Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
GFLPraxis said:
I still got the basic architecture right (PowerPC core + several Cell units = Cell chip). You're probably right that it won't run natively though, but it should be fairly easy to port. Sorta like the special edition Jaguar they released for the G5, remember? While it's not the same thing, they might release a cell-enabled version of OS X just for the Cell systems.

That's true, but one distinction that should be stressed is that the PowerPC core is not an "option." Your post made it seem like the PowerPC wasn't integrated into the Cell, when the PPC core itself is an integral part of the Cell.

So if Apple were to use the Cell, it wouldn't run alongside the PPC 970 already in the G5 machines - the Cell would probably replace the G5 chips entirely. For one thing, the Cell requires very high bandwidth, specialized memory from Rambus (that's not even on the market yet). The G5's memory architecture, as robust as it is, would not be compatible with the Cell at all.

On that note, this is actually an argument against the Apple adopting the Cell. Relying on proprietary Rambus memory would make things VERY expensive, as history has shown. It was just a few years ago that some PC makers were using Rambus memory for some versions of the P4 (because Intel made those Pentiums compatible with only Rambus memory). That made adding memory to those PCs two or three times more expensive than using the DDR memory you find in every PC and Mac today.

Disappointing that Sony would work with a company with such an questionable and convoluted corporate history as Rambus....
 
Didn't I read an article that Rambus is changing their business practices so that their next memory technology won't go down crashing that was RDRAM memory.
 
Lacero said:
Didn't I read an article that Rambus is changing their business practices so that their next memory technology won't go down crashing that was RDRAM memory.

Yeah, I saw that on Slashdot...
 
Lacero said:
Didn't I read an article that Rambus is changing their business practices so that their next memory technology won't go down crashing that was RDRAM memory.

The proof will be in the pudding, as they say. Or, we won't know until it actually happens. Let's hope that partnering with Rambus won't be a huge mistake. It's possible Sony partnered Rambus for legacy reasons - I think the PS2 also uses specialized Rambus memory.
 
Cell-Powered Macs, Pros and Cons

I've listed some pros and cons to Apple adopting the Cell for Macs

Pro:

  • PC-busting performance (30 gigaflops for a dual 2.0 GHz PowerMac G5 vs 256 gigaflops for a single Cell)
  • Economies of scale (they will be pumping out tens of millions of Cells each year, compared to the few million 970s that are used in G5 machines)
  • Ability to run multiple OSs simultaneously (although it's not clear how this will be implemented yet)
  • Scalable architecture (the first Cell will have one PPC core and 8 SPEs....imagine a 2nd or 3rd generation Cell with 4 PPCs and 32 SPEs....or imagine a PowerMac that can do 1000 gigaflops....1 teraflops of power on your desktop!)
  • Marketing heaven (imagine the PR coup Apple would net at announcing a "Cellular" PowerMac)


Cons:

  • New architecture (unknown how feasible it will be to port OS X to the Cell, even with the Power instruction core. Also, Cell is not a proven technology)
  • Rambus memory required (sure, 512 MB in a PS3 may be more than enough for a console, but how expensive will it be to stick 8 GB in a Mac?)
  • Compatability (even if OS X is ported to the Cell, is that enough to guarantee 99% compatability with existing software and hardware?)
  • Timing (unknown when the Cell will see the light of day)
 
vitaboy said:
I've listed some pros and cons to Apple adopting the Cell for Macs

Pro:

  • PC-busting performance (30 gigaflops for a dual 2.0 GHz PowerMac G5 vs 256 gigaflops for a single Cell)
  • Economies of scale (they will be pumping out tens of millions of Cells each year, compared to the few million 970s that are used in G5 machines)
  • Ability to run multiple OSs simultaneously (although it's not clear how this will be implemented yet)
  • Scalable architecture (the first Cell will have one PPC core and 8 SPEs....imagine a 2nd or 3rd generation Cell with 4 PPCs and 32 SPEs....or imagine a PowerMac that can do 1000 gigaflops....1 teraflops of power on your desktop!)
  • Marketing heaven (imagine the PR coup Apple would net at announcing a "Cellular" PowerMac)


Cons:

  • New architecture (unknown how feasible it will be to port OS X to the Cell, even with the Power instruction core. Also, Cell is not a proven technology)
  • Rambus memory required (sure, 512 MB in a PS3 may be more than enough for a console, but how expensive will it be to stick 8 GB in a Mac?)
  • Compatability (even if OS X is ported to the Cell, is that enough to guarantee 99% compatability with existing software and hardware?)
  • Timing (unknown when the Cell will see the light of day)

I think I can solve the 1st and 3rd con you list.

