October 2, 2002
Dean Tullsen is predicting a revolution in desktop microprocessing, thanks to new technology he
helped develop. Intel recently announced that this quarter, it will begin shipping the world's first PC microprocessor
with a powerful new capability called hyper-threading that can greatly enhance a chip's performance. And Tullsen, a
Computer Science and Engineering professor at UCSD, couldn't be prouder. That's because his work and a patent on the
technology (begun at the University of Washington in the mid 1990s) are widely accepted as critical to its commercial
adoption. "For a 5% increase in hardware cost," estimates Tullsen, "semiconductors equipped with this technology can
process almost twice as many instructions in the same amount of time as the same chip without the technology."
Hyper-threading is Intel's copyrighted name for what Tullsen calls "simultaneous multi-threading," SMT for short.
"Threads" are streams of instructions (e.g., programs) that a processor executes. "Many software applications
already break their code up into separate threads, but even those that don't can take advantage of it, as long
as the system has multiple programs to be run," said Tullsen. "An SMT processor executes instructions from these
multiple threads/programs at once, as if they all came from a single thread. The central processing unit (CPU)
duplicates the architectural state on each processor, while sharing one set of processor execution resources."
Essentially, SMT makes one processor appear as two to the operating system, and Intel claims an average improvement
of roughly 40% in CPU resource utilization providing greater throughput and improved performance.
Earlier this year, Intel became the first company to use hyper-threading, when it began building the technology
into its high-end Intel® Xeon™ microprocessors for servers. But in September, Intel announced that it will
begin shipments in the October-December quarter -- ahead of schedule -- of a Pentium 4 processor running at 3.0
gigahertz (3 billion clicks per second) enabled with hyper-threading. It is the first time the technology will be
available on a desktop computer, and it is already supported under Microsoft's Windows XP operating system.
The idea of hardware multi-threading (in less aggressive forms) has been around for years (in research, and
in a few high-end, exotic architectures such as the Cray MTA). But, Tullsen explains, "the SMT execution model,
and the research we did, demonstrated that: (1) you could achieve the performance of multi-threading without
sacrificing single-thread performance; and (2) it could be done with minimal changes to the way we were already
building modern processors. Both of these were critical to convincing the mainstream processor community that
this idea made sense for them. Now we're showing them that they can even get better single-thread performance
with an SMT processor than a processor dedicated to single-thread execution."
For now, Intel is the only semiconductor company building the technology into its chips. But Tullsen predicts
that it's only a matter of time before the technology spreads from high-end chips to all microprocessors. And
as co-director of the Processor Architecture and Compilation Lab at UCSD, Tullsen is already working on SMT-related
improvements, including operating systems, compilers, and the next generation of multithreaded processor architectures.
Links
Dean Tullsen Research Profile
Dean Tullsen Home Page
Story Adapted from Jacobs School Press Release
Full story can be found in the
Jacobs School News Room
Media Contact: Doug Ramsey, (858) 822-5825
