Cell Architecture Explained

IdiotOnMyLeft writes "OSNews features an article written by Nicholas Blachford about the new processor developed by IBM and Sony for their Playstation 3 console. The article goes deep inside the Cell architecture and describes why it is a revolutionary step forwards in technology and until now, the most serious threat to x86. '5 dual core Opterons directly connected via HyperTransport should be able to achieve a similar level of performance in stream processing - as a single Cell. The PlayStation 3 is expected to have have 4 Cells.'"

Seeing is believing

It's not like we haven't heard it before. It usually turns out to be halfish-truish for some restricted subset of operations in a theoretical setting, you know where you discount busses, memory and latencies.

Re:Seeing is believing

No, this sort of architecture is a general trend towards paralellization. It is smart, and it is known to work, and I would expect some bright Sparc wise people to chime in and say "u-huh" and some SGI wise people to chime in and say "I've seen some of this before." The OS people [dragonflybsd.org] are starting to move things in this direction, and I've heard that Darwin has had the asynchronous messaging type threading model for a while (RTFA: the article explicitly mentions Tiger's GPU leveraging techniques). If you have the head for it, try reading up on NUMA and compare that with SMP.

The math is simple. CPUs are CPUs, and anyone can make one that is the same speed as the competition, and if they do it second they can do it cheaper. The guy that can make 20 CPUs work like one CPU that does 20 times the work in a given time will win because he can always just throw more hardware at the problem. The SMP guys have to go back to the drawing board. In this case, the only way to beat-em is to join-em. Maybe doing the specific "Cell" computing design isn't it, but the ol' PC is dead. If these things start hitting the commodity price-points.

That's a big, fat IF. So, don't bet on it (yet), but it's even worse to ignore it.

This is beginning to sound more like

a DBZ reference: "Part 4: Cell Vs the PC"

Re:This is beginning to sound more like

a DBZ reference: "Part 4: Cell Vs the PC"

The 45 episode saga in which:

Bill Gates becomes a cyborg and summons the forces of evil.

A new Cell is constructed out of unsold Itaniums (Not to be confused with the Cell built by Sony, which is a friendly robot that is found out to be good. ( Until he is found out to be evil when the heroes notice he is under the control of the cyborg Bill Gates who has been behind the charade the entire time) and challenges the world to a rematch of earth shattering proportions

Second string characters have meaningless conversations that take up entire episodes

There is hilarious comic relief from common citizens in various towns as their cities crumble around them

Krillin dies

The dragon is summoned

Goku gets a haircut ...Good lord I should have my anime viewers license revoked for knowing all that crap.

if it sounds too good to be true..

..it probably is.

was the ps2 the supercomputer it was said to be...?

the author goes on to suggest that cell workstations would smoke x86 counterparts.. but says at the same time that there probably wont be that many of them.

wtf? though in-between the lines you can read at the end that he also thinks a single g5-cpu workstation would 'smoke' x86's...

Re:if it sounds too good to be true..

When it was released, the Emotion Engine in the PS2 actually would have been pretty wicked for supercomputing applications if Sony had sold a version with faster interconnects and more RAM. The processors in the PS2 are designed almost entirely to crunch vector operations, which is what most scientific codes rely on. It's really an excellent computer, it just sucks at graphics. The 4MB of uncompressed video memory and lack of hardware texture support are particularly ugly.

I suspect that the main reason there was never an Emotion Engine based cluster product was because the high performance market is tiny, especially compared to the console market, and Sony was already having trouble meeting demand with their exotic chipset when it first came out.

Anyways, I think the guy does go overboard about this new architecture. It probably will be a lot faster than PCs at certain tasks but you can only fit so many transistors in a chip. The cell stuff is cool though, it seems to fit a lot better with what most computers spend their time processing unless you're doing a lot of compiling or database operations.

fanboy article

This was not a technology article. That was a "I for one, welcome our new cell processor overlords.." article.

