For those of you not aware of the history, Tukwila was the next major slab of the IPF concept. All of the existing processors to date have been more or less variations on a fairly conventional theme, at least conventional for Itanium. Tukwila was going to change that, and not be too subtle about the transition.
The chip was planned to have it all, coupled with a bunch of things that most people didn't even contemplate. Eight cores? Sure, no problem, and toss one in for a spare while you are at it. CSI links? Sure, throw in more than enough to differentiate it from the desktop parts. FB-DIMMs? No problem, how many do you want, with eight cores you need a lot. Toss it on a slab of silicon, bake until late 2007 to give Montvale its day in the sun, and you have Tukwila.
Throw in the Alpha designers (ADT) and you had a winner. Rabid fanbois could not wait for it to come along, the spec sheets proved it would obliterate anything foolish enough to attempt a challenge. Wait until 2007, we were told, you'll see. And, to be honest, I really did want to see what these guys whipped up, if they came close to the goal, it would have been interesting. Sun's Niagara may be a much lower goal to shoot for, but it is silicon now, and will be a product soon.
Then, with little warning, Intel knifed it and replaced it with a modified Montvale core - let's call it M++. No reason was given, and to date none have surfaced, but of course, speculation has run wild. Either way, M++ is looking to be a yawner. The team, probably one of the finest CPU design teams in existence, did not take the killing as a good thing. We know that several went down the road to AMD, CVs in hand. You could call this a bad week for IPF.
But the questions remains, is it a blessing in disguise? I can see two plausible scenarios here, with little middle ground.
On a positive note. Intel has been getting kicked up and down by the reputable press of late. Management made some really hard decisions that will mean a very hard 18 months, but it seems to have done the right things for the right reasons. Projects that were not panning out were cut, money which was slated to be peed away can be saved. Good calls mostly, even if Intel only had bad options to consider. It is easy to pick from three good choices.
Intel has been canning unworkable things lately, and is taking a lot of lumps for that. It took a long hard look at Tukwila and saw that for technical reasons, it wasn't going to do what it thought it would. Perhaps the die size was too large, or it ran hot, or whatever. Maybe aggregate performance was less than a four core Montvale. Whatever the case, it saw something was wrong, and rather than beating a dying horse until it died, then beating it more while burning piles of cash for warmth, it shot the beast and sent it to the knacker's yard.
Instead, of empty promises, Intel green-lighted M++ and away the teamwent. The customer would have a solid performance bump, CSI and FBD numbers would all remain the same, and everyone would go home happy. The teams that were working on the 'failed' part would be reconstituted and put on the next core, Son of Tukwila (SoT) or Project Z, and it would be everything that Tukwila might have been, or at least the next fanboi fantasy. Hearty pats on the back all around.
On the negative side, let's start with M++. Why would it can a core that was the future? Imagine that the ADT people were 'all that' and some more, which by all accounts they were. Let's say they were producing a core that was on time, on budget, and did what it was supposed to. Why would Intel summarily can it?
Underachievers, Underperformers, Overpricers
That one is simple, money. IPF is a chronic underachiever as an architecture. It was way way overpromised, overpriced and it underdelivered in a spectacular fashion. By the time Intel delivered a chip that wasn't risible, the ecosystem would have shrunk. OEMs had moved on, and wagons were being circled. There were two major vendors left, HP and SGI and the others were in the category of 'other' or 'rounding error' depending on the ambient temperature more than anything else.
When Madison hit, its performance didn't astound, but it didn't embarrass. Montecito, due in about a year, will probably be a solid step up in every way. It should give Power 5 a solid run for the money, but IBM may have Power 5+ out by then, and may still be ahead by a nose. The numbers I have heard say Monty will at least catch Power 5 in all benchmarks including TPM-C.
But who really cares? Is the Itanium market performance sensitive right now? When you buy an Altix, do you care about raw numbers, or do you care about vendor stability, cost, tools, and support? Doubling speed won't affect those parameters, and right now, the big questions are all around those things. I don't think faster will buy them nearly as much as pundits think.
The core constituency of IPF vendors is interesting. SGI is doing some remarkable things with system scalability and interconnects. It has some very interesting products in the pipeline from the desktop to the ten thousand plus CPU arena, and people rarely say anything bad about the products.
HP, on the other hand, sells the vast majority of the IPF chips, and is not all that solid at the high end. With the recent roadmap changes, it has moved the debate to the point where it needs to prove to clients it will be there, and its actions of late speak loudly, and badly for IPF.
SGI is watching
This has not been lost on Intel, and a few weeks ago, word us that it shifted priority on Itanium from HP to SGI. Then HP gave up the IPF engineering teams, and Tukwila was cancelled.
So Intel has the costs and responsibilities for IPF chips now. It crafts the roadmaps, it forges the chips. HP is stuck with those chips and rumour has it the contracts say that HP must buy until 2008, and Intel must produce until 2008. What started out as a mutual love-in has turned into a bitter divorce.
Intel is faced with a partner that doesn't really want to be in the market that Intel wants it to be in. It is forced to make chips so HP can fulfil its obligations. Deviation means a lot of lawyers buy big Audis with the overtime they will get. What is poor Intel to do?
One answer is that you you fulfil your obligations on paper and enforce the letter of the contract, not the spirit. Screw me will you? Take that. In the mean time, Intel can put a small crew on M++, basically a dumb die shrink with a few more cores, and put the brains on the real 'next big thing', Project Z. Funding a new core with a projected one year lifespan and an apathetic partner would not be a wise use of money.
If you can't put two and two together and come up with four, this would mean that both parties are fighting to see which one can put a knife in the back of IPF first. If this is why M++ came into being, it also shows that Intel management is making a good call. From my vantage point, it comes out looking pretty smart either way.
Barring official sources from Intel actually telling us what is going on, that judgement needs to be deferred. While we are currently leaning toward the doomsday scenario, we could be convinced otherwise. Pete Bannon leaving a few months ago does not add to my optimism, and the reported brain drain from reliable sources is not confidence inspiring either.
One thing that Intel desperately needs to do, and do now, is to assure partners and customers. We hear that SGI is scared white right now, the purple colour drained from their cheeks. HP appears too busy plotting the next colour iPod by Apple to sell, but we assume at least a few people there are not ecstatic. Intel needs to show everyone with a stake in the IPF infrastructure that there is a plan, and a solid plan at that, beyond M++. There were two named projects beyond Tukwila on the roadmaps, Sot and SoT2, are they still there?
If there is a confidence inspiring roadmap that makes the rounds over the next few days, we are ready to believe everything is well. But barring that, we fear the worst. µ
Sign up for INQbot – a weekly roundup of the best from the INQ