Is AMD's 65nm SOI process broken?

Since news first broke that AMD's Barcelona will be six months late (confirmed by AMD CEO) and six hundred megahertz short on clock speed, the microprocessor community began to wonder aloud whether AMD's 65nm SOI (Silicon On Insulator) process is fundamentally broken.  The fact that there are absolutely zero production AMD processors beyond the 65nm Athlon X2 desktop processor at 2.6 GHz is adding fuel to the fire. 

To calm the analyst community and Wall Street, AMD put out some hand-picked demos of a 3 GHz "Phenom X4" which is a desktop variant of Barcelona quad-core CPU built on AMD's 65nm SOI process" at last month's analyst meeting.  That news was greeted with some controversy and debate broke out whether the existence of 3 GHz Barcelona demo CPUs settled the question.  Yesterday, Charlie Demerjian (contributor for The Inquirer) has put forth a theory that may explain this 65nm shortcoming.  Demerjian pieced his theory together based on bits and pieces of information from AMD and concluded that this was nothing more than a case of botched timing rather than a scaling problem.  I actually discussed this subject with Charlie at DEFCON 15 earlier this month and I thought it sounded plausible at the time, but other industry experts I've spoken to were less accepting of this theory.

Charlie's explanation is that AMD never "taped out" (completion of design phase and moving in to manufacturing) a 2-MB cache version of the K8 (class of processors including Athlon and Opteron) dual-core processor.  There was only a six month window of opportunity to sell those higher-end 65nm chips since Barcelona was thought to only be six months away at the time.  Therefore there was no pressing need to ship the higher-end low-volume parts on 65nm for such a short life cycle nor do server vendors want to validate an entire new line of microprocessors for a product that has a 6-month life cycle.  As a result, only a 1-MB cache version of the 65nm K8 CPU was taped out which is only used for the mainstream high-volume desktop market.

Note: The high-volume higher-yielding 65nm parts helped cut costs and increased AMD's gross margins during the last quarter despite declining ASPs (Average Selling Prices).

Anything higher than 2.6 GHz on the desktop parts and the entire line of Opteron server processors would simply have to ride out the 90nm process to the end until the 65nm K10 Barcelona CPU takes over in mid-2007.  Since mid-2007 came and went with Barcelona delayed, AMD was stuck with no choice but to keep cranking out high-end desktop parts and Opteron server processors on the same old low-yield 200mm wafer 90nm process.  Besides, as Charlie explained it, the whole 90nm process was already "paid for" and there was no point in cranking up the clock on 1-MB cache 65nm Athlon X2 processors to 2.6 to 3.2 GHz with half the cache of their 2-MB 90nm siblings since that would result in lower-performing parts at the same clock frequency.

I spoke with respected CPU analyst David Kanter (Real World Technologies) and he had a different take on the situation.  Yes it's true that there was no way AMD could have taped out a whole new 65nm K8 processor with 2-MB cache despite the possibility that AMD "probably had some idea that things might not be totally on target" with Barcelona.  AMD would have had to design a whole new L2 cache controller and L2 cache which is hardly trivial and there are far bigger engineering priorities for AMD to tackle.  As Kanter puts it:

"AMD was faced with a tough decision - they could spend engineering resources on improving the frequency for the 65nm K8, or they could use those same resources to increase frequency and yields on Barcelona."

However, the missing 2-MB K8 design alone cannot explain the reason there aren't any high frequency 65nm K8 parts.  Kanter explained that even if a higher clocked 1-MB cache 65nm Athlon X2 was a little slower (probably less than 5%) than a 2-MB cache 90nm Athlon X2 chip, it still makes a lot of sense to offer the higher-clocked 65nm version at a lower price if it was at all feasible.  This argument makes a lot of sense to me because AMD's latest "black edition" X2 6400+ 3.2 GHz 90nm processor costs a whopping $260 (street price) yet it benchmarks noticeably slower than a $200 Intel E6750 2.66 GHz Core 2 Duo processor.

Note: AMD's newer 300mm wafer with 65nm parts produces 2.5 times more chips than AMD's old 200mm wafer 90nm process.  But Kanter explained that this doesn't mean it's 2.5 times cheaper to produce a 65nm K8 chip since there are still other fixed costs like "packaging, testing, etc".  However, a 65nm processor will be significantly cheaper to produce and have a noticeably cheaper price tag which is something that AMD could use for the high-end desktop.

The reality is that if AMD could simply crank up the clock speed of the existing 65nm 1-MB Athlon X2 processor to 3 GHz or beyond despite its shortage of cache, AMD would almost certainly do it because it will put them in a better competitive position and it will lower their production costs.  So there may indeed be some speed path problems in scaling the existing K8 65nm processor but those speed path problems may or may not be difficult to solve.  Even if they were easily solvable, AMD would almost certainly commit its engineering resources on solving the scaling problems of the K10 Barcelona processor and just ride out the 90nm process.

So the sixty-four thousand dollar question remaining on everyone's mind is whether AMD's 65nm SOI manufacturing process fundamentally broken.  Unfortunately that isn't such an easy question to answer and the lack of high clocking 65nm parts doesn't necessarily prove a serious problem with AMD's 65nm SOI process.  But the lack of higher clocking 65nm parts combined with the delay and underwhelming 2.0 GHz clock speed of the initial Barcelona parts due next month doesn't exactly inspire confidence in AMD's 65nm SOI process.  The existence of air-cooled 3.0 GHz Barcelona demo units tells us that at least there is no hard-barrier but there is a world of difference between showing off hand-picked 3 GHz demo units and producing 3 GHz parts with sufficient yield within allowable TDP (Thermal Design Power).  Whether these problems are temporary snags or a disaster waiting to happen is yet to be determined and we'll have a better assessment of the situation by the end of the year when we see where AMD's clock speeds on Barcelona are.