I don't see anything in the cell arcitecure that would fundamentally make the same number of transistors at the same speed operate faster. I see lots of bottlenecks, IO overhead and wastet transistors. If there is some magical powerful thing that these can do SO much better than the current X86 instruction set and hardware, guess what, it'll adapt.

x86 adapted to RISC being "wildly faster" and, in the end, became better RISC than RISC was by translating more memory efficent X86 instruction onto a RISC backend. It adapted to SIMD (Single Instruciton, Multiple Data) efficiency issues by adding MMX/MMX2/SSD/SSD2 and 3DNow. It adapted to the reality of 64 bit address space and the need for more registers with the new X64 instruction set extensions. AMD and Intel could add cell hardware and instructions too if they offered anything special, which I highly doubt they will.

What always confused me

Something that has always confused me in gaming consoles is that, despite incredibly powerful hardware (processors, graphical chips, etc.), the system developers seemingly always neglect to put in enough RAM for most games to perform to their potential. Many PC ports often have portions compromised due to the lack of RAM, and system speeds also suffer because of this.

Seeing how RAM is increasingly becoming cheaper, is it possible that new systems like the PlayStation3 might be able to provide RAM that actually allows games to reach their potential along with this new cell hardware?

I'll believe it when I see it

I'll believe it when I see it. Sony made outrageous claims with the PS2 in the year or so before launch, I see no reason to believe this will be any different.

On paper an Emotion Engine was supposed to destroy everything, but achieving maximum throughput was difficult and other contraints such as I/O and memory hampered performance. Programmers had to learn a very different way of programming to make full use of the processor and it's two vector units.

A Cell might be a killer chip on paper, but real-world hardware with I/O latency and memory contraints will bring things down to a more reasonable level. Don't forget that multiprocessor programming is *hard*.

Hopefully, developing software for Cell chips will be easier then the early days of the PS2, Sony has already said as much a few months ago.

Can this be taken seriously?

Quotes from article:

"GPUs will provide the only viable competition to the Cell but even then for a number of reasons I don't think they will be able to catch the Cell."

Did this guy forget that NVidia is designing the GPU for PS3? If Cell is so almighty, why does Sony uses NVidia GPU instead of using more Cells for graphic prosessing?

"There is another reason I don't think Nvidia or ATI will be able to match the Cell's performance anytime soon."

Of course, Cell based products won't be available anytime soon either. According to the current rumors, PS3 will be available in Japan in Spring 2006 and elsewhere in Autumn 2006. One and half years equals a generation in the GPU world...

I love this kind of articles where some future products are compared against current ones and declared as a clear winners...

not a new architecture, and it's going to be tough

This sounds like a little PVM-cluster-on-a-chip. It also sounds like it's a pain to program and will, in the short term, suffer from the same problems that Intel's Itanium suffers from: it tries to push too much work on the compiler or software developer.

In the long term, it's nice that companies are exploring these kinds of architectures. It's not nice that they are trying to monopolize what are pretty straightforward architectural choices with patents. This may be a new CPU, but there is little that is new about having a bunch of fast processors interconnected via a reconfigurable network; these just happen to be on the same chip.

Cool!

85 Celcius operation with heat sink

Well, perhaps "cool!" is not the correct response...

Is it just me?

Or does the logical extension of this chart:

http://www.blachford.info/computer/Cells/Cell_Dist ributed.gif [blachford.info]

Make it look a little more like a HAL than a Cell?

Compiler technology

One question which was not addressed fully in the article was how do you compile/test programs for this thing.

The potential of parallel architectures has never been in doubt since the early days of the Cray monsters - but how to compile code to use all the features efficiently has.

I don't believe that we see the full advantage of these types of architecture exploited without some similar break-through in software tools.

Mind you the hardware rocks...

next please

I'm sorry, but Sony can kiss my ass.

This is from the company that said the Playstation 2 would have Toy Story quality graphics, and be able to render FF8 quality FMVs in real time (thus making FMVs no longer required). It was essentially that bullshit hype that killed the Dreamcast... so yeah, now they're at it again.

Maybe I'll be proven wrong, but I doubt their system will be able to do anywhere near what they say it can in practical application.

Re:next please

Maybe this time Sony will see fit to include that really high-tech mipmapping [number-none.com] stuff, so their console isn't the King of Sparkle.

