What the Cell Isn't
No, it's not your next PC or server.
Information on the new cell processor from Sony and IBM has been somewhat available to the geekily inclined public for quite a while now. So, I was very surprised by the questions I was being asked after the Cell's more official debut at the ISSCC conference. And, I was even more surprised by some of the reports that came out of the conference. Reporters have been consistently trying to make the Cell a competitor to the venerable X86 architecture and the Pentium in particular.
For example, here is the sort of quote you might encounter:
"Analysts say that the difference [between the Pentium and the Cell] is likely to shake up the contemporary model of computing that Microsoft and Intel have fostered. [because] The new chip also has been designed to handle multiple operating systems and programs simultaneously."
The Cell is not a Pentium killer or a competitorit's a SIMD supercomputer, and just because it has an IBM Power processor in it, which has been demonstrated to run X86 code in an emulation mode, that's not the use or the future of this Cell.
Here's another one:
"Later this year, Intel and Advanced Micro Devices each plans to launch their own 'multicore' chips, which similarly up the number of commands that can be executed at once. IBM and Sun Microsystems already sell such multicore microprocessors, largely for business servers."
The Cell has an array of SIMD RISC processors, not multi-cell X86 processors. Nor will it be a super server chipit's just not built that way. It's built to process the same kind of data, which is highly predictable, not changing data like a GP CPU in desktop or a server deals with. Mix up the data stream and types and feed it to a Cell and you'll see it screech to a slow crawl. It costs 18 cycles any time a branch mis-predict that's not server-quality processing.
Is it your next workstation?
IBM Chairman, President, and Chief Executive Samuel Palmisano said that this year it'll sell a workstation containing Cell. That's true, it will be built at IBM's Boeblingen facility near Stuttgart (see Tech Watch, November 29). Such a system will be used for software development for the future Sony console. It's possible it may be used for scientific work, but special compilers and applications will have to be created to make use of it. That comes under the "someday" heading.
Is it a superomputer, on a chip?
Well yes, architecturally, it's exactly like the multi-processor SIMD vector machines that were built in the mid-eighties, and to some extent are still being built (only 4.2% of today's supercomputers are vector based). However, at 256 GFLOPS it's not, by itself, in the supercomputer class. (The slowest machine on the current list of the Top 500 supercomputers does 624 GFLOPS http://www.top500.org/lists/plists.p...5&M=06&Y=2004.
) But you can envision an array of Cells being made into a supercomputer, and given the expected low cost of this soon-to-be-mass-produced chip, that's entirely feasible, except that it might wrinkle the feathers of IBM's BlueGene group a bit.
It's not a Grid computer
The other popular misconception, which IBM and Sony are doing nothing to dispel, is the notion that the Cell will hook up to some type of grid (like Butterfly net) and either tap into more resources or allow millions of user to simultaneously play a game. Yes, online gaming can take advantage of the Butterfly net, and yes, IBM and Sony said they would use it. But it's not to share Cell resourcesit's to play against others just like any other online network.
Take a look at this quote from a story published by CNET:
"Cells can even roam over a network, allowing the processor to perform a type of distributed or grid computing, an increasingly popular enterprise technique in which demanding tasks are divvied up among a gang of networked computers."
Clearly, there are some major misunderstandings about what networked and Grid computing are all about. Grid is used for collaboration and uses the Internet, not the fastest data exchange network in the world and certainly not fast enough for distributed for real-time computer, the type of computing that epitomizes a game.
A Playstation 3 could borrow unused processing power from other consoles on a network, for example, to complete a demanding task such as delivering streaming video. Sony has done a masterful job at creating a science fiction scenario that ignorant journalists have latched onto. The only problem is some of those ignorant journalists have good memories, and in a year or two when the next-generation Playstation is delivered and it doesn't do all these suggested whiz-bang things, those same journalists are going to feel used, and then they will write searing stories about how Sony and IBM have failed to live up to the expectations and promise of the lauded Cell. This misdirected PR BS is going to haunt IBM and Sony. Lest we forget, when Sony introduced the Playstation 2 Emotion Engine it was promised to power a wide variety of consumer electronics, and so far it's failed to live up to that promise.
Is it your next TV?
Well, Toshiba has been quoted as saying it plans to incorporate it into high-end TVs. Given the crashing ASPs on TVs these days, it's not likely the Cell, in its current configuration, will show up inside a TV, high-end or otherwise. The vector processors could be useful for MPEG decoding/encoding, but you wouldn't need eight of them, so a stripped-down version, with maybe two SPEs and a Power CPU, might be made for high-end TVs sometime in 2006, maybe.
So what is a Cell?
It is an eight-IMD processor array with a powerful master controller and program execution GP CPU. Each SIMD has local memory, high-speed interconnects, and a segmented memory structure ideally suited for graphics and codec applications. It has a super-high-speed chip memory interface and an equally fast flexible I/O for the graphics processor. It is currently running Linux and could run other OS's on the GP CPU.
It will be a really great game console processor and general multimedia accelerator, and should cost out at a reasonable consumer price range.Jon Peddie