With the release of 28nm toolkits to members of IBM Technology Alliance, the attention turned to Intel and how Intel plans to counter the future 28nm chips from AMD/ATI [manufactured by GlobalFoundries] and other competitors.
We spoke with several sources close to the heart of the company, and unfortunately, we were able to get off-the-record statements only – hence the bracketing of this story as a "Rumor". According to one source, half-node such as 28nm process is typically never used with CPUs, due to custom libraries [transistor design] used in CPU manufacturing. It is too expensive to change masks in full-node/half-node/full-node rhythm, and Intel decided against it.
Intel’s 32nm chips
According to the schedule, Intel’s introduction of the 32nm process technology will mostly depend on the existing inventory of 45nm chips. We do expect an introduction of 32nm in Q4 of this year, even though WW roadmaps still position chips such as Clarkdale in Q1’10. You can expect numerous 32nm chips being introduced over the next two years: Nehalem’s die-shrink family is called Westmere and is consisted out of at least four chips: dual-core CPU [plus dual-core CPU+GPU], native quad-core, native octal-core and possibly a dual-die hex-core [16 core CPU].
This will be followed by Sandy Bridge, a new architecture [tick-tock]. Sandy Bridge will also consist out of native quad-core and octal-core chips, continuing in 2011. The destiny of Atom product line was not known at this time, since Intel isn’t fond of using latest Fabs for a product with a pretty low ASP, hence the recent signing of an agreement with TSMC. This ends the 32nm cycle for CPUs.
When it comes to its GPU line-up, Intel plans to adopt 32nm process for its Larrabee core as soon as possible. The chip giant is working on several 32nm Larrabee designs, with at least one being planned for "fusing" with the CPU die on a processor socket – succeeding Clarkdale. The company is keeping its tick-tock model for Larrabee as well, using a mature manufacturing process for new architecture, applying a die-shrink, followed by a launch of the new architecture using the same architecture as the die-shrunk one. In case of Larrabee, 45nm is a start, 32nm is a die-shrink, and real second gen part is 32nm, then 22nm die-shrink and so on.
Intel’s 22nm chips
In mid- to late-2011, Intel will introduce Ivy Bridge, a 22nm die-shrink of Sandy Bridge. Just like Nehalem-to-Westmere, Sandy Bridge-to-Ivy Bridge will be more than just a die-shrink, but we should still expect Socket compatibility between the two. You should expect a quad-core and octal-core parts, with quad-core parts succeeding dual-core 32nm in entry-level models.
In 2012, Intel will introduce a second generation of 22nm parts, and a completely new CPU architecture. The first product is known as a native octal-core chip named Haswell.
Intel’s 32-22-16nm vs. 32-28-22-18-16nm
If we take a look at Intel’s process, even though the company isn’t playing with half-node process, their long term gain is that with "full-node only" approach, fewer things can go wrong. We all remember leakage on 90nm full-node, but even more so 80nm [ATI Radeon 2900], or the current 40nm issues over at TSMC. Intel tweaked its full node + custom design library concept to perfection, while we saw that other players struggled at various points in time, such as nVidia’s path with GT200 GPUs and their 65/55nm path.
Traditionally, Graphics Processing Units use every possible manufacturing step, since they feature much more core-logic and die-shrinks result in significant power savings. CPU designs are limited by the large amount of cache, and a half-node process would not yield with a significant gain, either in clocks or power. Thus, here comes Intel’s decision to stick their upcoming GPU to full-node process. This way, AMD and nVidia will be in cost advantageous position every time GlobalFoundries or TSMC successfully introduce a new half-node process technology.
Intel heavily invested in Larrabee – LRB, and some figures we heard were measured in the billions of dollars. Intel, regardless of how stuffed with cash is or isn’t, would not invest sheer billions and not be competitive in the field. They will probably lose on cost-effectiveness and clocks, but there is an ominous warning for full-node/half-node companies.
Again, according to our sources, Intel used all of its manufacturing knowledge on Larrabee and the company is using vastly improved transistor design libraries when compared to TSMC’s best. At least, that’s what people wearing Intel shirts and T-shirts are saying.
There really can be only one: let the games continue. Intel’s tick-tock manufacturing/product symbiosis resulted in the Core 2 and Nehalem architectures. On the other hand, ATI proved how to utilize half-node manufacturing steps by utilizing 55nm to its full potential, with the Radeon 3800 and 4800 series. 40nm is also looking good for the company, with Radeon 4700 coming at sub-$100 price point and record performance in that pricing bracket.
Note: Have one very important thing in mind – just like CPUs; consoles traditionally do not use half-node process. And if Intel succeeds in its push of ******* [concept console platform] and wins the Microsoft Xbox contract, part of the reason will have been Intel’s proven manufacturing muscle and zero issues with die-shrinks in 12-18 month steps. That is still a long shot, since ATI is already taking wishes from developers for the next console parts. But this article is not about next-gen consoles :-]