(Stupid Sony, I've had my PS/2 for about a year now and I still notice it almost every time I play. Can't believe how unbelievably stupid they were not to include it. That one change, which by computer graphics standards is dirt cheap, would have massively improved its graphics. Anti-aliasing, on the other hand, is expensive to do right, so while I expect it on this next generation, at least while running in NTSC or PAL, I wouldn't have expected it in the PS/2 era. Though some managed, I think....)

After that, I don't trust them any farther than I can throw them. The PS/2's graphics subsystem wasn't an Eighth Wonder of the World, it was an incompetent disgrace. Fortunately most of their fanboys are so stuck up the ass with Sony that it took them years to notice, instead of it jumping out at them in 5 seconds.

I have it for the game selection, and I like the games, I like the controller, I like the case, etc... but the graphics are far, far worse than what they should have been. You have to reach back for years and years to find anything else that didn't do mipmapping.

(I've also played the Dreamcast some more lately. It definately pumps out fewer polygons, but equally definately, they are higher quality polygons, and the fact that the Dreamcast clearly has mip-mapping is no small part of that. The PS/2 was a step forward in some ways, but a big step back from the DC in others.)

How could they possibly do this cheap?

I'm willing to believe that a 4.6 GHz chip with 8 ALUs and high bandwidth memory would be fast, but even in bulk, there's no way they can afford to put 4 of them in a sub-$500 game console.

I've been reading PR about the Cell for years, and nothing I've ever read has seemed even remotely plausible. Is there any objective information that even comes close to substantiating any of these claims?

STI

i didn't understand any of the document, but damn it looks fast

multicore, stream-processing, vector-oriented BS

While I tend to agree the Cell is an impressive architecture, this article is a steaming pile of B.S.

No cache for CPUs? A breakthrough? Hello! Both PSone and PS2 have the so-called scratchpad, which is what the Cell seems to have: a cache which has to be managed explicitly by the programmer. Breaking news: This is a royal pain in the ass. And calculating bandwidth when reading from this tiny scratchpads makes about as much sense as calculating the speed at which a x86 processor can execute MOV EAX, EBX.

Magically "the OS solves everything", and, in an obvious attempt to automatically get OSS-crowd support (is that "slashdot-trolling" or "slashdot-baiting"?) the triumph of Linux is predicted, because it's portable. Good luck getting the Linux kernel and GCC compiled, let alone running well on a massively parallel array of tiny CPUs without cache.

Unfair comparison

Since the main goal of the chip is to pump through graphics, regardless of what device its in, a GPU is better grounds for comparison.

From TFA: "Existing GPUs can provide massive processing power when programmed properly, the difference is the Cell will be cheaper and several times faster."

Its supposed to do 250GFlops when? 2 years from now? Apparently the Geforce 6800 Ultra will do 40GFlops and thats today.... extrapolate with some doubling here and there it seems a lot more reasonable.

So the big thing is that it comes down to programming. It came up a few times in the article "Doing this will make it faster but will make for one hell of a time for the programmers" It may have a huge potential but may take a while to get everything efficiently as Sony would like. Reminds me of when the GF3 first came out and was beaten by the GF2U in some tests. IIRC it took a while for games to come out that took advantage of its programability. It'll be interesting to see how well the programmers can fair between now and Cell's release.

It needs some serious software to work!

There are several assumptions that lead to tremendous theoretical performance figures. The simple fact is that like the Itanium, the Cell processor depends on some rather complicated software that will solve issues like parallelism, coherency etc. The article clearly states that the Cell architecture is a combination of software and hardware (1st page). This is good because performance can always increase (via a better OS or microcode) but it is also bad because it means that initial versions may not stand up to their performance claims.

Also, let's not forget that developers will be unable to keep up, unless some highly sophisticated libraries and languages are made available. I really don't expect the majority of developers to be able to cope with massive parallelism from the beggining (not just 2x SMP or hyperthreading, this needs a totally different mindset).

To sum this up: the hardware will deliver, but the software is a critical unknown in the equation. I have faith in IBM ;-)

P.

Nicholas Blachford is an idiot. Please don't read.

