Chip makers to flex their bits at Hot Chips
Eight-core ante for CPUs
Next week, the IEEE's annual Hot Chips conference will be take place at California's Stanford University, and it looks like there's an eight-core ante for any microprocessor vendor who wants to make a presentation.
On August 24, AMD kicks off Hot Chips 21 with a discussion on blade servers based on the future 12-core Magny-Cours Opteron 6000 processors, due early next year.
Intel follows right after that with a presentation on the forthcoming eight-core Nehalem EX Xeon 7500 chips, which are due to start shipping at the end of this year and which will appear in systems early next year if history is any guide.
El Reg has told you all about Magny-Cours Opterons here and Nehalem EX Xeons here. Intel will also present some details on its next-generation of PC and server cores, code-named Westmere and due to start shipping later this year in PCs.
It's noteworthy that Intel has not seen fit to divulge more details on its much-delayed quad-core Tukwila Itanium server processors at Hot Chips 21 - back in February, the Tukwilas were pushed out until early 2010 by Intel. Maybe we'll see something on the Poulson kickers to the Tukwilas two years from now at Hot Chips 23.
But if history is any guide, maybe not.
Day two of the Hot Chips conference will see Takumi Maruyama of Fujitsu going over the feeds and speeds of the eight-core Venus Sparc64 VIIIfx processor, which El Reg told you about back in May. This Sparc chip is now at the heart of a $1.2bn, 10 petaflops supercomputer being sponsored by the Japanese government, and called Project Keisoku.
Sun Microsystems, which has been muzzled by the impending $5.6bn takeover by Oracle, is nonetheless expected to show up to give a presentation on its Rainbow Falls Sparc T3 processor, formerly known by the code-name KT and sporting 16 Sparc cores and up to 16 threads per core.
We told you all about the Sparc T3 back in June 2008, before Sun freaked out and started peddling itself instead of doing the necessary layoffs to get profitable and sticking to its engineering knitting.
The Rainbow Falls processors are expected to eventually end up in eight-socket boxes, and if so they will be the main reason why Sun - er, Oracle - doesn't feel compelled to continue with the 16-core Rock UltraSparc-RK processors. The rumors are that the Rock chip is dead, something that Sun has not confirmed and very likely will not talk about until the Oracle acquisition closes.
Maybe Rock is just sleeping. Anybody got a mirror?
Anyway, Sun will try to keep the conversation on the Sparc T3s and not say much about Rock.
The company will be giving two presentations on IBM's Power7 processors (code name unknown at this time). One presentation is on the "next generation Power microprocessor" while the other is on the "next generation balanced Power server chip."
Does this mean one variant of the Power7 chip is unbalanced? Or that one version won't be used in servers? Go figure.
Not much is known about the Power7, but IBM has talked a little bit about it. The company confirmed in July that Power7 chips will come in four, six, and eight-core versions and that customers using its high-end Power 570 and Power 595 servers would be able to upgrade their Power6 and Power6+ boxes to the Power7 processors without having to do a push-pull box swap. IBM has also said that Power7 machines will be able to support up to 1,000 logical partitions, almost four times as many as current machines can host.
Real World Technologies reports that IBM will be adding features to the memory controllers in the Power7 chips so it can, if programmers desire it, present the main memory in a cluster of Power7 machines as a global shared memory akin to the type that Silicon Graphics offers with its Altix line of parallel machines and their NUMAlink interconnect for server nodes. Real World Technologies also says it has caught wind that the Power7 will include 16MB of embedded DRAM (eDRAM) instead of static RAM (SRAM) as its L3 cache memory.
There was some talk a few years back that the Power7 chip would plug into Opteron processor sockets. In the summer of 2006 there was also talk that future Sparc processors from Sun would plug into Opteron sockets as well.
While Sun might be able to do this with Sparc T series chips, IBM would probably only be able to do this with a cut-down version of the Power7. And you could argue that what it really needs to do is get the PowerXCell 8i processor, formerly known as the PowerPC Cell chip and used as a math co-processor as well as a motor for game consoles and other electronics, to plug into Opteron-HyperTransport or Xeon-QuickPath Interconnect sockets.
Finally, niche supercomputer maker Convey Computer, which launched its debut product, the HC-1, last November at the SC08 trade show, will talk about how cool its Xeon-FPGA clusters can crack big HPC jobs. Convey has an FPGA math co-processor that drops into a plain old Xeon socket and links to the cache and main memories just like a regular Xeon chip does. ®