OS X can be modified (without completely re-optimizing it for Cell) to run on the PowerPC core for the cell processor. Of course, you don't get the speed boost except on specially optimized programs, though. But it guarantees compatibility with anything that will already run on a G5, and allows it to run Cell-specific programs at fast speed.

OR

Emulation. Design a Cell-centric version of OS X, and use emulation to run all older programs (heck, even Windows programs). If Cell is so vastly more powerful than current processors, heck, a Cell Mac could probably emulate faster than most x86 and current PPC processors RUN! As a result, a Cell-based OS X, using emulation, could run almost all current Mac programs with near-perfect compatability by emulating a G5, and specially designed Cell programs would run with an added boost. THat'd probably work better than the first.

Sorta like the 68k/PPC transition IIRC.

Furthermore, a Cell PowerMac could probably run VPC so ridiculously fast that it'd probably run faster than an actual Windows PC. :D
 
GFLPraxis said:
Cell_Arch.gif


See the bit labeled "Processor Unit (PU)"? That's where the G5 goes (as it says even in the picture). The APUs are the Cells.

No, the grey box is a Cell, a Cell consists of:

1) 1 SMT (two thread) capable PowerPC Processing Core
- i.e., a single core POWER5 derivative, maybe cut down further because it isn't using much die size from the available die photos
2) 8 APUs, i.e., vector processors
3) A lot of on-chip bandwidth between these components!
 
benpatient said:
here's a picture of a CELL chip with 8 SPEs and one 64-bit PPE (capable of 2 instructions per clock)...thus 10 simultaneous (max) operations per clock. That's theoretical peak, of course...which is not real life, but the potential is there.

details-on-cell-20050207005224717-000.jpg


10 operations per clock X 4 billion clocks per second=40 billion operations per second.

So something isn't quite being added up properly from the initial specs they've give us.

Vector operations. One instruction acts on multiple numbers. If the patent application as explained by Blachford (http://www.blachford.info/computer/Cells/Cell0.html) is how it's actually implimented, then each SPE instruction will actually operate on 4 numbers, hence 4 operations.

so ( 2 operations + 8 instructions * 4 operations ) * 4 billion clocks per second = 136 billion operations per second.

But if it has a multiply-add instruction, and you pretend that's all it does, then you get 272 billion operations per second, which is roughly the cited 256 GFLOP.

Correction: from http://www.electronicsweekly.com/ar...iChannelID=114&liFlavourID=1&sSearch=&nPage=1 Each 2.5x5.81mm SPE can issue two instructions per cycle to seven execution units using two pipelines. There is no out of order execution.
Thus, you don't to assume multiply-add operations to get to around 256 GFLOP

Correction: I works out somehow. I'm not a chip designer, I don't know how it all works.
 
maya said:
iPod Shuffle

iPod mini

iPod

iPod Photo

iPod Video a.k.a. iPod Cell. ;) :)



It had already been stated that iPod didn't have a powerful enough processor in the current and past generations and the Cell processor is scalable, so I think this will become a reality. :)

The Cell processor is too large for the current iPod enclosure. Maybe if they move up to say, a 7" LCD or so, the iPod Cell would be a strong seller as a replacement for portable DVD players (no need to lug around all the DVDs!)
 
vitaboy said:
Consider this: a dual 2.0 GHz Xserve G5 can do 30 gigaflops.

A single Cell chip can do 256 gigaflops.

All I can say is, Apple better be finding a way to utilize the Cell in its G5s. Gigaflops won't speed up Microsoft Word, but can you imagine 10x faster iMovies exports, Motion effects, and other computationally intensive operations than a top-of-the-line PowerMac today? And with Tiger pushing off virtually all the rendering of the interface to the GPU, you'd be looking at major, major performance in OS X from top to bottom.

Its much faster than a G5. Where did you find the 256 gigaflops number? Based on a news release in November, IBM stated: "The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second."

I expect that IBM will be using this in their future cluster supercomputers. I don't think that Apple will use the Cell, but they will use the PowerPC core from it, and they will likely benefit from the rest of the design.
 
When we migrated from 68K to PPC, I remember all sorts of tales about how emulation would allow computers to pretend to be other types and work faster than the real thing. We were told in glowing terms how RISC computers would be so fast that they could pretend to be anything the software writer wanted.

