On March 2, Intel and TSMC announced that they will â''split the Atom:â'' Intel will port its ultra-low power Atom microprocessor cores to TSMC, essentially licensing the IP to customers who will have their Atom-based designs manufactured by TSMC. This was a big deal mainly because Intel usually doesnâ''t share its painstakingly designed toys: Atom, its line of tiny x86-based microprocessors, was designed to be used in low-power applications and for that reason was never intended for power users. So it ended up in a lot of netbooks.
Now, Intel is offering TSMC the Atom for its embedded systems customers. The collaboration will let TSMC offer its customers systems on a chip built around the very low power Atom design, for use in devices like cellphones, set-top boxes and cameras. Intelâ''s giving TSMC everything they need, including process, IPs, libraries, and design flows.
Well, almost everything. Intel held onto its high-k metal gate process. Thatâ''s a bit like licensing the Big Mac without the special sauce.
But I was confused.
Wasnâ''t Atom expressly designed for the 45-nm process technology?
And didnâ''t Intel just about kill itself trying to squeeze its transistors down to the 45 nm node?
The answer to both of my rhetorical questions, by the way, is yes. Recall that the diet plan that got Intelâ''s transistors squeezing into their 45-nm skinny jeans was the high-k metal gate process, based on a ground-breaking hafnium oxide (HfO2) gate dielectric.
The result was, and continues to be, impressive. In November 2008, an EETimes article painstakingly poked around the innards of Intel and AMD 45-nm processors and found that â''Intel's 45-nm high-K metal gate transistors have the best peak drive currents on the market with 1.36ÂµA/Âµm for NFETs and just over a milliamp for PFETs.â''
So Intel is NOT giving TSMC HKMG? Isnâ''t that just a recipe for failure?
By the time I reached Intelâ''s spokesperson Chuck Mulloy, he had probably answered this question just one time too many. â''You are correct,â'' he e-mailed me, and I could read a heavy sigh between the lines. â''We are not sharing process technology with TSMC, or anyone else for that matter.â''
That raises a couple of questions, then:
1. How will the original 45nm HKMG Intel Atom compare to the 45nm non-HKMG
Walmart TSMC brand? Letâ''s confine it to speed and power dissipation. Show all work.
2. Is there any logic circuit in the Atom that would not work without the 45nm HKMG process? For example, the Penryn's new FP division system is probably the only thing in it that would not work at an earlier node. Is there a similar situation in the Atom?
3. What will be the process technology of the Atom-based product? Since Intel has been shipping 45 nm products for a while, it is a matter of speculation if that will be the process technology node of the Atom-based product.
Despite knowing there was no way in h- e- double-hockey sticks I was going to get the answers from them, I went ahead anyway and asked TSMC spokesperson Michael Kramer*, and Intelâ''s Chuck Mulloy.
Mulloy: â''I can't contrast the power or performance because Atom is currently only built on our 45 nm process with high-K metal gate so there's nothing to compare it to.â''
To the question of what technology node TSMC would debut their off-brand Atoms at, Kramer* replied, â''We can not comment on the details of the technology node that Atom cores will be ported to, but you can assume that it is an advanced process node.â''
So Atoms, which are designed for 45-nm node process technology, are being ported to TSMC, which gets at best 10 percent of its business from customers that require the most up-to-date manufacturing technology (the rest is split pretty evenly between one-generation-behind and really older nodes like 90 and 130). And thatâ''s during a good year, which this is not (ASMLâ''s Bill Arnold told me that the $248 billion a year industry is down this year to an expected $200 billion).
But then, the big reveal!
â''I would also like to point out,â'' Kramer* continued, â''that TSMC has its own high-k/metal gate technology at the 40/45nm node.â'' [italics mine]
Did you know that? I didnâ''t.
