AMD pulls forward six-shooter Opteron cannon
Guns for Intel on server turf
AMD is picking up the pace for its six-core, Istanbul family of Opteron processors, saying it will ship the chips to paying customers in May with server OEM partners shipping in June.
That brings the Istanbul chips ahead by several months and gives AMD a chance to leapfrog Intel in the two-socket server space, where Intel has just launched its quad-core Nehalem EP Xeon 5500s. It will also provide parity or better with Intel's six-core Dunnington Xeon 7400s for four-socket servers.
AMD has been trying to prove that it has gotten its Opteron mojo back after flubbing the launch of the Barcelona quad-core Opterons in 2007, a delay that cost AMD lots of revenue just as Intel was swinging its Xeon fists at full force in the market.
The follow-on Shanghai quad-core Opterons made their debut last November, somewhere around two months early and without any bugs. At the end of January this year, AMD kicked out some faster Opteron parts, right on schedule and including a bunch of low-power Highly Efficient, or HE, Opteron parts rated at 55 watts instead of the standard 75 watts.
And just today, AMD put a dozen new Shanghai's in the field, including new 105-watt Special Edition parts which hit 3.1GHz clock speeds as well as a bunch of other faster standard and HE parts and two new Extremely Efficient (EE) ultra low voltage parts that are rated at 40 watts with 2.1GHz and 2.3GHz clock speeds.
Today is a big day for AMD, with this being the sixth birthday of the Opteron processor and also Earth Day, which plays a bit into the whole energy efficiency theme that AMD has been using for marketing leverage against Intel since 2003. Between 2003 and 2008, AMD has moved from single to quad-core chips and delivered roughly ten times the integer and floating point performance in the same thermal envelope at the system level, according to AMD chief marketing officer Nigel Dessau.
"We're very proud of Opteron, and it is good enough to copy," Dessau said, taking a jab at Intel's Nehalem, which bears some strong resemblances to the Opteron family architecture in areas such as on-chip memory controllers, point-to-point interconnect, 64-bit memory addressing, power saving features and integrated virtualization. "But as someone once said to me, 'It's OK to clone the past, but the future has to be created.'"
To that end, AMD and its foundry spinoff, GlobalFoundries, have been working to move the six-core Istanbul chips ahead as part of a broader plan to get a new set of server platforms to market that AMD believes will better address the needs of the market. The Istanbul processors, which are made using a 45 nanometer SOI process like the Shanghai chips are, is expected to deliver about 30 per cent more performance within the same power envelope, and will hit the streets in May for both two-socket and four-socket and larger machines.
And because of the confidence that AMD has after it rejiggered its design and testing processes in the wake of the Barcelona bug, the company took Istanbul from tape out to production on one take, which Pat Patla, general manager of the server and workstation chip business at AMD, said was the first time the company had done that. This could be risky. We'll know about July or so.
These six-core Istanbul chips will support HyperTransport 3 interconnects, which have more bandwidth, and apparently some of the Shanghai models announced today will also support HT3 links. I was unable to find out which ones at press time. AMD is not talking about clock speeds for the Istanbul chips, but did say that integer performance for the chips would be about 16 times greater than for the original Opterons and floating point performance would be just under 14 times that of those ancient Opterons. The Istanbul chips will have three HT3 links with an acceleration technology called HT Assist and will go into the same sockets used by Barcelona and Shanghai Opterons in two-socket and four-socket variants.
Moving ahead Istanbul means that AMD can also move ahead its high-end Magny-Cours twelve-shooter, which is two Istanbuls put in the same package, side-by-side, with four HT3 links coming out of the chip package. Patla says that Magny-Cours is being sampled to OEM partners right now, and they will get seed units for testing and validation in the second half of the year in preparation for a launch in the first quarter of 2010. The Magny-Cours chips will include a new suite of virtualization and power management technologies that are being called AMD-V 2.0 and AMD-P 2.0, respectively.
Next page: Cores times three
TeeCee: the cores will have their own RAM
I'm pretty sure that, even with existing AMD multi-core processors, there's more than one group of cores, each with their own dedicated memory interconnects and RAM. AMD seems to quite like this sort of design. It appears likely that each chip in the Magny-Cours package will have its own set of memory interconnects, just like if it was two 6-core processors.
re: re: n=6?
ISTR that round here recently there was a report on some research showing that there are indeed diminishing returns on multi-core CPUs* as the core count increases. They'd come up with 8 as a limit beyond which handling core contention for resources starts to consume a significant amount of the grunt that the extra core(s) are there to provide.
I guess the 12-way Magny-Cours is AMD's way of sitting with their fingers in their ears and going "La, la, la, we're not listening" to that one.
*There's a world of difference between multiple CPUs (with their own RAM, cache, IO channels and such) and multicore CPUs (with their unfortunate limitations in these areas).
@Tim, Re: For the average PC
These are big, expensive, multi-socket processors. They are most definitely not aimed at the average PC, and I rather suspect they are not aimed at runnig a single multithreaded task. This sort of box would be ideally suited to be a virtual server host, for example. There isn't really a situation where you can have too many cores and too much IO bandwidth for that sort of system.
That said, software bloats to consume available resources. I have no doubt that in the near future we'll see desktop applications which take full advantage of a 64bit operating system's vast address space, and require half a dozen cores with a gig or two of ram apiece in order to run smoothly. I expect they'll do useful things like let you write documents and add up columns of figures.