Design: March 2009 Archives

In the Twitter conversation that followed the deal between Intel and TSMC, in which the number-one chipmaker will let a foundry make some system-on-chip (SoC) parts for favoured customers, I made the rash prediction that Intel would probably not allow a customer to resell the parts it designed, or at least not OK a deal where that was likely to happen.

There was nothing made public to back that up, other than Sean Maloney, executive vice president of sales and marketing, mentioning several times that Intel would remain the company in charge of deciding who can build an SoC around an Atom core.

Intel wants to break into some markets that it cannot get at because some customers do not want the standard silicon that Intel makes or they do not want Intel to make a part to their specification.

However, Intel is not going to break into a market at any cost. CEO Paul Otellini has already taken a pasting over Atom cannibalising sales of its mainstream PC processors. Can you see him being eager to get it in the neck from the analysts a second time around, accusing him of letting other chipmakers cannibalise Intel's direct Atom sales?

Conversely, who'd want to drum up sales for Atom-based products in the knowledge that Intel can take the ball away just as you get successful and go for the second generation.

Let's take one of the most obvious candidates for licensing an x86 core for use in portables: nVidia. OK, you can stop laughing at the back. It would be a funny move to make when you are fighting in the courts just to bolt a coprocessor onto the side of Intel's Nehalem. But nVidia is a big customer of TSMC, could easily wrap a GPU and chipset around an Atom and has signalled its intention to do something similar through a tie-up with Via, although that has gone a bit cold recently.

But nVidia is telling anybody who cares that the CPU is turning into an appendage – it's the GPU that has all the horsepower. I can construct in my head a byzantine scenario where Intel sues nVidia to get the GPU maker's tanks off its lawn so it can tie up a deal around Atom that boxes in nVidia, but it's not pretty. If you wanted something to cannibalise Atoms in netbooks, the device would be similar to something nVidia would like to make.

What about one of the handset silicon suppliers, like Qualcomm? Intel has been banging its head against the wall trying to do something in handsets for years. Qualcomm's engineers could cheerfully slide an Atom alongside an ARM baseband processor for a Windows-running smartphone. But it might be too good at it and we're talking about a company that loves its lawyers.

ST-NXP is a possible alternative to Qualcomm in handsets. ST is a company that is good at doing deals and has managed to work with Intel in the past, notably through the Numonyx mash-up. If I thought Intel wanted to do a deal to get into handsets, ST-NXP would be my choice. Texas Instruments doesn't want to play anymore in baseband and its applications processors compete head-on with Atom. Freescale and Infineon have problems of their own right now and Broadcom is not a major player in a business that demands bulk.

However, an Atom with a 3G baseband bolted to the side would be pretty attractive in a netbook. Oh dear, we're back on Intel's core market for Atom. Do I hear the C-word again?

Industrial and automotive provide some opportunities for a company out of Intel's sphere to get involved. But it's hard to picture a situation where the volumes justify doing a special Atom versus buying the standard product and sticking it alongside a custom I/O device.

So, I keep coming back to the systems houses. And not many of those are really in the running to do a 40nm, 32nm or 28nm chip just to get PC compatibility. Right now, only two names come to mind. One is Apple, inevitably. The other is Cisco. Both have extensive experience in designing their own silicon. Cisco, in particular, has run a fair number of chips through TSMC.

Apple? Let the netbook speculation resume. Intel offering Steve Jobs the opportunity to build a custom SoC might be the carrot to stop Apple from using the ARM Cortex-A9 as the guts for a souped-up-iPhone take on the netbook. However, the more you look at how much control Apple wants over its new markets, such as phones, the company is better off just sticking with the ARM infrastructure. A netbook that is incompatible with a MacBook will cannibalise Apple's computer business less than one that is modelled after an Asus Eee.

Cisco looks a bit more promising. The company keeps making noises about getting into consumer electronics, although the results have been less than dazzling. An Atom-based home gateway or home-tablet thingy are the kinds of products I can imaging Cisco wanting to make.

There are other consumer-market players I could list. Sony, for example. But it's hard to think of companies who are desperate for x86-compatibility outside of those who have proven, over the years, to be happy with the standard silicon – I'm looking at you Dell and HP.

