Transmeta posts Q4 loss, shows $107 million deficit for ’04: “Transmeta has not exactly said it would stop selling standard, x86-based microprocessors. But the troubled company will move away from selling standard products and focus on IP, as speculated earlier this year”
At ISSCC this week, IBM, Sony and Toshiba presented some papers on the Cell chip. Sony plans to use Cell as the core of its Playstation 3 console. But the three companies are promoting Cell as a “supercomputer on a chip”, to be used in a wide range of consumer and communication products.
The news media has largely repeated IBM’s press releases with the same breathless claims. The Merc gushes in New chip called a threat to Intel: “The partners say the first Cell chips, which can simultaneously juggle multiple computing tasks, will have 10 times the processing power of comparable Intel chips. Eventually, the technology could pack the power of a supercomputer in a handheld device.”
Right. Maybe a handheld device with a heat sink the size of a toaster.
According to the folks at CNet:”While IBM didn’t disclose the exact heat statistics, some at ISSCC said it could run as hot as 130 watts…. Cell contains 234 million transistors and takes up 221 square millimeters in the 90-nanometer process. That’s about double the size of the 90-nanometer 3.6GHz Pentium 4.”
That’s a big, expensive chip, built in a new, low-yield process. It also contains a new proprietary Rambus memory interface, further inflating system cost. As CNet points out, that might not matter to Sony, selling Playstation at a loss, but it might severely limit use of Cell outside of the console market.
More worrisome for me is the implication that Cell requires an entirely new programming model to effectively make use of that power.
Cell is a System on Chip (SOC) design, containing 8 vector execution units, and control processor based on PowerPC. It supports multiple threads of execution, and is supposed to run at 4 GHz in IBM’s new 90 nanometer Silicon on Insulator (SOI) technology. Naturally, the developers claim it will have “10 times the performance of PC processors”.
OK, Cell is certainly an interesting architecture. But listening to the claims reminds me too much of the past 20 years I’ve spent working on multiprocessor systems. Often it’s much easier to build innovative hardware than to get good performance out of those systems. And any system that requires substantial changes in how programs are written is a long shot.
CNet reports on Sun’s next-generation Sparc processor: “Niagara, with eight cores each able to run four threads, can run a total of 32 threads in parallel.”
In the grand tradition of Burton Smith’s Tera and Cray X-1, Niagara uses cycle-by-cycle multi-threading to hide memory latency.
And like those machines, it requires a lot of parallel threads to feed the beast.
I don’t know if Sun will actually produce effective systems using this processor. And maybe users will not be able to exploit its full performance. But I’m encouraged to finally see some new processor architectures introduced after years of x86 induced drought.
Yesterday’s Mercury News has all but written off Transmeta, at one time the Darling of the Valley.
Transmeta — how a great idea, brilliant minds and big investors equaled a big flop.
The Merc expects Transmeta to quit the business of selling processors, after several years of unsuccessfully trying to compete with Intel.
Instead, they’ll focus on their IP business … and probably lay off half their employees.
But all is not lost. Their radical ideas will undoubtably live on in later, successful ventures.
“The most innovative companies rarely succeed,” said Paul Saffo, a director of the Institute for the Future. “But Silicon Valley is built on the rubble of failure, not on the spires of success. Their intellectual property will get recycled by the valley and continue to move it forward.”
I’m sure that’s a big comfort to all those workers about to get “recycled” in the valley.
Transmeta says that it is considering leaving the business of making x86 compatible processors, and becoming as an IP business instead.
This week, the Santa Clara, California, company said it now definitely plans to increase its licensing activities, and that it would “complete a critical evaluation of the economics of its current business model of designing, developing, and selling x86-compatible microprocessor products.”
Analysts expect Transmeta to announce job cuts at their January 21 conference call.
“The cuts could be as low as 100 or as high as 200 people,” Piper Jaffray analyst Richard Shannon said. Transmeta has about 325 employees, he said.