Not quite PS3

[Darkseider]

But I have no idea where to post this thought/question. Here goes...

With Toshiba incorporating the Cell into the Qosimo F50/G50, Sony with the PS3 and Zego and the many other implementations of the Cell. Will we ever see a Cell based desktop? Seeing that the Cell at 65nm cost approx. $80 to produce I don't see a reason why this couldn't become a reailty. For that matter a dual processor Cell desktop w/ 2 gig RAM, your GPU of choice, etc... just like a regular ole desktop. Yes it would be a niche machine but no more a niche than a Fixstar PowerStation. Cost wise they should be comparable as well. Any thoughts? Could this happen or am I just wishing for the impossible?

[billb]

How about a Powerstation with a GigaAccel 180? And all for under $8K ... But of course if I had one sitting in front of me right now my SPE's would be just as idle as they are now.

[Darkseider]

See that's what I don't get. Granted the GigaAccel 180 is a PowerXCell 8i chip but I can't understand why plain ole Cell B/E processors would be priced so high? They are produced in sufficient quantity that the cost on them has come down dramatically. That coupled with the smaller die size reduces costs even more. $1500 - $2000 should be more than attainable for a system similar to the PowerStation but with 2x Cell B/E processors and a commodity GPU. The only high(er) cost pieces would be the XDR RAM.

[ppietro]

But I have no idea where to post this thought/question. Here goes...

With Toshiba incorporating the Cell into the Qosimo F50/G50,

Ah - but that's not a Cell - it's a SpursEngine. Basically, it's only a cluster of SPEs used as a DSP. The IBM part of the Cell - the PPE - isn't in that chip. The main processor is still an x86 or x86_64, and is off-chip

Check out this link:
http://en.wikipedia.org/wiki/SpursEngine

Sony with the PS3 and Zego and the many other implementations of the Cell. Will we ever see a Cell based desktop?

I seriously doubt it. I'll explain why in a sec.

Seeing that the Cell at 65nm cost approx. $80 to produce I don't see a reason why this couldn't become a reailty.

The main reason is that the PPE core of the Cell doesn't do out-of-order processing - or have branch prediction - which most modern programs and operating systems require. This is the exact reason Apple dropped what would have been the G6 - the PowerPC Core of the Cell - and switched to Intel. (That's from the book "The Race For a New Game Machine" by Shippy & Phipps).

For that matter a dual processor Cell desktop w/ 2 gig RAM, your GPU of choice, etc... just like a regular ole desktop. Yes it would be a niche machine but no more a niche than a Fixstar PowerStation. Cost wise they should be comparable as well. Any thoughts? Could this happen or am I just wishing for the impossible?

It's really niche, to be honest.

Look - it's fun messing around with Linux on the PS3. But - we're enthusiasts, and we can appreciate the quirks in the system and optimize around it.

But - and I hate to say this - the average user would never unlock the full potential of the Cell. Instead - they'd just wonder why the performance of this new desktop system they bought reminds them of an old G3 mac.

I never thought I'd say this - but with the current architecture of the Cell - I actually think Toshiba's SpursEngine is the right way to go. You get the full x86 speed, plus SPEs to experiment with. Outside of supercomputers or game machines, I think the Cell will remain extremely niche until a full PowerPC core is used to replace the PPE.

As a side note: The PowerPC variant used in the 360 is interesting - it's called the Xenon. http://en.wikipedia.org/wiki/Xenon_(processor). It still has the in-order execution and non-branch prediction issues that plague the PPE - since each of the cores is a PPE - but since it's a tri-core design, a multi-threaded operating system and program could possibly make up enough speed to make it useful for desktops. However - I'm guessing it would be a cold day in hell before Microsoft makes that chip available to anyone.

(Also - it has no SPEs - although the VMX128 sounds interesting)

Cheers,
Paul

[Darkseider]

Granted the Cell is an in order processor but that doesn't make it slow. With 2x Cell processors and a GPU the machine should be more than usable. Think of it this way the PS3 has one Cell @ 3.2 Ghz w/ 6 SPEs available for use and YDL/Ubuntu/Fedora are all pretty damned snappy on it especially with only 256 meg of RAM to play with. Now I agree that not every user would be able to squeeze the performance out of the SPEs on this box but Sony and game devs are doing a damned good job as are other third parties. Think about the DEC Alpha when it was first introduced and how radically different it was. Once the architecture gained acceptance and started being put into all sorts of machines from servers to desktops even Microsoft made their OSs for it. So just like any other new processor out there it needs time to mature and applications written for it to take advantage of the SPEs. Same goes for compiler technology and optimized libraries, etc... Point is if they build it they will come and I am sure I am not the only one that thinks this way.

[ppietro]

Granted the Cell is an in order processor but that doesn't make it slow. With 2x Cell processors and a GPU the machine should be more than usable.

