Intel's many-core version of Itanium to skip 45nm
And then comes Kittson
Intel will apparently keep selling the Itanium processor through at least four more code-names. Yes, world, the company today revealed the follow-on chip to Montvale, Tukwila and Poulson. Say hello to Kittson.
How many cores will Kittson have? When will it arrive? Will it be made of fish or fowl? Intel refuses to say. It's simply uttering the two syllables and nothing more.
Intel, however, did provide a number of key updates to the nearer-term Itanium products, during a press briefing today.
For one, Intel confirmed some things you already know. Montvale - the refresh of today's dual-core Montecito - will ship at "the end of this year."
Next will be Tukwila - a four-core processor that supports 2 software threads per core and that ships at the end of 2008. Unlike its predecessors, Tukwila will sport an integrated memory controller and Intel's so-called CSI interconnect that replaces the aging front side bus. These additions will let customers create a common chipset for their Itanium-based and Xeon-based systems.
After Tukwila, you'll find Poulson. The most interesting thing so far about this part is that it will be crafted on a 32nm process. That means that Intel will jump from Tukwila at 65nm straight to 32nm without passing 45nm and collecting $200. That should help Intel catch up IBM on the high-end chip manufacturing front.
Again, Intel refused to say how many cores or threads will arrive with Poulson. Although, the chipmaker did confirm that we're talking about a brand new microarchitecture with the chip and "many more cores and high performance threads." The chip sounds an awful lot like what Sun plans to ship at the end of 2008 in the form of the 16-core Rock.
Even though we're years away from Poulson, Intel has started work early with its "don't fear the recompile" campaign. Company officials promised that existing software will work well on the new chip. You'll also see gains by optimizing software for certain new instructions but won't need a recompile as you do for today's Power6 chip, if you're looking for absolute best performance.
From a hardware standpoint, the microarchitecture revamp should be relatively painless. The many-cored Poulson parts will slot into Tukwila-based servers.
Speaking of Tukwila, Intel wants to be damn sure that you know about the Double Device Data Correction technology shipping with the chip. DDDC, as it's known on the street, allows a server to detect and disable two DIMMs, which Intel say is a huge plus over today's systems that can only deal with a single DIMM failure. Multi-core chips and virtualization software demand more memory, so you want a RAS feature or two to handle DIMM meltdowns.
Let's see. What else?
Oh, yes. Itanium is a smashing success apparently. To prove this, Intel keeps rolling out a slide showing that Itanium-based server sales equal about 63 per cent of total SPARC-based system sales and 54 per cent of Power-based system sales. We're still trying to figure out why you'd want to boast about that.
After all, Sun, IBM and HP split the high-end server market just about evenly. So, all things being equal, Itanium should be just about even with the rival chips, since it has most of HP's business. Add to that the entire Itanium ecosystem - SGI, Fujitsu, Hitachi, Bull and whoever else - and we're talking about a RISC butchering, right?
Overall, today's announcement confirms that Intel and Sun are heading in one, mostly similar direction - many low powered cores - while IBM will go it alone with speedy dual-core chips. We fail to see how IBM will be competitive come 2009/10 unless Power7 brings with it a major change in design philosophy.
So far, Intel has said that 32nm production will begin in 2009, and the company is sure to lead with mainstream PC and server chips with the new process. That should put Poulson in the 2010 area.
This leaves us wondering if Sun won't pull off a minor miracle and actually make its mainstream UltraSPARC Rock line the most competitive high-end server gear on the market. Sun has pegged late 2008 for Rock shipments. You have to believe it will miss that date if history is any kind of teacher. If, however, Sun can hit the target, it would appear to take a one-year lead in the massively multi-core race. ®
Will Poulson ever see the light of day?
Changing Poulson a to 32nm could mean it does not require a mature process because it is no longer an Itanium instruction set, but really an x86 chip. Or it could be a deliberate plan to put Poulson far enough out so it so nobody will notice when Intel pulls the plug on Itanium.