Nicholas Blachford is an idiot. Do not read any of his articles. Just to give you the best of Nicholas, read his antigravity article and visit his web site:

http://www.blachford.info/quantum/gravity.html

Also, look at the nose pictures of him ;)

http://www.blachford.info/other/me.html

Seriously, the guy has burned most of his sane braincells.

For serious laugh, read his article series 'building the next generation' from osnews. I really got good laughs from that 4 part series.

Also, it didn't take long to spot a totally idiotic statement from todays slashdotted article:

> Parallel programming is usually complex but in this case the OS will look at the
> resources it has and distribute tasks accordingly, this process does not
> involve re-programming.

Here Nicholas misses the core problem of parallel programming. The program algorithms _always_ have to made parallel. The OS can't do it.

x86, Apple etc Vs Cell my arse

I'm not actually surprised that so-called journalists, especially the technical kind, get good salaries. If you look at the painful clowns running the show at ZDNet, and most technical publications for that matter, including such wonder rags, such as the Register, you know that the Agenda is almost the most important thing. The actual realities of the tech world be damned as long as you have someone passing you your monthly wad of cash.

And this story is no different.

As many have noted, Sony did exactly this kind of hyping the last time around when the PS2, with its emotion engine, was supposed to be the future of all things computing. As everyone knows, the PS2 was a real pain to code for, and the actual performance was not better than the PC's of the day. The Cell will undoubtedly suffer from the same problems when it comes to coding real applications. Concurrency and parrallelism do not an easier coding experience make.

I have no doubt that this thing will be good, but I absolutely doubt that it will have much or any effect on the x86 world of computing. The G4 processor, when it came out with the Altivec SIMD processsor, which was apparently better than SSE at the time didn't turn Apple into the next Microsoft overnight either, did it?

So, I expect that the x86 world will continue to thrive and that Apple will stick some of these Cell processors, having as they do a PPC 970, aka G5, in their core, in some of their machines and will make the usual wild RDF claims about how hot it is while it will be used by only a small fraction of actual Mac developers in reality, the Mac having to maintain backward compatibility only slightly less then the x86 world does.

In other words, it'll be business as usual.

Re:What's that? Microsoft isn't supporting it?

You have not read it. It will be on a specific class of tasks. It is similar to modern GPUs. They are faster then 10 opterons on a specific task.

Back to the article. The guy seems to understand hardware, but he does not understand shit about software. Once he got past the first 3 parts he started babbling. Linux on cell, so on, so fourth. If he just read his previous parts he should have hit himself on the head. The only type of linux this can run is mcLinux. There is no memory protection as such. So no Linux, no Windows past 2000, no MacOS past X, so on so fourth.

Similarly, it is all nice and well about cell software beasties making herds by themselves and cooperating on a task. I am going to be a spoilsport and ask a nasty question: Err.. What about a security model? Memory protection? Privilege model for communications? So on so fourth...

To continue on this, the power of a modern general purpose OS is the task switching. How long does it take to load and store the context of the vector processing units? Doing so requires moving their dedicated memory to main memory. This will take ages.

Overall, this is a design similar to Cray 1 initial design. Cray initial design smashed the IBM, DEC (and lesser fish) monopoly on big computing iron to bits. Unfortunately the next thing the people buying the Cray asked for was "can we share this resource between two people?". The answer was provided eventually, but by the time Cray could do all the nifty time sharing and memory management tricks necessary to do this its advantage was no longer phenomenal. And all people who could use Crays for single tasks with manual scheduling actually continued to use it that way. But it did not even dent the general purpose big iron market.

Re:What's that? Microsoft isn't supporting it?

Back to the article. The guy seems to understand hardware, but he does not understand shit about software.

This part I agree with. His statements regarding abstraction are just flat out incorrect. Is this going to be programmed in assembly only? I think not...and if not there is significant abstraction involved. The thing that's closest to his point is that multiple *layers* of abstraction tend to add significant overhead. That doesn't mean that program-level abstractions do.

Once he got past the first 3 parts he started babbling. Linux on cell, so on, so fourth. If he just read his previous parts he should have hit himself on the head. The only type of linux this can run is mcLinux. There is no memory protection as such. So no Linux, no Windows past 2000, no MacOS past X, so on so fourth.