The deal with TSMC is not the first time that Intel has tried to manoeuvre into the embedded space by signing up a company to make chips for customers to whom it cannot sell a standard product. However, it did take me a while to remember and then find the reference.

Way back in the summer of 1992, Intel cut a deal with VLSI Technology – now part of NXP Semiconductors and the original contract supplier of ARM processors to Acorn Computer and Apple – to help customers develop application-specific ICs (ASICs) based on the 386SL. Although the 486 was in full swing by then and the first Pentium was due a year later, Intel thought its 386SL could make inroads into the market for PDAs.

Intel invested $50m in VLSI, providing the chip giant with a 20 per cent shareholding – somewhat different to today's deal with TSMC – and a licence to use the 386SL. Although VLSI planned to make an integrated processor called Polar, the companies called the deal off in 1994.

The initial customer was to have been Compaq and everybody was talking the deal up through 1993. But things did not turn out as planned: Apple's ARM-based Newton was floundering by mid-1994 and the market forecasts for handhelds were not looking realistic. From VLSI's 10Q financial statement for 3Q94:

"To date, the handheld market has not developed per initial expectations. The Company now believes that significant uncertainty exists as to the revenue potential of the handheld market in general. As a result, the Company and Intel, its partner in the Polar development effort, have canceled further production of the Polar product and in early November 1994, terminated the amended July 8, 1992 Technology and Manufacturing Agreement between the companies. In addition, the parties are exploring alternatives to carry out Intel's expressed desire to dispose of its equity ownership position in the Company."

TSMC will be the first foundry* to make Intel x86 processors. And rather than use Intel's manufacturing process, Intel is working on a port of the Atom core to an, as yet, unspecified TSMC process.

Anand Chandrashekar, senior vice president and general manager of Intel’s ultra mobility group, confirmed the company will not pass on its manufacturing technology: “We are porting Atom, some of them, to TSMC’s process libraries using their libraries and their flows.”

As Atom was designed for the Intel 45nm high-k metal-gate process – and the number-one chipmaker described a variant for system-on-chip (SoC) designs at last year's IEDM – it seems likely that Intel will port the core to TSMC's own flavour of high-k, metal-gate. However, on the foundry's current schedule that will not be ready until next year as the first commercial process to sport that kind of gate will be the 28nm technology.

TSMC has run prototype wafers on a 32nm high-k process and was slated to talk about that at IEDM in December but switched gears on the day and reported its experiences on the 28nm version. Right now, TSMC is working with standard and low-power versions of a polysilicon-gate process. In principle, there is nothing stopping Intel from trying to port Atom to the existing 40nm TSMC process. But given how much Intel has claimed about the power advantage of high-k over polysilicon and that it is in control of the porting process, it seems unlikely.

Sean Maloney, executive vice president of sales at Intel, claimed at today's press conference that there is "a sense of urgency" at the two companies but also that they favour leading-edge processes.

Although TSMC will make Atom-based SoCs, Intel will still call the shots on who gets to use it. And Maloney was clear that the deal is there primarily to get access to markets that Intel would find difficult to enter on its own. The chipmaker will still make Atom-based SoCs for netbooks, consumer electronics and handhelds, he claimed.

“We will maintain full control of who we sell to,” Maloney said. “We have markets in our mind that we are going after.”

Maloney claimed work on products has started: "The very first products are currently in definition. But we are not talking about that today."

Insisting that the deal is not about securing additional capacity but attracting a wider range of customers to the Atom, Maloney said: “Intel is a manufacturing company. During the course of the past 16 years, when people questioned whether we should invest in manufacturing, the answer was yes. It is a central core competency.

“So, why port a core to TSMC? Our belief is that as we look at the next three or four years, [customers] will need to embed PC functionality into these devices,” Maloney added, indicating the growing market for portable and embedded electronics. “We also want to go after markets where we are absent.”

Assuming that it will take at least a year to get customer designs from the block diagram to physical design, TSMC is probably going to be in a position to start production on a high-k process when that happens. That points to consumer products being ready for the Christmas season of 2010.

* But not the first company to make x86 processors under licence to Intel. A bunch of people could make the old 8086s. AMD and Harris had access to the 80286. IBM was able to make everything up to and including the 80486 but only for use in its own computers, although it wound up making processors for Cyrix.