About a billion papers for IEDM, ISSCC, SPIE, and the rest of the alphabet soup, have TSMC introducing HKMG at the 32 nm node (and in TSMCese, that means the 30 nm and/or 28 nm node. Letâ''s just pretend itâ''s 32. Now my head hurts.) â''TSMC and everyone else won't have high-k dielectrics and metal gates until 32nm,â'' Real World Technologies guru David Kanter confirmed later. â''For TSMC and IBM, that would be late 2009 or early 2010.â'' Kanter is an expert, and he kind of knows everything. Some kind of savant, maybe. So I almost trust his word over that of TSMC.
As you might imagine, I emailed Michael Kramer* immediately with a message that can be summed up as â''Say what now?â''
I havenâ''t heard a peep from the man since. I imagine him in the witness protection program, which would be rough on a Kramer* in Taiwanâ''heâ''ll have to wear a ski mask in perpetuity and have his vocal chords surgically altered.
As for the specifics of a 45-nm Atom microprocessor without Intelâ''s fabled high-k metal gate, Kanter says it is likely that without high-k metal gates, the transistors in TSMC-produced Atoms will run perhaps 20-30 percent more power consumption. â''But in some applications, you donâ''t care about that,â'' he says.
But then I went poking around on the interwebs some more and found this!
In December 2008, a paper published at EDSSC (motto: â''the other IEDMâ'') was called â''National Project on 45 to 32 nm Metal Oxide Semiconductor Field Effect Transistors for Next Century IC Fabricationsâ''â''by a university consortium in Taiwan! Which is working with TSMC!
When pressed, one of the authors, K.S. Chang-Liao, responded with great brevity: â''I will simply state the fact that this is a Project sponsored by MOEA, and work was done independently by professors in National Tsinghua University and NCTU and good discussion was underway with TSMC for adopting in using our IP.â''
The person who finally cleared all this up was Dr. Rajarao Jammy, a veteran IBM semiconductor insider who is now vice president of emerging technology at SEMATECH.
He basically told me to stop focusing psychotically on a single aspect of a complicated device. "In some senses, yes, there would be higher leakage,â'' he said. â''But it's not just about transistor leakage. There would be a whole range of other leakages and power related issues.â''
If you take the transistor itself, the leakage comes from many places. One might be junction-related leakage, threshold leakage, power dissipation from the metal lines themselves. All of this is power dissipation. Leakage is one issue, power is another.
The overall power management of the device comes from several factors. Gate leakage is only one of those. Controlling gate leakage, of course, helps keep overall leakage down--and Atom is one of those chips thatâ''s really aimed at really low standby power operation. However, when it comes to what kind of leakage a device maker, or rather an end-user might tolerate--it would all depend on the application.
â''The design cores could essentially be the same,â'' Jammy explained. â''How the system inside the chip operates might be the same, but they donâ''t necessarily need to be a mirror image of each other.â''
Chuck Mulloy has made it clear that the Intel-TSMC MOU isnâ''t clear yet on how their arrangement will function. It's possible that there are certain cores in the Atom processors that are going to be imported into another design TSMC might have. Different companies could be making their chips with Atom cores.
But, unlike most knockoffs, the chips will not be identical in every way.
Well, but regardless of whether my questions barked up the wrong tree, I still think it's interesting that TSMC has 45-nm HKMG but isnâ''t talking about it.
I wish you the best of luck, Michael Kramer!*
After I posted this entry, TSMC released Michael Kramer from his carbonite prison just long enough for him to pen the following note.
Dear Ms. Adee,
My apologies for my late reply and I hope it's still useful to you.
Double-checking on the status of our 40/45nm HKMG took a little longer than
While we have not put 40/45nm HKMG on our roadmap or made any official
announcements, we certainly have the capability to offer this if customers
have the need for it.
Having said that, unfortunately at this point we really can not confirm to
you whether the ported Atom SOC cores will use HKMG or not, or even if it
will be at the 40/45nm node. We welcome you to provide Spectrum readers
with your analysis and insight into the details of this agreement, but
again I apologize that TSMC can only say that the Atom SOC cores will be
ported to an advanced technology node.