The problem is - although the Cell is quite fast - it is SERIOUSLY gimped as a general purpose processor by the weak performance of the PPE.

Let's look at a real world example. We'll take QEMU. QEMU is an emulator - not a virtualizer. So - even if you run it on an x86, it still has to translate.

So looking at my PS3 and my Core2Duo Lenovo laptop, we can boot the little test Linux under QEMU and run nbench. I get these results for the Original Bytemark Results. These are in comparison to a Pentium 90.

PS3:
Integer Index: 0.713
Floating Point Index: 0.783

Lenovo:
Integer Index: 3.891
Floating Point index: 2.445

So - the PS3 QEMU is .7 as fast as a Pentium 90, and the Core 2 Duo is 3, nearly 4 times faster. Both take a big hit with QEMU - it is an emulator - but until the PPE posts this kind of performance, there isn't going to be great demand for this as a general purpose processor. Not with the price points of x86.

Think of it this way the PS3 has one Cell @ 3.2 Ghz w/ 6 SPEs available for use and YDL/Ubuntu/Fedora are all pretty damned snappy on it especially with only 256 meg of RAM to play with.

The snappiness you see is more a factor of Linux than any raw speed of the PPE core of the Cell. It's roughly equivalent to a G3 Mac - G4 if you're being charitable. Considering how fast Macs are now, that's not great.

Now I agree that not every user would be able to squeeze the performance out of the SPEs on this box but Sony and game devs are doing a damned good job as are other third parties.

But - they're doing that in games - not in OSes. The Cell is a game processor, multimedia decoder and (sometimes) supercomputer core. It's not a general purpose processor - no matter what IBM tells you.

I highly encourage you to read "The Race For A New Game Machine" by David Shippy and Mickie Phipps. David was the lead IBM designer of the PPE. He lays out pretty clearly its strengths (low power, great for setting up SPEs) and weaknesses (lack of out-of-order processing, unsuitability for general purpose OSes).

Think about the DEC Alpha when it was first introduced and how radically different it was. Once the architecture gained acceptance and started being put into all sorts of machines from servers to desktops even Microsoft made their OSs for it.

It wasn't as radically different as the Cell. I worked at a company that used DEC Alphas - they were great. I ran Win NT 4.0, TRU-64 Unix and Debian Linux on them. But - they were also standard-ish architecture processors that ran super, super fast. You could work the compiler technology around it.

So just like any other new processor out there it needs time to mature and applications written for it to take advantage of the SPEs. Same goes for compiler technology and optimized libraries, etc... .

Ah - but that's the rub. Writing applications to take advantage of the SPEs requires you to code specifically for them. SPEs can not run PPE code and vice versa. Most Linux programmers program for a unified thread model - x86 - they're not going to write a separate program for the Cell. Even the PowerPC compiles you see out there are re-compiles from x86 code - a true PowerPC - like a G5 - isn't radically different from an x86 from a programmer's point of view. (Yes - I know it's RISC vs. CISC - but that's primarily a compiler difference unless you're writing assembly code.)

Point is if they build it they will come and I am sure I am not the only one that thinks this way

Obviously you're not - we wouldn't be discussing this on this great forum if you were.

But - I don't see a market for this machine. Not until IBM/Sony/Toshiba release a more general purpose version of the Cell. Right now, it's designed for Game Machines (PS3) and specialty servers (Zego, Blade).

Those that are interested - like myself - bought PS3s. As far as the market is concerned, they're covered. They're probably figuring that anyone else who's really interested can build them from scratch.

Cheers,
Paul

[Darkseider]

Question about the QEMU benchmark. Does QEMU take advantage of multiple cores? If so is it possible to lock it down to one core to then see the difference between a Core2 Duo and a Cell. Also remember that you are benchmarking a 3.2 Ghz Cell as opposed to a full on 4.0 Ghz Cell which can be readily produced. If it does scale linearly with clock speed that would be an additional 25% performance gain. So the .78 would be more in the realm of 1.0 or on par with a Pentium 90. Now under this assumption, which this is all it is, if the Core2 Duo scored what it did because it utilized both cores and using a dual Cell system scaling linearly or as close to you would be at about 50 - 55% the performance of the Core2 Duo you mentioned. That wouldn't be a half bad general purpose computer and with the addition of the SPEs to offload multimedia, transcoding, encoding tasks to along with a decent GPU I would argue aside from using OOo it would be superior to a Core2 Duo machine with a good GPU when gaming, watching HD movies/broadcasts, upscaling streams, etc... which is most of whats done on a home PC anyway.

[ppietro]

Question about the QEMU benchmark. Does QEMU take advantage of multiple cores?

Nope. It's single threaded. Also - this isn't even a particluarly fast core2duo - it's only a 2 GHz part.

