Intel’s recent announcement that it has developed hyper threading technology and will begin shipping Xeon processors in 2002 is being called an important breakthorough for the company by one industry analyst.
Hyper threading or simultaneous multithreading (SMT), as it is usually referred to, is designed to recapture lost system’s resources. Essentially, it allows there to be multithreads of execution, and allows for separate stacks for each individual threads and switching between threads. Jonathon Eunice, the principal analyst for Illuminata in Nashua, N.H., said the feat puts Intel ahead of Sun Microsystems’ SPARC processors.
“A SPARC processor doesn’t even use SMT,” he said. The technology is not about adding vast amounts of resources to the system but about recapturing resources that have already been lost, he said. He added that as far back as Intel’s release of the Pentium Pro, Intel has done an exceptional job of integrating new technologies into their mainstream business.
In processors today, what typically happens is if a processor issues an instruction, such as a load memory instruction, the memory sub system cannot respond immediately. In a best case scenario, a load instruction would be issued, and on the next instruction the data from the next load would be used. But as Eunice pointed out, even with high-speed caches on processors today this is not the case.
“The memory sub system typically cannot get the data in just one clock cycle. Often times the processor has to stall (or) do nothing until that data becomes available.” He said even with the most advanced compilers, the decode and fetch units are unable to fill the slots, leading to frequent stalls. SMT, he said, is a technique that’s moving beyond the operating system level, some context information is moved into the processor itself and letting the processor know of several different threads of execution it can work on.
Doug Cooper, country manager for Intel in Toronto, said the technology is designed to improve workstation performance, and tries to take advantage of application parallelism. He said instructional level parallelism has existed for some time, and it allows for more than one instruction at a time within the chip. What they have accomplished is making a single CPU look like two CPUs to an operating system. “We’ve fooled it so that it thinks a single CPU actually looks like two virtual processors. You can actually run two threads on the same CPU.” When switching between threads, the state of the thread is maintained by the operating system in the processor, he said.
The move by Intel to introduce the technology is innovation driven, said John Enck, the senior research director for Gartner in Loveland, Colo. He said their market will initially be servers and workstations but added that the operating system is an important aspect of the equation. “For the right combination of operating systems application there is a real performance boost, around 20 to 30 per cent for a regular application. It does require an operating system that can handle multithreading well, like Windows 2000 or XP.”