At the time it seemed entirely plausible, but the reality never quite lived up to the expectations.

This all seems totally plausible too, probably rather more plausible than before. Can we be sure that the reality will be close to the promise on this occasion ?
 
nek said:
Its much faster than a G5. Where did you find the 256 gigaflops number? Based on a news release in November, IBM stated: "The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second."

I expect that IBM will be using this in their future cluster supercomputers. I don't think that Apple will use the Cell, but they will use the PowerPC core from it, and they will likely benefit from the rest of the design.

One rack is not equal to one chip. Basic math :rolleyes:

256 x 64 = approx. 16 teraflops. So one rack has 64 of these processors.

Actually, a Cell processor can have as many cell units as needed (these figures are for one with eight cell units and a PowerPC PU). So if you made a big one...

The Cell processor is too large for the current iPod enclosure. Maybe if they move up to say, a 7" LCD or so, the iPod Cell would be a strong seller as a replacement for portable DVD players (no need to lug around all the DVDs!)

A slightly larger enclosure just big enough to actually see the video should do it easily. We don't need all 8 APUs, one should easily do the trick for an iPod video.
 
AlanAudio said:
When we migrated from 68K to PPC, I remember all sorts of tales about how emulation would allow computers to pretend to be other types and work faster than the real thing. We were told in glowing terms how RISC computers would be so fast that they could pretend to be anything the software writer wanted.

At the time it seemed entirely plausible, but the reality never quite lived up to the expectations.

This all seems totally plausible too, probably rather more plausible than before. Can we be sure that the reality will be close to the promise on this occasion ?

Because we now HAVE emulation, and we SEE how it works.

Look at Virtual PC...we have a visible example of the type of performance you get.

Now, if you multiply that speed by 10x...you get some VERY useful emulation speeds.

So the only question is, will Cell perform as good as hyped? If so, then the emulation should be ridiculously fast.
 
Cell should be G6 AND the death of VPC

I found through the Altivec mail list this Sony news announcement site.

For me, the most important paragraph found in the details doc at the bottom ( open the word doc ) is the following:


• Contains 64-bit Power ArchitectureTM with VMX that is a dual thread SMT design – views system memory as a 10-way coherent threaded machine

Since the news point out that a target for this chip is for embedded applications, I think we found our PowerBook proc.

I also have recently read more on the OS independence. Essentially, the cell can also emulate in hardware an instruction set. Couple that with the ability to run multiple OSs at once, and you now have a single chip that can run PPC and x86 instructions simultaneously.

Should be fun!!

Max.
 
nek said:
Its much faster than a G5. Where did you find the 256 gigaflops number? Based on a news release in November, IBM stated: "The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second."

I expect that IBM will be using this in their future cluster supercomputers. I don't think that Apple will use the Cell, but they will use the PowerPC core from it, and they will likely benefit from the rest of the design.

It's in the News.com article:

PlayStation 3 chip has split personality

Cell can process 256 billion calculations per second (256 gigaflops), falling a wee bit short of marketing hyperbole calling it a "supercomputer on a chip." The slowest machine on the current list of the Top 500 supercomputers can do 851 gigaflops.

Where does IBM get the 16 teraflops figure? "The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second."

The biggest standard-sized rack you can get is 48U or enough to stack 48 XServes on top of each other.

48 * 256 gigaflops = 12,288 gigaflops or about 12 teraflops.

Or if you had a dual processor Cell server, you would have

48 * 2 * 256 gigaflops = 24,476 gigaflops or 24 teraflops

But you have racks that are 24U, 25U, 40U, etc. Apple uses a 42U rack on its Xserve page. So IBM could have used whatever size rack to come up with the 16 teraflops number, but here's the money line:

A single rack of Cell-powered single processor 1U servers would theoretically have the same processing power as the hundreds of Xserves that make up System X at Virginia Tech.

Also, there is no way Apple can "use the PowerPC core from [the Cell]."

The PPC core is integral to the Cell's design. It is not "separate" from the Cell, as seems to be the misconception that several people have stated in this thread. You might as well say that Apple can use the PowerPC core from the G5 and strip off the Altivec unit (which is not an option when ordering G5s from IBM).

