Chip makers will describe plans to deliver server processors with eight CPU cores at the Hot Chips conference at Stanford University this week, though there is some debate about what the products will mean for end-users.
For a long time, vendors boosted the performance of their microprocessors by increasing the clock speed, but concerns about power consumption and heat dissipation have steered them toward adding more processor cores to each chip instead.
On Tuesday, IBM will give the first detailed look at the Power7, a follow-on to the Power6 introduced two years ago for IBM's Unix servers. The Power7 marks a big shift for IBM, moving it from a dual-core design to a new architecture that will be offered with four, six and eight cores, each able to process four instruction threads simultaneously.
IBM has said the Power7 will be available in the first half of next year. The chip will be manufactured on a 45-nanometre process, and IBM has said customers will be able to use the chips in existing Power 570 and 595 servers.
Fujitsu engineers will discuss plans for an eight-core Sparc64 processor, an update to the four-core Sparc64 VII released last July. Fujitsu mentioned the chip, code-named Venus, briefly at the end of its presentation at last year's Hot Chips but has said very little about it. The Sparc64 chips are sold in servers from Fujitsu and Sun Microsystems, though Sun's plans are up in the air since it agreed to be bought by Oracle.
AMD, meanwhile, will give a talk on Monday about using its Magny-Cours processor in blade servers. Magny-Cours is a single-threaded, 12-core chip that combines two six-core processors in a single package, linked by AMD's Hyper Transport interconnect. The chip, named after a French motor racing track, is due for release early next year.
Intel will give an update on its Nehalem-EX chips, which will have eight dual-threaded cores and are due in the first half of next year. There will be no presentation at Hot Chips about Tukwilla, the quad-core update to Intel's Itanium processor that has been delayed several times and is now expected next year.
Also absent from the agenda is Sun's 16-core Rock processor, which had been due to ship later this year but has now reportedly been scrapped. Sun engineers will be at Hot Chips to talk about Rainbow Falls, the third generation of its multithreaded Niagara design that will be a follow-on to Sun's Ultrasparc T2.
The eight-core server parts promise to deliver vast amounts of computing power. "These chips are just incredible in terms of their scale," said industry analyst Nathan Brookwood of Insight64. "Power7 is truly an awesome processor."
But there's some debate about how much today's software - and by extension end-users - will be able to take advantage of the multicore products. Software applications need to be written in a way that makes them smart enough to break up tasks and run them in parallel across the cores.
Gartner analyst Carl Claunch issued a bleak report on the topic in January. Chip makers are driving microprocessor cores "to peaks well above the levels for which key software - including operating systems, middleware, applications and virtualisation products - have been engineered," he wrote.
Claunch said much of today's software has limits that will prevent it from taking advantage of all the cores being offered. And IT users will be forced into upgrading their software more frequently to keep pace with the evolving hardware, he wrote.
IDC issued similar findings in June. With few exceptions, the research company said, "most applications today see a stagnation of performance where there are more than four logical processors".
Others are more optimistic. They argue that much of the software that is restrained by CPU performance today is already highly parallelised, because it has been designed to run across clusters of servers or on large SMP (symmetric multiprocessing) machines that have multiple processors.
Virtualisation and partitioning offer other ways to make use of the cores. An application that doesn't scale well beyond four processors can be run in a virtual machine assigned to four CPUs. IDC expects companies like VMware to develop smarter tools that can automatically assign VMs to additional cores when it makes sense to do so.
"I don't think it's going to become an impediment, and anybody who thinks it is going to become an impediment is probably being pessimistic," Brookwood said.
A bigger challenge, according to some, is writing programs that can spread work across a mix of CPUs and graphics cores. Nvidia, which makes graphics chips, is naturally a big fan of these heterogeneous architectures. "That's where the real cutting edge is," says Kevin Krewell, Nvidia's marketing director and a Hot Chips organiser.
Nvidia's CEO, Jen-Hsun Huang, is expected to address the topic in the opening keynote at Hot Chips on Monday morning. The conference also features a whole tutorial dedicated to OpenCL, an attempt to develop a standard way of programming for mixed CPU-GPU environments.