The Inquirer had an article the other day reporting that one consulting firm, Insight 64, thinks AMD is being downright negligent in not talking about its 90nm Winchester chips.
With all due respect, we think AMD has very good reason for staying quiet now and is quite wise in doing so, for now.
This article will explain why, again, with some new supporting information, and in the end, some new developments that ought to eventually be very good news for AMD fans and give a good indicator as to when the best time to go to Hammer ought to be for most.
Don’t Stop The Avalanche
AMD has been doing a remarkable PR job the last few months, taking full advantage of Intel’s woes to get positive publicity.
There’s only one problem: this isn’t a snowjob; this is an avalanche. AMD has managed to convince an obviously not-too-inquisitive media that it has succeeded where Intel hasn’t and can make cool-running 90nm chips.
What has escaped the media’s attention is that AMD is making relatively slow, cool-running 90nm chips.
What happens when you try to run them faster?
The IBM Experience
As you probably know, IBM is AMD’s partner in developing SOI technology. Like AMD, they, too, have a 90nm SOI processor, which they call the PowerPC 970FX. Unlike AMD, they’re pretty open about talking about their problems.
As you may know, IBM thought it could deliver a 3GHz processor to Apple by now, but found out that 2.5GHz was all they could deliver.
Go here and you’ll find IBM talking about the power consumption of the 970FX in quite a bit of detail, which tells you why IBM could only deliver 2.5GHz.
Read the text and the second chart, and you’ll find that the 970FX is indeed a very cool running chip up to about 2GHz. It chews up only 40-50 watts at that speed (and needs only 1-1.1V to do so).
Make it run at 2.5GHz, though, and you find that this processor now needs 100 watts at 1.3V. Quite a difference, isn’t it?
Why does this happen? Read a bit more, and look at the first chart, and you’ll find that once you push the chip past a certain point, it starts leaking power, a lot of power. Maybe not quite as much as Prescott, but still enough to qualify it as a relative.
As IBM puts it:
“This total power is very sensitive to voltage. Each power component has a steep response to voltage, for example, ac switching power is proportional to V2, whereas the subthreshold power is proportional to V3 and gate leakage power has an exponential (our emphasis) relationship to voltage.”
Obviously, the PowerPC represents a different design than Hammer, though it is closer to Hammer in general design (relatively short pipeline, relatively large amount of work done per cycle) than Prescott.
Nor does this mean SOI is no good. If IBM or AMD tried to make their chips without SOI like Intel does, they’d probably melt.
What this does mean is that SOI works well up to a point, then it falls prey to the same kinds of problems as Prescott, if not for exactly the same reasons.
An Even Better Reason Not To Push It
IBM says the following about voltage for the 970FX:
“At the same time, the 1.3-V application condition can only be applied to products with <50,000 power-on-hours (POHs). To accommodate the reliability requirements of a 100 000-POH system, the upper voltage must be limited to 1.2 V."
In English, increase the voltage past 1.2V, and you shorten the life of the CPU considerably. No doubt the equivalent numbers for AMD are different (and higher), but this tells us that SOI is probably more sensitive to damage from excessive voltage than most.
Certainly a likely reason for AMD not to put out fast 90nm processors chewing up 1.5V, for instance.
Hiding The Power Surge
You may say, “But people have been running tests and Hammers don’t get a whole lot higher when you overclock them a bit!”
There’s a very simple reason for that. SOI is very sensitive to voltage. If you look at the second chart again of the linked article above, you’ll see a thin blue line extending from about 50 watts at 0 Ghz to 100 watts at 2.5GHz.
That represents power consumption at 1.3V. At that voltage, power leakage is at a maximum, so the wattage at 0GHz is pure leakage. The power curve doesn’t go up all that much, but all that means is that until you get close to 2.5GHz, you’re just shoving more power into the CPU than it needs to work.
That’s essentially what any power curve for Hammers show. They look relatively smooth and stable, but that’s because voltage stays high and constant. Run a Hammer at 1.8Ghz or slower, and you’d probably be able to reduce voltage quite a bit and still have a functioning CPU.
P.S. If any of you with Hammers want to try running them slowly while lowering the voltage (though anything below 1-1.1V is probably not a good idea; IBM says “Today, a functional limitation exists in the 90-nm design, preventing operation below 1.0 V.”), I’d be happy to get and report on the results.
Conclusion: Part One
AMD is putting out relatively low-speed 90nm chips because that’s all they can safely and responsibly put out for now, while claiming credit for something (i.e., being relatively cool compared to Prescott). A 2.4GHz 90nm Hammer likely wouldn’t be so cool, and a 2.6GHz chip almost certainly wouldn’t be.
They’re not trumpeting this because if they did, one of the first questions would be, “Why don’t you put out faster ones?” and the true answer would be “We can’t.” Of course, they wouldn’t say that, but why be deceptive when you can avoid the question all together.
And you think you can cover your tracks in a few months and make it look like a normal, natural progression.
There is now some good reason to think AMD can do just that.
Covering The Tracks
Most of this article has talked about the limitations of the 970FX. However, the 970FX isn’t the last word in 90nm technology from IBM.
Recently, this place reported that IBM plans a follow-up to the 970FX called the 970GX. It is expected to be ready around the first/second quarter of 2005, and come in at speeds of around 3GHz.
Perhaps not-so-coincidentally, we recently got this blurb from AMD engineers via XBit Labs.
“However the second-generation 90nm production process, which should be put into life next year, will include both: SOI and strained silicon. AMD engineers are very excited about this combined technology aka Strained Silicon Directly on Insulator (SSSDOI), since they expect it to open up new opportunities for further clock frequencies increase.”
Actually, the blurb about strained silicon alone probably doesn’t mean all too much, the 970FX already uses SSSDOI. However, combine that with a year’s worth of advances and tweaks on SOI, and you could well see a 20% or more jump in maximum performance.
And when are we likely to see that? Hmmm, the AMD roadmaps talk about a new generation of 90nm SOI processors showing up the first half of 2005. Up to now, no one’s been sure what they were supposed to be about, we guessed a long time back that these were DDR-2 capable chips.
Taking all this new evidence together, though, and while DDR-2 capability may well show up in at least some of these chips, I would bet that most of whatever will make the IBM chip run faster will be in these AMD chips, too.
And when are these due out? Around the same time as the IBM chip. What a coincidence.
I don’t mean to imply at all that AMD is just sitting around waiting for IBM to improve Hammers for them; I’m sure this has been a joint venture.
A 20% improvement could mean 3-3.2GHz overclocked chips. If it does, AMD will have no problem fulfilling (and perhaps overfulfilling a bit) its roadmap for the rest of the year.
Conclusions: Part Two