Cell computer details announced. Don't hold your breath!

IBM, Sony Corp., Sony Computer Entertainment, and Toshiba Corp. said today that the companies had "powered on" the first workstation using silicon designed around the collaboratively-designed "Cell" processor, which will begin pilot production next year.

According to the Wall Street Journal, a 2002 patent described the idea of software "cells" using chunks of data that are bundled with programming code needed to process the data. The Cell computer announced today uses a dispatch computer and eight execution engines, each of which can simultaneously work on its own data. To me, the software cell concept fits the fifteen-year old definition of an "object", also defined as a package of data and processing instructions. So, let's think about the Cell computer as an architecture designed to execute object-oriented (OO) code extremely efficiently. Today's computer architectures do not run pure OO code very effectively -- the overhead costs of managing the tens of thousands of objects found in a decent sized enterprise application result in 30-second response times on even the fastest server processors available today, my sources say. So enterprise C++ and Java oriented application designs are often truncated to be sorta object-oriented in a very linear, transaction-oriented fashion that maps better to today's leading computer architectures.

Enter the Cell processor. IBM says the one-rack Cell processor-based workstation demonstrated today will perform at the 16 teraflop level -- something that required a multi-million dollar Cray supercomputer a decade ago. The IBM Power architecture-based chips are scalable. Sony invested $325 million into IBM fabrication technology to use the Cell processor in its next-generation Playstation 3 console, expected in a year.


My first reaction is that the workstation using the 8-way Cell processor is not what will appear in the placated 3 -- not by a long shot. It will be very hard to drive 8 cores at 4 GHz in a production manufacturing product in 2005. If Intel and IBM cannot do 4 GHz on mass market (or even specialized) microprocessors like Prescott, what am I missing about clock/heat issues?
The cell computer die size of 221 mm on 90 nm fabrication unveiled by IBM is absolutely HUGE and therefore hard to make -- costly too. Note that Intel's Prescott is 122 mm at 90 nm, almost 100 mm smaller per side. The cell processor chips will cost a fortune to make! Too many defects in chips that big lead to lousy yields per wafer, the basic economic measure of semiconductor production. Expensive fabrication is OK on, say, an Itanium or Power processor retailing for $3,000 for an enterprise big-iron server, but horrific on a CPU for a $150 retail Sony game console. I really doubt that Sony is going to use anything like 8-way, 221 mm dies and a 4 GHz clock in their Playstation 3 design -- the technology risk of getting the placated 3 out by early 2006 with these specs is huge. Expect a (very) slimmed down version of what IBM showed today in your next Playstation.

What Sony can do with the Cell processor is combine the functions of two complicated, expensive game console parts -- the central processor and the graphics processor unit (GPU). It should be clear now why nVidia walked away from Sony last year. The Cell processor will do the work of nVidia GPU silicon, saving chip and integration costs. In addition, the on-chip I/O capabilities of the Cell processor should reduce the chip count compared with traditional (i.e., Microsoft Xbox) designs.

What am I missing to make this pipe dream a reality? Software development tools! The mature ecosystem around any of today's (surviving) microprocessor architectures allows thousands of programmers to churn out the applications that run the world. Cell processors will have to build that ecosystem. While the learning curve is steep, the business costs and risks are highest. Say you want to develop a game for the Playstation 3? Expect to invest millions in training, development tool R&D, and the costs of porting your existing intellectual property to the new Cell processor architecture. That suggests Microsoft is betting its next generation Xbox, built with traditional software tools, will be more attractive to game developers than than the sexier Cell processor and a dearth of tools -- at least circa 2006.

All is not from scratch, however. There are numerous OO language compilers available through open source and IBM which can generate the pure OO code that will make cells hum. IBM was correct to demonstrate the first Cell processor prototype on a workstation application, as opposed to a transaction-based server application. Lots of scientific code has already been "parallelized" to run on multiple machines at once. So, the second analogy for the Cell processor is an 8-way computer "grid" on a chip. This will be attractive in visualization and desktop- and departmental number crunching -- everything from computer-aided design to life sciences to financial derivatives and Monte Carlo simulations.

This grey-haired writer remembers the 1960s announcement of the System/360, an architecture first that later became the mainframe business bonanza that generated IBM billions. If IBM and its partners do a tenth as well as the System/360, they will have hit a home run. I think they can. But the payoff of the Cell processor architecture will be measured in decades, not in 2006 market share.

Peter S. Kastner

Intel's 90 nm Prescott silicon specs.