Monday, February 2, 2004 |
11:30 - "I know-- we'll dig our way out!" "...No, dig up, stupid!"
http://www.lostcircuits.com/cpu/prescott/
|
(top) |
I've been hearing underground info about this for some time now, and now the reports are released (so I can make mention of this without incriminating anybody).
Prescott, Intel's new CPU, is out. It's on a 90nm process instead of the old 130nm one, and... the performance gains over the old chips are negligible at best.
Why? Well, because Intel, in their infinite wisdom, decided that the P4's 20-stage pipeline wasn't enough... so they took it to 31 stages in this one. Any branch prediction failure, as occurs in any application except for SPEC2000 benchmarks, will lock up the CPU for up to 31 cycles.Which means they had to double all the caches, which makes Prescott's die size barely smaller than Northwood's. Which means: heat? Oh yeah... heat! Apparently Prescott doesn't use any less heat than older P4s, even though it's on the smaller process.
In other words, this is one embarrassing chip for Intel. As Ars Technica says:
Anand maps out the performance delta here while trying to suggest that Prescott's future will be bright. You can easily see what's going on here: the performance gap narrows (not widens!) as the two CPUs ramp up in speed between 2.8 and 3.2 GHz. Why this is a secret isn't entirely clear... some would call it the point of a core revision. But one thing is clear: as speed ramps up, you tend to see Prescott get closer and closer to being almost as fast as Northwood. It's only an assumption, although a logical one, that the trend will continue in a positive fashion, and at some point (say, 3.6GHz) Prescott will be shipping consistently faster than Northwood, presuming that you could clock Northwood that far.
In other words, Prescott is being released at a speed below its ideal debut point, which suggests that Intel is indeed having some serious problems with the CPU. The problem isn't the 90nm fab process, though. It looks as though Intel has this down. Rather, one must consider the 40+ million new transistors into the mix. Indeed, the problem with Prescott (almost sounds like a movie title) is heat, heat, heat. That's the oddity of the situation: usually a shrink in die size results in less heat, but Prescott has extra mojo on board in terms of the expanded L2 cache (now 1MB), some additional SSE logic (13 new instructions for SSE3), HyperThreading improvements, and more.
The word that I heard a couple of weeks ago, through the unnamed grapevine, was that Intel knew months ago that Prescott would be a dismal performer for what they're putting into it-- but contractual obligations had to be met, fabs were already built, and there was no backing out. It's no accident that P4s have remained pretty much at the same speed since June, after that flurry of speed-bumps that seemed to happen every week in the preceding months. The farm's been bet on Prescott, and Intel's legitimately worried that the cows have all died.
A lot of reviewers seem to be putting as positive a spin as they can on the news, but Prescott has pretty much been a laughingstock in insider circles for some time now. I know a few people who are watching the headlines today with no small amount of glee.
Something tells me AMD's going to take the uncontested lead in fairly short order now. ...At least as far as PC chips go.
|
|