John Crawford, a senior fellow at Intel, presented many of Intel's stunning insights into the performance per watt and multi-core areas during a speech given to researchers at PARC yesterday.
He told a great story, and, for awhile, the PARC crowd seemed to buy it. Then, however, we asked why Intel had in fact been so far behind the competition in both marketing and product delivery on these fronts. Didn't Intel think it could push GHz for a couple more years?
"Why were we last?" Crawford said, repeating our question. "Why were we last? Why were we last?"
This mantra started to make everyone uncomfortable.
"IBM certainly beat the market to the punch," he said.
Crawford, having not come close to addressing our question, moved on to the next topic.
Intel actually supplies its own evidence to show how far behind it was on the power consumption and multi-core fronts. It spent most of the sessions at IDF talking up tricks done to combine instructions, tweak the front side bus speed higher, keep idle transistors quiet and improve memory sharing in an effort to boost the performance of an aging architecture. Why promote the tricks instead of the actual architecture unless that's all you've got?
It may seem silly to talking about an aging architecture given that Intel just revealed the new "Core" designs for Woodcrest (server), Conroe (desktop) and Merom (notebook). Intel also talked up more sophisticated future multi-core designs.
But, as much as Intel likes to downplay things like not having an integrated memory controller and not putting multiple cores on the same die, the fact remains that this is what everybody else is doing. Intel is still stuck in the past.
For the moment, Intel seems to have narrowed the performance gap with AMD and others. Although, it will start to see problems crop up again in the four-core and beyond era due to the issues mentioned above, particularly the lack of an integrated memory controller.
"At some point in the future, we will have an integrated memory controller," Crawford confessed. "That's something we are wrestling with."
(Intel could apparently use some help on the Itanic front as well. "Our challenge there is to maybe catch up on some of the frequency angles in that product line," Crawford said. Although, he did hold out hope that Intel could eventually play in the high-end chip market. "I suppose that some day we will need a 128-bit architecture and then maybe the game could open up again.")
Intel too faces issues as it moves to FB-DIMMs, which provide some nice benefits but also add about 6 extra watts each to a server board via an AMB (Advanced Memory Buffer) . As usual, Intel has a concrete plan for dealing with these issues.
"Intel is working really hard on that," Crawford said. "NEC has an FB-DIMM that is down to 5 watts. Maybe then we can get it down to 4."
(Standard, registered DIMMs eat up about 3 watts.)
This is the same kind of language Intel once used to address the heat issues of its chip. It would point to partners working on new cooling technology and publicly pray that someone discovered something to help the situation.
Closing on a positive note though, Intel has managed to jump on one power-consumption theme at the same time as rivals. Both Sun and Intel have started calling for better, more expensive power supplies.
In a rush to buy the cheapest supplies possible from China, vendors have more or less gone too far. Power supplies that cost just a few bucks extra deliver huge gains in overall energy savings.
Currently, customers have to watch as about 52 per cent of their power evaporates between the wall socket and the actual machine. By 2010, Intel expects major improvements in power delivery to the point where 85 per cent of the energy gets to the box. ®
Sponsored: Webcast: Simplify data protection on AWS