« "Reviewing" the Mac mini | Main | "The Obvious Guess Would Be 'Worse'" »

The Economist on the Cell

As is often the case, the most readable coverage of a technical subject can be found in the Economist -- in this case an article on the new Cell microprocessor from IBM, Sony, and Toshiba:

As its name suggests, the Cell chip is designed to be used in large numbers to do things that today's computers, most of which are primitive machines akin to unicellular life-forms, cannot. Each Cell has as its "nucleus" a microprocessor based on IBM's POWER architecture. This is the family of chips found inside Apple's Power Mac G5 computers and IBM's powerful business machines. The Cell's "cytoplasm" consists of eight "synergistic processing elements". These are independent processors that have a deliberately minimalist design in order, paradoxically, to maximise their performance.

A program running on a Cell consists of small chunks, each of which contains both programming instructions and associated data. These chunks can be assigned by the nucleus to particular synergistic processors inside its own Cell or, if it is deemed faster to do so, sent to another Cell instead. Software chunks running on one Cell can talk to chunks running on other Cells, and all have access to a shared main memory. Since chunks of software are able to roam around looking for the best place to be processed, the performance of a Cell-based machine can be increased by adding more Cells, or by connecting several Cell-based machines together.

All of this means that programs designed to run on Cell-based architecture should be able to fly along at blistering speeds—and will run ever faster as more Cells are made available. The prototype Cell being discussed this week runs at 256 gigaflops (a flop -- one "floating-point" operation per second -- is a measure of how fast a processor can perform the individual operations of digital arithmetic that all computing ultimately boils down to). A speed of 256 gigaflops is around ten times the performance of the chips found in the fastest desktop PCs today; the Cell is thus widely referred to as a "supercomputer on a chip", which is an exaggeration, but not much of one. On the top500.org list of the world's fastest computers, the bottom-ranked machine has a performance of 851 gigaflops. A machine based on only four Cell chips would easily outrank this...

Cell's debut will be in Sony's next-generation games console, the PlayStation 3, which is expected to contain four of the beasts.

If the PlayStation 3 does include four Cell processors, and if they run at 256 gigaflops, and if a PlayStation 3 were available today, it would place 387th on the Top 500 list. That's staggering.

After talking with my friend and colleague David Smith, I'm convinced the Cell has the potential -- if it lives up to the promises made for it -- to be an industry-changing event. Ray Kurzweil and many others have long argued that at some point, Moore's Law will continue through the use of highly parallel architectures, as opposed to continually increasing the clock speed and word length of today's microprocessors. Much evidence exists for this. Most recently, Apple's chief financial officer called a hypothetical PowerBook equipped with a G5 processor "the mother of all thermal challenges". The Cell addresses such challenges by providing high levels of performance using large numbers of efficient RISC-based processor cores. Instead of one very fast processor, how about eight that are moderately fast? And that's just on one chip.

I'm thinking through the larger implications of this. I'm sure they're not good for Microsoft. The open question is, for whom are they good?


TrackBack URL for this entry:



You are correct the Cell will not be good for Microsoft

This story is about the Cell, but we need a little history to set the stage.

IBM designed the IBM PC in early 80's in response to Apple II. Shortly after the release of the original IBM PC, IBM published the XT technical reference manual. This book contained schematics for building a complete PC, all the "magic addresses", and the source code literally printed out for the original BIOS. IBM then bought 8086 chips from Intel and MS DOS operating system from Microsoft.

Others discovered this book, built copies, the copies worked and the PC business was born. They all used Intel chips and Microsoft OS. The funny thing is IBM never complained. IBM so far as I know never sued anyone to try to stop them from knocking off the PC design. They even called them PC's or "clones"

Intel and MS went on to become pillars of the world economy. Microsoft was viewed as so important to American economic dominance that they were allowed to maintain their monopoly status even after being found guilty of monopoly practices.

IBM lost billions on the baby they created. I have never done business with IBM directly. Never bought any equipment from them or any software. I started buying PC's in about 1983 and have bought 1000's for use in automation systems. They all had Intel cpu's and Microsoft OS but never actually bought an actual IBM (I have seen a couple so I know they existed).

As I said IBM lost billions on their Direct PC business, however, I suspect IBM did make money on the PC. The PC replaced all the dumb terminals used in big business and big government. the rs232 link to these terminals was replaced with Ethernet. All these PC's made IBM servers and database software much more valuable. The fortune 500 now all rely on IBM technology to run their vast empires and that has benefited IBM and the world.

I would really like to know if any of that was a planned out strategy or "just happened".

Fast forward to 2006

IBM has a new plan. This time it is an actual plan. I do not have any inside information, but when I add 2+2 this is what I get.

IBM has recently sold its PC division
IBM has announced it is (will) open source the PowerPC CPU -- www.power.org
the Cell Processor contains one PPC core plus 8 math accelerators called SPE's (synergistic processing elements)

IBM will open source the entire Cell Processor technology. This time IBM will open source the System Design, BIOS, Cpu, and OS.

The Cell Processor design makes much more rational use of the transistor budget than the X86. By FLOPs / transistors or flops/watt, or Flops /$ the cell will beat the X86 by 10 to 100 times.

So any one cell chip will outperform the X86 by a large margin and the Cell chips are designed from the ground up to work in groups.

The Cell design includes an I/O subsystem called FlexIO. FlexIO is very very fast. FlexIO allows multiple cell chips to be directly connected with no "glue chips". FlexIO will replace all the legacy and current PC IO standards such as PCI AGP ATA USB 1394 serial parallel with one standard. This will eliminate much of the clutter, confusion, and complexity of the PC. FlexIO will greatly simplify the OS.

There has been talk about the Cell using a more straightforward, understandable approach to software development in which direct control over the system is returned to the programmer. The multiple layers of abstraction found in X86 systems were designed for a time when memory and other system resources were scarce and expensive. The X86 is so complicated that nobody understands it. there are benchmark examples of simple X86 code running several times faster when compiled with Intel compilers than when compiled with Microsoft or Linux compilers. It is likely that complex abstraction schemes were always part of X86 Wintel strategy to discourage competition. Well it backfired.

Of course the Cell comes out of the gate in a large volume consumer appliance with a 15 year track record - the Sony Playstation. The Cell instantly has economy of scale and media attention.

Their will be a new enthusiasm for computers, everyone has a chance of being a part again. Opportunities will be everywhere. Existing applications will run 50 times faster. New applications previously not practical will appear. AI - real AI will be the near term target

The entire world will defect en-masse. There will be a herd mentality, nobody will want to be left behind, People and business will be tripping over themselves to get their hands on anything cell.

Intel will take advantage of the fact that the cell cpu design is open source and will take a license. What does Intel care what kind of chips they make so long as they are selling. I think the Cell chips will be in short supply for years and will command prices comparable to X86 cpu's - Intel will have its pride injured but not the bottom line. Intel might be able to fire most of their CPU designers and just contribute their share to power.org to gradually evolve the cell.

Microsoft will not be so fortunate.

IBM will do what it has always done. Supply and support the large scale IT systems needed by the fortune 500 and world's governments. IBM will use the cell as part of that strategy. It does not matter who builds the hardware. IBM will show big organizations how to use it to best advantage. This strategy knocks Wintel completely off balance and keeps them far away from IBM's customers.

IBM will give Sony a running start with the Playstation. I figure an open source cell design will be a year or so after the playstation3 release. I think the Playstation will be locked down and will not be progammable.

I have put up www.cellsupercomputer.com I plan to build cell systems and write cell software. I am running a poll to see what people think about the cell - go vote. If you are interested there is contact info at my site.

Post a comment