I would argue aside from using OOo it would be superior to a Core2 Duo machine with a good GPU when gaming, watching HD movies/broadcasts, upscaling streams, etc... which is most of whats done on a home PC anyway.

That's very doubtful. Remember that the Core2Duo (Merom family in this laptop) also has SIMD extensions like MMX, SSE, SSE2, SSE3 , etc. I can already watch HD Movies/broadcast, upscale streams, etc. just fine with this laptop. And - this has a graphics card that's just barely above the framebuffer display of the PS3.

But - a bigger question - in the scenario you're describing - why bother with the Cell at all? You're actually describing the approach Toshiba is doing with the SpursEngine/Core2Duo combo right now - which doesn't need any special conversions to Power processor architecture and can run Windows right out of the box.

http://asia.cnet.com/reviews/notebooks/ ... 80p,00.htm

This would make an excellent desktop computer, don't you think?

Look - IBM isn't even positioning their Cell based Blade servers as general purpose machines:

http://www-03.ibm.com/systems/bladecent ... index.html

IBM BladeCenter QS 22 (PowerXCell): "High Performance computing blade ideal for accelerating key algorithms such as 3D rendering, compression and encryption with double precision floating point computing power."

IBM BladeCenter QS 21 (2 Cell): "High Performance Computing blade ideal for tasks that require single-precision processing such as image processing, signal processing and graphics rendering."

This is in contrast to their other Power & x86 BladeCenters which are

"Durable blade ideal for small database and application serving and for replacing traditional AIX®, IBM i or Linux® rack-mount and tower servers." (IBM BladeCenter JS12 Express - PowerPC based)

or

"Energy efficient mainstream blade ideal for collaboration, Web serving and general-purpose applications." (IBM BladeCenter LS21 twin AMD Opteron).

I'm not saying that a Cell based workstation wouldn't be a blast, mind you. But - I think it would only appeal to a very small group of folks. People who had specific SPE applications in mind. Things like 3D games, complex floating point computations, scientific research, etc.

I think it would be a hard sell to release this as a desktop to the general public.

I think this section of Wikipedia bears quoting:
http://en.wikipedia.org/wiki/Cell_processor

The first major commercial application of Cell was in Sony's PlayStation 3 game console. Mercury Computer Systems has a dual Cell server, a dual Cell blade configuration, a rugged computer, and a PCI Express accelerator board available in different stages of production. Toshiba has announced plans to incorporate Cell in high definition television sets. Exotic features such as the XDR memory subsystem and coherent Element Interconnect Bus (EIB) interconnect appear to position Cell for future applications in the supercomputing space to exploit the Cell processor's prowess in floating point kernels. IBM has announced plans to incorporate Cell processors as add-on cards into IBM System z9 mainframes, to enable them to be used as servers for MMORPGs.

The Cell architecture includes a novel memory coherence architecture for which IBM received many patents. The architecture emphasizes efficiency/watt, prioritizes bandwidth over latency, and favors peak computational throughput over simplicity of program code. For these reasons, Cell is widely regarded as a challenging environment for software development. IBM provides a comprehensive Linux-based Cell development platform to assist developers in confronting these challenges. Software adoption remains a key issue in whether Cell ultimately delivers on its performance potential. Despite those challenges, research has indicated that Cell excels at several types of scientific computation.

Cheers,
Paul

[Darkseider]

Well this stuff sure as heck sounds alot like what I mentioned.

http://www.techradar.com/news/computing ... &artc_pg=3

So when is the Cell desktop coming?

[ppietro]

Well this stuff sure as heck sounds alot like what I mentioned.

http://www.techradar.com/news/computing ... &artc_pg=3

So when is the Cell desktop coming?

Actually - that's the anti-Cell. Seriously. The greatest roadblock to adoption of the Cell will be increased programibility of GPUs from nVidia and ATI. We've already seen it with the Folding@Home project - the GPU clients have higher performance than the Cell clients.

Check out this page, and notice the nVidia stats:
http://fah-web.stanford.edu/cgi-bin/mai ... pe=osstats

The ATI specs don't look too good - until you read thru their FAQs and realize that, due to a hardware issue, the ATI's have to do two calculations for every one of the nVidia & Cell.

For Cell fans - that article represents the four horsemen of the apocalpyse.

Cheers,
Paul

[Darkseider]

I disagree, apparently the folks really interested in this are willing to write code specifically tailored to these units to take advantage of them. For some reason they have no problem with it yet as you pointed out earlier people will have a problem writing software to take advantage of the SPUs? They also mention throughout how this will enable real-time ray tracing, scientific research, CAD/CAM, being the strengths of a PC like this. The Cell already exhibits these strengths in these areas. The whole article is pretty specific when saying that for the GPGPU to work people have to write code in order to take advantage of it, same goes for Cell. The OS on Cell, Linux, runs well but not stellar. An example from the article

"Though GPUs are extremely efficient at streaming and processing data, most PC apps are serial in nature. A GPU can't turbocharge your word processor, for example, or speed up your anti-virus package."