Again: the Cell is 1 PowerPC core + 8 SPE ("Altivec-like") cores. These 9 cores make up one chip that is called the Cell. The PPC core by itself is useless for Apple's purposes - you wouldn't have an FPU nor Altivec. The last time we saw a chip without an FPU was, hmmm, the Motorola 68000 or the Intel 286. If Apple uses the Cell, it needs to commit to the entire chip.
 
When the term was originally developed, a 'supercomputer' was a computer that did over a billion calculations per second.

That's why Apple even calls the G4 a "supercomputer on a chip"- even the older 1 GHz eMac can hit 7 gflops.

So they can't deny that it's a 'supercomputer on a chip' ;)
 
vitaboy said:
Also, there is no way Apple can "use the PowerPC core from [the Cell]."

The PPC core is integral to the Cell's design. It is not "separate" from the Cell, as seems to be the misconception that several people have stated in this thread. You might as well say that Apple can use the PowerPC core from the G5 and strip off the Altivec unit (which is not an option when ordering G5s from IBM).

Again: the Cell is 1 PowerPC core + 8 SPE ("Altivec-like") cores. These 9 cores make up one chip that is called the Cell.

Correct me if I am wrong, but if you have a dual core processor, single-threaded apps or OSes will only use the first core, correct? What would be the difficulty in only using one core (the PPC one)?

Either way, emulation would be the best solution. Make a Cell-optimized OS X and add PPC emulation...
 
GFLPraxis said:
Because we now HAVE emulation, and we SEE how it works.

Look at Virtual PC...we have a visible example of the type of performance you get.

Now, if you multiply that speed by 10x...you get some VERY useful emulation speeds.

So the only question is, will Cell perform as good as hyped? If so, then the emulation should be ridiculously fast.

But that was really my question.

We were previously sold the idea of emulation because PPC chips were going to be so much faster than CISC chips and a PC could be emulated.

In reality, the CISC chips got faster and for a while the PPC ones didn't.

Virtual PC certainly emulates a PC, but it does it at a rather low speed.

Chips running 10x faster sound fantastic right now, but I just get this Deja Vu feeling.

I just hope I'm wrong.
 
GFLPraxis said:
Correct me if I am wrong, but if you have a dual core processor, single-threaded apps or OSes will only use the first core, correct? What would be the difficulty in only using one core (the PPC one)?

Either way, emulation would be the best solution. Make a Cell-optimized OS X and add PPC emulation...
They'll only use one core - not necessarily the first one, since the operating system scheduler could be running some system process on core 1 and assign the single-threaded application to core 2. There isn't really any difficulty with using one core if we assume that the operating system is Mac OS X - it intelligently allocates applications to processors/cores, thus preventing idle resources.
 
wrldwzrd89 said:
They'll only use one core - not necessarily the first one, since the operating system scheduler could be running some system process on core 1 and assign the single-threaded application to core 2. There isn't really any difficulty with using one core if we assume that the operating system is Mac OS X - it intelligently allocates applications to processors/cores, thus preventing idle resources.

so with that, couldn't Apple make a special version of OS X (like the 64-bit extensions version of Jaguar for G5s) that will use the PowerPC core for older PPC apps, guaranteeing compatability? (aka no emulation required)
 
vitaboy said:
Where does IBM get the 16 teraflops figure? "The companies expect that a one rack Cell processor-based workstation will reach a performance of 16 teraflops or trillions of floating point calculations per second."

As said before, all you have to do is cram 64 cells into one box and you have 16 teraflops. It's very concievable that IBM would do something like this, especially with the cell.
 
AlanAudio said:
But that was really my question.

We were previously sold the idea of emulation because PPC chips were going to be so much faster than CISC chips and a PC could be emulated.

In reality, the CISC chips got faster and for a while the PPC ones didn't.

Virtual PC certainly emulates a PC, but it does it at a rather low speed.

Chips running 10x faster sound fantastic right now, but I just get this Deja Vu feeling.

I just hope I'm wrong.

Yeah, you have a point. But at the point we're at now, PPC chips are barely behind x86 chips. Even a 2x speed boost would be a big deal, and a 3x or 4x boost would be excellent. A 5x boost or bigger would virtually guarantee that anyone doing heavy work would use a Mac, and would also see a large surge in Mac-only games. That would utterly CEMENT the scientific users (UNIX compatabilitiy + far better performance) and video editors (who are mostly Mac already, this would just pull the few still on Windows) in the Mac platform, and begin to attract the gamers (the only market Apple has failed to touch).

It'd be a huge deal for Apple.
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.