There is memory protection if the PU is in fact "something like a G5". IBM would have to be insane not include a MMU, and it has already stated that it's going to build workstations based on the Cell architecture.

All in all, interesting stuff...we'll see how it plays out. :-)

To continue on this, the power of a modern general purpose OS is the task switching. How long does it take to load and store the context of the vector processing units? Doing so requires moving their dedicated memory to main memory. This will take ages.

This, of course, depends on how many cells are in the box (with 8 vector units per cell) and how many tasks need vector units. The main purpose of the vector units in an interactive workstation will be multimedia processing. How many multimedia applications can you view at once? For me, the answer is one. The vector units may be useful for other things like engineering simulation and pattern matching, but once again how many different tasks using those features will be running at once? Plus if the processors are cheap enough to put 4 in a Playstation, one hopes the workstations will have 8 to 32 of them.

Overall, this is a design similar to Cray 1 initial design. Cray initial design smashed the IBM, DEC (and lesser fish) monopoly on big computing iron to bits. Unfortunately the next thing the people buying the Cray asked for was "can we share this resource between two people?". The answer was provided eventually, but by the time Cray could do all the nifty time sharing and memory management tricks necessary to do this its advantage was no longer phenomenal. And all people who could use Crays for single tasks with manual scheduling actually continued to use it that way. But it did not even dent the general purpose big iron market.

Two points. First, this is based on an already successful processor - the Power series. It already multitasks :-) and is used in a wide range of applications. Second, this will be a low-cost part. Crays were a super high-end system, which cost millions of dollars. Your analogy doesn't work.

Re:What's that? Microsoft isn't supporting it?

"No Apps"? Try every single video game publisher in the world.

And besides, this isn't about "Office" style apps. Its about games, and more importantly: its about home media centers. I think the Windows MCE is going to have its rear-end handed to it by the PS3.

When you consider that a cell-based PS3 could have a computational power of *several times* a 3 GHz Pentium...

You have to ask, what's more likely: that Intel can get around IBM/Toshiba patents in time for Windows to conquer the living room with a faster box? (That's if they can even build a secure, stable OS with a decent UI). Or that Sony, now armed with the worlds fastest consumer-computing platform, an enormous user base and years of TiVO experience, will own the living room media center market.

If I had to bet on who builds a better media-center PC .. Sony or MSFT... I'd say its absolutely no contest. Sony would crush MSFT. They have better interface design, fewer conflicting platform goals, and they'll put a PS3 in your living room for a fraction of what MSFT could.

Is that really that revolutionary?

The idea of having many processing units in a personal workstation is not new. They thought that Moore's law was going to fail years ago and predicted that by now we would all have massively parallel machines at home on our desks. Well it turned out that Moore's law didn't fail and most importantly that many of the software algorithms are not easily parallelizable. So what if I can have 100 cells at home in my workstation. I could run SETI, weather or some other kind of simulation but I couldn't really play my video games much faster or have a more responsive user interface if I ever install Longhorn. I just can't think of too many programs run on home user's machines that would benefit form a parallel architecture.

Now if the can be made very fast and have only a few (2-8) coupled together...well,as it was said, that is what a nice Opteron machine does anyway nowadays.

Re:Microsoft isn't supporting it? Who Cares?

don't give me crap about NT running on Alpha. It ran on 32bit version, and there was a early beta of W2k that ran 64bit native, but the Win32 API and everything else you use on your computer is and always has been x86-only

It's not crap; we produced release versions of our graphics software for Windows on x86, PowerPC, MIPS and Alpha at one point. Shipped some, too. We had machines for all four architectures (still have them, in fact, though the Alpha and PowerPC's are mothballed), development tools, and working Windows OS's on all of them, and they all ran Windows NT, approximately the same version. Perfect, definitely not -- but Windows under x86 isn't perfect either. It worked well, certainly no worse than the x86 versions. We still use one of the MIPS machines as a backup file server. It refuses to die.

