Celebrating the sixth anniversary of the Opteron server CPUs that put it ahead of Intel in 2003, AMD claimed that the design for 6-core 'Istanbul' Opterons is working out so well that they'll be on sale in June.
Originally the manufacturer had stated a "second half" launch date for the chip. Istanbul offers 30 per cent more performance for the same power than current quad-core Opterons, says Pat Patla, AMD vice president and general manager of its server business, "but in the jump from the 2009 timeframe to 2010, you'll see the biggest Opteron jump in a single generation."
Article continues below
That comes with the 8 to 12-core 'Magny-Cours' (say it out loud) Opteron 6000 which will be available in the first quarter of next year. The next step is by the 12 to 16-core 'Interlagos' Opteron the following year, when AMD switches to 32nm production. Patla says "we are looking for a similar type of performance gain from 2010 to 2011 – the Bulldozer core, our next big implementation of x86 going into server products, goes to 11".
Single threaded future and DirectConnect 2.0
Although AMD is sticking to single threading instead of Intel's two-threads-per-core HyperThreading, by putting four 12 to 16-core CPUs in a system, AMD can match the number of threads. "We think we're pretty well covered in the 48-thread to 64-thread environment over the next couple of years," claims Patla, adding that "12 cores looks like the ideal amount for our architecture as long as it's balanced with the right amount of memory and HyperTransport".
That combination is the next version of AMD's processor architecture, DirectConnect 2.0, which will be in Magny-Cours. This has a four-channel memory controller instead of two channels, twice as much cache meory and nearly double the HyperTransport bandwidth, with four links instead of three, plus new features for AMD-V virtualisation support and AMD-P power management.
DirectConnect was part of AMD's original innovation in Opteron, moving the memory controller from the chipset to the CPU and using direct connections between cores. Some of these features have appeared in Intel chip architectures since, although AMD's chief marketing officer Nigel Dessau didn't name Intel when he commented that "it's ok to clone the past but the future has got to be created".
Patla also contrasted the fact all the variants of the 4000 and 6000 Series Opterons have the same set of features with Intel offering more features on more expensive processors in the same range; "Our technology is available in every one of our series, on every one of our sockets.
Istanbul is launching in two, four and eight sockets at same time; they have the same feature set so IT managers don't have to make a compromise if they want to have power efficient systems." That's important for virtualisation, whether it's a large data centre where virtualised workloads need to be migrated from one server to another regularly, or a small business with a second, cheaper server that can take over without any changes if the main server has problems.
To prove the point, AMD demonstrated live migration between Istanbul Opterons and models dating back as far as 2003, without any interruption to video the server was playing.
The range includes the HE (Highly Efficient) 55 watt processors and the EE (Extremely Efficient) 40 watt processors as well as the 105 watt SE (Special Edition) series. The EE range is designed for the data centres that power cloud computing services, says Patla, but he says Opteron also saves power in the two-processor servers that make up 70 per cent of server sales; "that's where thousands of servers get deployed and you can really have an effect on power efficiency."
Part of that is down to extending the CoolCore technology in the less-than-successful Barcelona Opterons to the memory cache explains Senior Product Marketing Manager Brent Kerby; "Barcelona knows how to expect what kind of instructions it will get and shuts of parts of itself down to save power.
In Istanbul we've extended CoolCore to the L3 cache. The cache can understand whether it's writing or reading and reading takes less power than writing. We're putting those kinds of algorithm in the chip itself."
And what will all those cores mean for what servers can deliver? It's not just about running programs faster, Patla says. Take search. "You're going to be able to not only analyse searches but do analytical work across searches [to find out] why the search is occurring; not what you're searching for but why you are searching for it.
You can analyse those trends – and then what's triggering those trends; you can be doing analysis on what's creating the demands behind the searches."