The cell does just that, streaming and processing data. But when applications that can take advantage of the Cell and are tailored for it whether it be real time ray-tracing, folding@home, scientific research (MIT gravity well research), medical imaging, terrain mapping, etc... it kicks some serious ass, same as the article points out about GPGPU based systems. Like I said, build it and they will come, seems like ATI and nVidia understand this why doesn't S.T.I.? Everything you mentioned as a CON for a Cell desktop this article mentions as a PRO for a GPGPU based system.

[ppietro]

I disagree, apparently the folks really interested in this are willing to write code specifically tailored to these units to take advantage of them. For some reason they have no problem with it yet as you pointed out earlier people will have a problem writing software to take advantage of the SPUs?

It's a different programming model. In this case, it's easier to invoke the GPU to handle code. There's a lot less of the "throw/catch" mentality that you need when you program the PPE/SPEs of a Cell.

See here for some nVidia CUDA programming examples:
http://en.wikipedia.org/wiki/CUDA

Also - it just seems more programmer friendly. The very fact that there are Python, FORTRAN and Java implementations of CUDA speaks well.

Another difference may be RAM size. Each SPE currently has only 256k that it can use for programs and local data. nVidia cards that can run CUDA must have a minimum of 256M. That's quite a difference in the way you create your code.

Like I said, build it and they will come, seems like ATI and nVidia understand this why doesn't S.T.I.? Everything you mentioned as a CON for a Cell desktop this article mentions as a PRO for a GPGPU based system.

Blame IBM & Sony's Ken Kutaragi for this one - they may have made the Cell just too difficult to program. Apparently, they weren't expecting GPUs to achieve this kind of performance, and tried to pull some of those functions back to the CPU. I think they were wrong.

According to "The Race For a New Game Machine" (Shippy & Phipps), Sony didn't even want a GPU in the PS3. They were designing a simple framebuffer chip in-house, but ran out of time and went for an off-the-shelf nVidia design. That's why the RSX in the PS3 is weaker compared to the ATI part in the 360.

This is the same crazy model - CPU + Vector DSP ignoring the GPU - that Sony did with the PS2's Emotion Engine + Graphics Synthesizer. Then, it worked, although it drove normal programmers crazy. Now, I'm not so sure they can get away with it.

Obviously, a lot of my pessimism is personal opinion.

But - I've been in the computer industry for a long time now, and I just don't see this architecture - as cool as it is - surviving the Intel/AMD/ATi/nVidia onslaught in any major capacity. I personally think IBM/Toshiba/Sony have created the Emotion Engine mark 2 with the Cell. A cool piece of technology to be sure - but one that will end up driving mostly Sony game machines, TVs and the odd supercomputer/blade server.

I'm not even sure that Toshiba's SpursEngine will survive - although they have to recoup their investment somehow. Check this post out:
http://www.reghardware.co.uk/2008/06/17 ... ursengibe/

Cheers,
Paul

[ppietro]

Hiya!

In the interest of having a good discussion, I started reading some posts here. I don't know if they're pro-GPU or pro-Cell - just more data to think about:

http://gpgpu.org/forums/viewtopic.php?t=4500

http://www.gpgpu.org/forums/viewtopic.php?p=17055

http://forums.nvidia.com/lofiversion/in ... 58846.html

http://www.ddj.com/hpc-high-performance ... /197801624

http://web.mac.com/sfj/Jenks_Parallel_P ... mming.html

http://forum.beyond3d.com/showthread.php?t=26200

http://www.nvnews.net/vbulletin/archive ... 60855.html

http://www.anandtech.com/video/showdoc.aspx?i=2453&p=4
(I really like this one for comparing 360 to PS3, even though it's slightly out of date. It lays out the design decisions well.)

Cheers,
Paul

[Darkseider]

All very interesting threads, good job Paul! After reading through them, except the Anand one I had problems pulling it up , seems that there is a split opinion on Cell and GPGPU. Although I do have to say that the threads do support some type of demand for a Cell based desktop, especially when people are comparing the $7000 PCI-E Cell cards and the $10k IBM blades to a PS3. Granted the PS3 can fill some of that niche but lacks in memory, 2 SPUs and honestly a second Cell processor. It would be really interesting to see a unit like this come to market. Some folks did mention the lack of ability of the Cell B/E to handle DP floats well and in answer to that is the PowerXCell 8i. Hey anything can happen, right?

[ppietro]

By the way - did you see our post over here? How does $2000 (U.S.) grab you?

viewtopic.php?f=13&t=5012

Cheers,
Paul