Chris Edwards: November 2007 Archives

At a workshop on advanced architectures earlier this week, the ARM's director of research Krisztián Flautner made an unprovable but seductive assertion. He said upfront that he had practically no evidence for it. But it's one of those statements that has the ring of truth about it. And at its heart is something that may prove to be the big problem that faces microprocessor makers as they try to work out how many cores they should put on one piece of silicon.

To make use of tens or hundreds of cores, you need applications that are "embarrasingly parallel". But Flautner is not so sure such things exist. His rule? That there is no such thing as an algorithm that is embarrassingly parallel, just an algorithm that hasn't been optimised enough yet. He pointed to situations in graphics where operations that used to be parcelled up and split across many processors in a dumb way have been replaced by approaches that apply more local intelligence but which are much tougher to parallelise.

I can think of one or two algorithms that remain embarrassingly parallel - they just aren't that useful. Anyone roving an electronics show in the late 1980s will remember people trying to push the Inmos transputer. If they weren't displaying scenes of shiny balls rendered by ray-tracing, they were processing the good-old Mandelbrot set.

People found lots of other ways to render scenes since then that don't parallelise so well. It's hard to think of a smarter way to generate Mandelbrot set images than the way they were processed then. But why would you want to?

Intel is spending a lot of time telling people, once again, how much of a breakthrough it has made in introducing metal gates - reversing a trend that Intel's founders began back in 1967 when they developed the first silicon-gate process at Fairchild. The Penryn processors are not just 45nm processors, they are "45nm hafnium-based high-k metal gate" chips.

The processors have already entered production: reverse-engineering specialist Chipworks is etching away the top of its sample right now to see what lies underneath. So the 45nm generation is underway. And, based on the combination of metal gates and 45nm, Intel looks to be way out in front. But it's not necessarily the case. Foundry TSMC has started running customer wafers through its most advanced fab. We can expect the chips from those wafers to go on sale soon. Matsushita has already got manufacturing runs of its Uniphier chip.

Where Intel does seem to be going it alone is on metal gates. So, it should come as no surprise that the company is focusing attention on that part of the process. However, metal gates are not necessary for making 45nm devices, only those that are likely to burn a lot of power, such as PC processors. For those operating in the low-power segments, the metal gate is something of a luxury. And it's a luxury that a lot chipmakers are avoiding, at least for this generation. Only IBM has said it will use them and it has yet to show off any chips made on a 45nm process.

Although the combination of metal and a dielectric with a higher dielectric constant than silicon dioxide stops current flowing out of the transistor out through the supposedly insulating gate, that form of leakage is only one small part of the overall power loss. In a high-speed chip such as a PC processor, most of the leakage is down to the transistor passing electricity even when it is supposedly turned off. These days, transistors don't so much switch between on and off; they go from on to less-on. As a result, most of the changes needed to deal with leakage have to happen during the design phase. And, those changes cut gate leakage as much as the other, more bothersome form: subthreshold leakage.

So, other chipmakers do not share Intel's enthusiasm for metal gates. They don't want the extra cost of shipping in sophisticated and slow equipment that lovingly deposits the novel high-k dielectric one atomic layer at a time. Even Intel may not choose to use metal gates on the products it plans for the consumer-electronics market.

TSMC last year said it was developing a metal-gate process for 45nm but, as the process got nearer to production, found that a more conventional gate structure was all that it needed. The foundry is looking at the metal gate option seriously for the 32nm process and will, if a customer really wants it, think about putting a metal gate on the existing 45nm process. But they are really going to have to want it.

However, Intel's move won't shove up the cost of making Penryns that much, it seems. In comparison with previous desktop processors, the die is pretty small. The penny you see sitting on a wafer full of Penryns in the publicity shots gives a very good idea of how small the chip is. That means Intel can get a lot more onto a wafer, reducing its overall production cost. This, most likely, more than offsets the increase in cost caused by the use of the more complex metal-gate process. And it gives Intel a potentially strong lead in any price war that it might conduct with AMD to push the advantage it currently holds. There are other companies with 45nm devices, but AMD is not one of them.

Microchip Technology, the company that stole the lead in 8bit microcontrollers from Freescale Semiconductor (back in the days when it was part of Motorola), is now going to have a go at the 32bit market. The company has licensed the M4K core from MIPS Technologies rather than give any of its own architectures another workover. Two years ago, the company turned its digital signal processors into general-purpose microcontrollers to try to break into the 16bit market. Now it has another architecture to push.

What does it mean? Basically, the decision provides designers inundated with choices of low- to mid-range 32bit micros with...more choice. Yay.

Now, this is a stunt that Microchip has pulled off once before. But only once. If you rewind time by about 12 years, you can look at the circumstances that gave an aggressive relative newcomer the advantage over a seemingly unassailable market leader. The problem is that, this time around, the circumstances are different. And Microchip has yet to demonstrate that it can turn success in the 8bit market into big sales in the 16bit sector, let alone the even more competitive 32bit space: a market space that a number of chipmakers have made the battleground for control of the future microcontroller business. And they all started earlier.