Cores times three
The chips will sport three times as many cores as today's Shanghais, as well as twice the number of memory channels, 3.3 times the memory speed, 1.9 times the HyperTransport bandwidth, and 2.2 times as much cache memory. The twelve-shooter is expected to offer more than twice the performance of the Shanghai chips that are shipping today. (When you do the math, the Magny-Cours chips will have more than 25 times the oomph of the original Opterons.)
And because of the tweaks in the Magny-Cours electronics, Patla says AMD will be able to cram a whole lot more virtual machines on a processor - something server buyers want to do a lot these days. Magny-Cours will include AMD-Vi, which virtualizes I/O, among other new features.
AMD is not talking clock speeds on either the Istanbul or Magny-Cours chips, but it is hard to imagine the speeds deviating much from the 2GHz to 3GHz range of the Barcelona/Shanghai quad-cores.
Another interesting bit about AMD's announcement today was that it is changing the way it makes server platforms. AMD will deliver two different server platforms in 2010, which will target different parts of the server market and smear some of the distinctions between single-, dual-, and quad-socket servers that are separate items today. These changes will create two different Opteron processor families instead of the current three: 1000, 2000, and 8000 series.
Starting in 2010, AMD will offer Opteron 6000 Series Magny-Cours processors, which will be aimed at the biggest and heaviest workloads and which are designed for two-socket and four-socket servers. These chips will use AMD's G34 chipset and will have four channels of DDR3 memory and up to 12 DIMMs per socket. The Opteron 6000s will come with nine different SKUs in standard, HE, and SE variants.
The target thermals for these are 80 watts for standard Magny-Cours parts; Patla didn't say where HE and SE parts would be. Magny-Cours is expected to come in eight and twelve core versions, by the way, and will use a 45 nanometer SOI process. These chips and their related servers are aimed at server consolidation/virtualization, database, and HPC workloads. The server platform using the G34 chipset goes by the code-name Maranello.
Also in 2010, the Opteron 4000 Series of Magny-Cours chips, which will make use of the C32 chipset and which will be aimed at single-socket or two-socket servers. The C32 chipset will support up to four DDR3 DIMMs per socket and will have two memory channels. The Opteron 4000s will come in nine SKUs as well, but this time there are standard, HE, and EE parts and no SE parts.
The servers based on the Opteron 4000 chips will be aimed at HPC, cloud computing, and infrastructure workloads such as print, file, email, and web serving. The Opteron 4000 series chips, code-named Lisbon, will come in four and six-core variants and will feature higher clock speeds that are possible because the lower core counts allow for more clocks in the same thermal envelope. The Opteron 4000 platform goes by the name "San Marino."
In 2011, AMD will move to a 32 nanometer process and rev the Opteron 6000 series with 12-core and 16-core variants code-named Interlagos, and at the same time, the Opteron 4000s will get six-core and eight-core kickers code-named Valencia. These chips will be based on a new Opteron core named Bulldozer, and they will plug into the existing Maranello and San Marino platforms.
The 16-core Interlagos chip - which is probably two eight-core chips sharing a package, but AMD did not say - will have about 47 times the performance of the original Opterons on floating point work and about 37 times on integer work. That's about a 40 per cent boost in integer performance and 80 per cent on number-crunching compared to Magny-Cours chips with a dozen cores.
And looking out beyond that, AMD is working on a future set of chips that will plug into a next generation of AMD platforms in 2012 with a rev in 2013. Patla and Dessau did not say anything more about these processors and their related platforms, except that AMD did not believe that simultaneous multithreading - Intel's implementation is called HyperThreading, and Nehalem and Dunnington chips have it - was necessary.
"We think we're pretty well covered in the 48 to 64 thread area," Patla said, referring to what happens when you plop four Magny-Cores or Interlagos chips into a four-socket box.
One more thing: AMD says that neither Broadcom nor nVidia will make chipsets that compete with the G34 and C32 chipsets, or any future processors that come from AMD, but they are expected to continue to sell and support their existing Opteron chipsets. ®