Plenty of Sun users to skip Fujitsu march
Ready to be Rocked instead
A large chunk of Sun Microsystems customers may skip right over the upcoming SPARC experiment with Fujitsu, according to server chief John Fowler.
In the coming months, Sun will launch a line of SPARC64-based servers in conjunction with Fujitsu. These servers will replace boxes based on Sun's own UltraSPARV IV+ chips and help the vendor compete against the likes of IBM and HP in the Unix market. The rub, however, comes when you realize that the Fujitsu gear will only out pace similar systems based on new "Rock" silicon from Sun by a few months.
"The timing of the products is closer together than normal for this class of technology but that's just an engineering reality," Fowler told us. "We've had worse problems like not enough products."
Sun originally turned to Fujitsu to share chip and system costs after it canceled the UltraSPARC V project in 2004. Back then, Fujitsu's aid looked like a godsend with Sun projecting a massive gap between speed boosts to its UltraSPARC chips and the, at the time, unrefined release schedule for Rock systems.
Hindsight, as you all know, can be a cruel master.
The UltraSPARC IV line of processors has been tweaked past 2GHz and sales of servers based on the chip have gone better than anyone could have imagined. Meanwhile, Fujistu has slipped on the release of fresh SPARC64s, and Rock - due out in 2008 - has arrived right on time.
"Engineering works in funny ways," Fowler said. "We got more out of UltraSPARC IV than we expected, and Rock is happening in the time frame we hoped for. It could have easily been very different."
As a result of the overlap, Sun will have two high-end server lines aimed at very similar tasks - back office stuff such as Oracle, SAP and the like. So far, the company has been unwilling to admit this overlap, although we coaxed such a confession out of the good-natured Fowler, who was kind enough to give us an interview just days after we poked at him in a story.
"A lot of people, especially given that we have gotten to 1.8GHz, 1.95GHz and 2.1GHz with UltraSPARC IV+, will just stay there until Rock," Fowler said. "Other people will be interested in the features and performance of the combined Sun/Fujitsu products.
"We aren't making people transition at all. That's part of the value proposition. Strong choices and no forced march."
Just like the Itanium rollout. Oh, wait . . . (To see just how good of a sport Fowler is, check out this video. No Carly doll in the tub?)
So, for Sun customers, the presence of two high-end server lines looks like a nice option. The Fujitsu gear, based on dual-core chips, won't blow you away, but should be decently competitive with Power- and Itanium-based servers. We're sure you're all too familiar with such a scenario.
Then, in 2008, Sun looks set to give rivals some more serious competition with the release of the 16-core Rock chip.
For the first time in many years, we're feeling bullish about Sun's high-end system prospects. That said, one does wonder about the costs tied to supporting two high-end server lines. Sure, Sun has deals in place with Fujitsu for some system and chip work and with TI for chip manufacturing to offset a portion of the R&D load. That doesn't make this stuff free or even cheap. ®
When will 32nm matter?
Sun's current processors (US-IV+ and US-T1) are 90nm designs. SPARC64-VI (dual-core Olympus chip) is a 90nm design. Niagara2, Rock (2008), and Jupiter (SPARC64-VI+, also 2008) are 65nm designs. Beyond these are any 45nm SPARC chips, which I would expect would be Niagara3 and Rock2. Realistically, these are 2010 systems offerings. Then beyond any Niagara3 and Rock2 would be 32nm processors, (Niagara4 and Rock3?), in 2012 or later.
Sun has plenty of time, and plenty of options. First of all, TI did not say it was exiting the 32nm chip market, only the 32nm fab business. So any fab partners would have to be using process compatible with TI's semicondutor design. These TI partner fabs may be able to fab 32nm Sun SPARC chips.
Second, Fujitsu is a major semiconductor fab, and to my knowledge, has not announced much around 45nm. Perhaps Fujitsu could skip 45nm and pursue 32nm.
Intel is certainly an option, as is AMD, and even IBM. I think Freescale has some fabs as well.
And as for your comment about 32nm vs. 45nm, one process revision does not a fight make. Intel always brings out its next process on desktop systems, while its more powerful Xeon server processors lag one generation. Similarly, Intel's more powerful Itanium lags Xeon by one generation (Itanium is still 90nm, and will not go to 65nm until Tukwila in 2008). Also, AMD's very capable Opteron is still 90nm, and competed with 65nm Xeon very well. Opteron is going to 65nm and true quad-core with Barcelona, only about six months ahead of Intel's 45nm Xeon transition.
One last thing, the very powerful IBM POWER5+, Intel Montecito dual-core Itanium, and Sun UltraSPARC T1 processors are all 90nm designs, and crush any 65nm processors out there.
High-end server processors are different than desktop PC or laptop processors, and always will fall a generation or two behind the state of the art.
SUN to use Intel fabs - I don't think so
TI laid off 500 engineers because their 45nm process is completely developed and TI has decided not to pursue 32nm.
The recent EETimes front page article sums it up nicely:
You need to have a certain revenue (EETimes estimates 13 billion dollars PER YEAR) to be able to amortize a 32nm fab and literally only a handful of companies (Intel, Samsung and the consortium led by IBM) are expected to ever get a positive ROI out of a 32nm fab. (Note the absence of AMD from this list - the EETimes article implies that it is very unlikely that AMD will ever build a 32nm fab of its own.)
Intel has no interest to nurse an architectural competitor on its figurative breasts - after all, SPARC is one of the reasons why Itanium is not taking off as hoped. If at all, Intel would try to rent out only EXCESS fab capacity (i.e. amid one of the periodical industry slumps), but even then they'd think twice to rent it out to a competitor in the processor industry.
The reason the SUN - TI partnership worked so nicely is because
(1) TI is not a competitor in the 64 bit microprocessor market
(2) TI is (up till now) not trailing far behind with respect to fab technology
With TI gone from the 32nm market, SUN has few, if any, options to find a fab partner who is not a direct competitor.
The logical option for SUN seems to be to abandon SPARC, probably a better option than seeing it (stuck at the 45nm node) becoming outclassed by Intel's and IBM's 32nm processors.
The saying is: "Don't bring a knife to a gunfight." Trying to compete against 32nm processors with a 45nm processor is like bringing a knife to a gunfight.
Sun to use Intel fab?
It looks like Sun's unexpected partnership with Intel came around the same time that TI chose to fire 500 fab engineers. Looks like Intel could be the one cranking out Sun's new CPU's at 45nm and beyond.