April 4, 1996

Computer Engineers Develop Prototype of Next Generation Microprocessor

The next generation of computers, predicted to hit the market in about two years, will be very fast--finishing one instruction every five hundred millionth of a second--but the necessary technology could leave behind the software running on current 486-based machines.

Dr. Thomas M. Conte, assistant professor, and Sumedh Sathaye, research assistant and Ph.D. candidate, both in electrical and computer engineering at North Carolina State University, have found a way to solve the problem through developing their own processor, TINKER.

Their research into VLIW (very long instruction word) technology, which is supported by Altera, AT&T, IBM, Intel Corporation and the National Science Foundation, led to a best paper award for Conte and Sathaye at the International Symposium on Microarchitecture last November.

According to Conte, a few groups of researchers started companies in the early 1980s to develop VLIW for supercomputers, but the market was too limited and the technology impractical at the time. The companies failed, and very few researchers since have investigated VLIW. Conte and his TINKER group are filling the void by re-examining the early problems with the technology and solving them.

Computers perform tasks by executing instructions, and if hardware were designed to perform functions in parallel, instructions could execute simultaneously. VLIW encodes independent instructions by combining them into a long "word" of instructions, hence the name "very long instruction word."

Unlike current microprocessors, VLIW has the capability to run the number of instructions in parallel needed to process a command every five hundred millionth of a second, but it is not compatible with current software. Therefore, upgrading to a VLIW computer would hit the consumer's pocketbook twice: once for the machine and again for all the new software it would need.

The programs running on current machines will not run on VLIW because current machines expect instructions to run in a sequence one after another. Conte said that these programs do not include the necessary information to work efficiently when instructions can run in parallel.

VLIW computers are able to process the information so quickly because the work is done in the compiler, a function in the software that translates a program's high-level computer language into the low-level machine code the computer requires, rather than in an extra piece of hardware.

"You don't want that hardware because a computer's like an assembly line where the slowest worker in the line determines how fast the product comes off the line," Conte said. "Performing work in the compiler has several advantages. For example, it removes complexity from the hardware, reducing costs and power consumption."

VLIW is more efficient, then, but what about the problem of incompatible software? With TINKER, Conte and Sathaye have found a solution.

"We figured out how to make computers speak the same language with the smallest amount of work possible and still get very high performance," Conte said. "That meant we had to do some tricks in the hardware to reduce the cost (the amount of space required on the chip or the length of time needed) of doing that translation and also some tricks in the software."

The trick they devised is to have the operating system, such as Windows 95, perform a little of the work and translate the code somewhat when the program is first executed. They have also discovered a way to store this translation information for later use. After a program runs once in their system, it will run more quickly the next time it is used because the translating has already been done.

The researchers' findings could one day have an impact on new products produced by the computer industry.

"Changes to architecture and implementation of modern processors are occurring at a rapid pace, and these changes are highly disruptive to the installed software base," said Mike Schlansker, a scientist with Helwett Packard in Palo Alto, Calif. "Technologies to promote object code compatibility are very important to future products."

Conte acknowledges that several problems must be solved before the technology is embraced by the computer industry, so he and his team of nine graduate students are addressing not only the compatibility issue but others as well.

For example, software programs will require a type of tuning for VLIW, called profiling, prior to release. According to Conte, profiling raises at least two issues. First, a program that has been profiled for VLIW runs two to 30 times slower than normal. Also, testers might use the program differently than consumers, so the software's performance once it is sold is not predictable.

The TINKER group is tackling both problems. They have devised a method to make profiling unnoticeable to the user, and the group is finding ways to postpone the tuning until after the consumer has bought the software.

/ News Index / News Archives Index /