Now, I'm no fan of Windows, but if you think MS couldn't port Windows to another architecture beyond x86, you're only fooling yourself. They can any time they want to, they have already, three times that I know of for certain, not counting whatever credit you want to give Windows CE ports, if any, and there you have it. For all I know there may have been ports to 68k archtectures... I wouldn't be in the least bit surprised.

You have to consider that MS has more money than anyone, and if they decide to go this route, there is no reason to think they cannot do it. I doubt there is any market force, including Sony and the largest governments in the world, that could put a serious roadblock in front of them in this arena.

Re:Microsoft isn't supporting it? Who Cares?

Microsoft is still dicking around with porting Windows to AMD64... a platform mostly compatable with x86. (don't give me crap about NT running on Alpha. It ran on 32bit version, and there was a early beta of W2k that ran 64bit native, but the Win32 API and everything else you use on your computer is and always has been x86-only)

There are two operating systems Microsoft have developed called Windows. DOS/Windows, the original one, was based on an x86 clone of CP/M that Microsoft bought. The first version, "Windows 1.0", was released in 1985. The last version, called "Windows Me", was released in 2000, IIRC. This OS was always x86-only, originally ran on archaic CPUs without memory protection and never supported full protected memory, symmetric multiprocessing or other (now) basic OS features.

The second OS developed by Microsoft that's marketed as Windows is Windows NT (now just called "Windows"). It was started in 1988, and never had any relation to DOS/Windows, except insofar as it can (to some extent) emulate it for compatibility reasons (including an x86 emulator on hardware that can't natively execute x86 code). Windows NT was developed on the MIPS platform, not the x86. The original plan had been to use the Intel i860 (an LIW architecture completely different from the x86) as the development platform, but the i860 hardware never met its promise, so MIPS was chosen instead.

The first version of Windows NT was released in 1993, and called "Windows NT 3.1" (3.1 was used for marketing reasons, since that was the latest version of DOS/Windows at the time). Like UNIX, it was mostly written in C, with assembly at the low level to handle hardware dependencies. At its release, Windows NT 3.1 ran on 32-bit MIPS (the development platform) and 32-bit x86 (the first port).

The second version of Windows NT (3.5) was released in 1994, and planned to add 64-bit Alpha (in a semi-crippled, 32-bit mode) and 32-bit PowerPC. However, IBM and Motorola ran into problems with the hardware (in part because of ongoing disagreements with Apple, who wanted to use their own, proprietary platform), so Windows NT 3.5 only added Alpha support. In 1995, after IBM and Motorola had managed to (mostly) sort out their problems (but with Apple declining to follow the IBM/Motorola PReP standard), the PowerPC port of Windows NT was completed, and released as version 3.51. At this point, the OS ran on MIPS, x86, Alpha and PowerPC.

In 1996, the user interface of Windows NT was upgraded to match the user interface of the popular 4.0 release of DOS/Windows (called Windows 95). Windows NT 4.0, which copied the user interface of DOS/Windows 4.0, ran on MIPS, x86, Alpha and PowerPC.

By the late 1990s, as Microsoft continued work on version 5.0 of Windows NT, the market had lost confidence in non-x86 systems for general-purpose PCs (apart from Apple Macs, which didn't follow the PReP standard, so couldn't run OSes ported to it, like AIX and Windows NT). As a result, Microsoft and the vendors of MIPS and PowerPC workstations agreed to cease development and marketing of NT 5.0 for those platforms. Windows NT 5.0 continued to be developed for the x86 and DEC Alpha architectures, into the beta releases.

DEC (which was taken over by Compaq) had continued to have hope for the Alpha as a general-purpose alternative to the x86, but financial difficulties led to the project being abandoned towards the end of the developent cycle for Windows NT 5.0 (marketed as "Windows 2000"). As a result, Windows NT 5.0, completed at the end of 1999, was the first version of NT that only ran on one platform (the x86).

A port of Windows NT 5.0 to the 64-bit Intel Itanium, including 64-bit versions of the Windows APIs (unlike the earlier Alpha port), was released in 2001, but only to select customers.

Windows NT 5.1 (marketed as "Windows XP) was also released in 2001, and again only ran on the x86, apart from another 64-bit limited release for Itanium (in 2002, IIRC).

Windows NT 5.2 (marketed as "Windows Se