007 4:43 am Post subject: Possible benefits - optimization for PowerPC

--------------------------------------------------------------------------------

Hello,

Mostly all Linux applications are developed in C or C++. People often believe that C compiler are good enough to guarentee good performance. This is unfortunately not the case, especially on PowerPC manual optimization can make a huge difference.

Here an example of a memcpy on PowerPC...

a) Normal C routine working on Byte

150 MB/sec

b) Normal C routine working on Long (32bit)

800 MB/sec

c) Normal C routine working on quad (64bit)

1000 MB/sec

** This is best performance that you can archive by algorithm design, using C language **

d) Normal C routine working on quad (64bit) + with two ASM Cache-instruction added.

1380 MB/sec

e) ASM routine better optimized for this PPC architecture

2750 MB/sec

From 150 MB/sec to 2750 MB/sec is quite a difference.

As you can see by using optimized code you can achieve 20 times better performance!

Gunnar

----------------------------------

"Probably not "typical desktop Linux" at all, eh?"

Of course not!!

But several people here brought up performance under various circumstances including desktop. I wanted to explain that it's not designed for such a workload, but it's fantastically fast floating point abilities make it fantastically well suited to other workloads.

---

The comments that it's not suited for games are just dumb. The PS3 is easily the most powerful of the game consoles, even if it's power hasn't been fully realized.

Note that this is different from saying "The PS3 is the winner". In the last two generations of game consoles, the top spec system sold the worst and the lower specced PSX and PS2 kicked ass on the market.

The PS3 is reminding me of a lot of systems, none of them winners on the market.

The Sega Saturn, for example, also had fantastic processing power in a bizarre architecture that would require special investment to take advantage of.

The Apple Pippin too. It was more PC like, and had the same sort of ill-conceived banana controllers that often go with really expensive systems that do things nobody really cares about. (Sony did dump those banana controllers though.)

If developers are going to try to target multiple platforms, that will only make it less likely that a bizarre architecture will be utilized effectively.

You can't really brush this off by saying graphics aren't important because if graphics aren't important, than the Playstation 3 has no competitive advantage against the cheaper to buy, cheaper to develop for Nintendo Wii or X-Box 360.

Yes, I should probably have mentioned that even the official Sony docs suggest that you avoid using that painfully small bandwidth from the CPU to the RSX's memory, and get the RSX to write the data to main memory where the CPU can read it at the normal (very, very quick) speed.

Although, this is discussing the specific case of trying to use RSX memory when the CPU memory pool is already full - at which point there isn't much space to do such a thing. A small 'paging' area would work, somewhat like the old memory pages on the 128k Spectrum, if you're elderly enough to remember working on that.

I don't know the guy's application code, but it might well be that 256Mb is enough anyway. There are plenty of clever things you can do in that space, without needing any more.

It'll obviously perform better $ for $ than a desktop PC because this is 16 playstations being used in a way that would effectively make a huge loss for Sony, even if Sony hadn't donated.

It's ironic that they are helping this guy, not the least because of their efforts at stopping folk using PSPs to run their own software, and also because if we all rushed out and get PS3s to use as linux machines Sony will lose even more.

Obviously they are hoping we'll believe the cell is a supercomputer. D'oh.

> The comments that it's not suited for games are just dumb.

> The PS3 is easily the most powerful of the game consoles

Today, that's like being the most powerful flea. Unless you wave your arms and say "If a flea were human size it could jump really high!" but the truth is, fleas don't jump that high relatively speaking.

And relatively speaking a moderately priced gaming PC can easily trounce the PS3's gaming ability today.

That will only be more true as time progresses. Even if any developers do get around to getting at this mythical "potential" that everyone waffles about developers using "in a few years", it's really to justify what is nothing more than developers chucking out crap games, at a high price, and getting away with it because the platform is new.

It's a bit stupid to wait until your hardware is obsolete before showing what it could have done. Especially when your competitor [themselves stupid enough to sell 733mhz desktop linux boxes at knock down price, in much the same way as this article describes, with their earlier xbox] has actually made it a doddle to use the performance in the 360.

Don't get me wrong, a 360 is still an expensive way to keep the door open so your real computer doesn't overheat...assuming it's not keeping a door open at the repair centre.

But it has demonstrated quite obviously and clearly across a wide number of games and across a wide number of different developers that the PS3's mythical extra power has made no difference or impact at all - indeed, if anything, often the reverse is true.

Doing it in the future, even if that is possible, will be as pointless as Microsoft bringing out a bug free version of windows 95 next year with a "See, we told you it could have worked!" tag.

Highlander believes that if the Cell processor is upgraded in the PS4 then the upgrade will be to clock the Cell processor at a higher frequency and not to increase the number of SPEs. I think that this is unlikely. This is because I understand that doubling the clock frequency of a processor increases the power consumed by a factor of four.

A far better use of that extra power would be to quadruple the number of transistors in the Cell processor chip. This is the approach that intel has take with Core Duo processors where extra performance in gained by increasing the number of processors on the chip and not by increasing the clock frequency.

I extimate that quadruping the number of transistors in the Cell processors chip would increase the number of SPEs from 7/8 to around 40.

According to Moore's law the number of transistors on a chip doubles every 18 months. It is therefore perfectly possible to imagine a Cell processor with 40 SPEs being available 3 years after the manufacture of the original Cell processor.

IBM Cell processor documentation does not specify the number of SPEs which indicates that future Cell processors will most probably include more SPEs.