A computer's central processing unit, for instance, puts out far more heat than it needs to destroy itself. What's more, the heat is accelerating along with the speed of computers. Each new generation of processor runs faster and generates more heat than any of its predecessors. Things apparently can only get hotter, because no one in the business is predicting an end to the demand for increasingly faster computers.

NEC Computers Inc. of Boxborough, Mass., fell into an unorthodox design after it charged its engineers to come up with a heat sink that would protect a processor running at 600 megahertz. When the company took up the task, it was designing a network server from scratch based on a new version of a Pentium 3 processor. As it turned out, the engineers working on the heat sink did much the same thing, rebuilding the part from scratch.

Because they were designing the device specifically for a system server, the computer that would be at the heart of a network, they had to be certain that the heat sink would be capable of cooling the central processing unit if one of two redundant system fans failed. It is too risky to rely on both active heat sink fans to keep the server's processor cool, because if one fails, the whole system fails with it.

The system contains protective circuits that shut it down when the processor overheats.

A computer-generated model of a heat sink shows the pins arranged in four rows, with their size, for the most part, increasing gradually in width and height.

NEC's engineers had to rethink the problem completely, and their answer turned out to look quite a bit different from anything they originally expected. It also allowed a bit of headroom in the matter of heat transfer. According to Gregg Hamlin, the project manager at NEC, the design doesn't stop at 600 MHz, but is expected to serve in the computers that NEC will build around a new gigahertz chip that Intel plans to introduce later this year.

That chip is still to come, according to Intel. George Alfs, a spokesman at Intel's Santa Clara, Calif., headquarters, said some computer manufacturers are already shipping desktop computers with gigahertz processors, but the Pentium 3 Xeon chip aimed at network servers has yet to be delivered with a billion hertz speed.

Network servers use chips that have large memory cache. In late May, Intel introduced a large-cache chip that operates at 700 MHz. The chip is offered in two versions, one with 1 mega-byte and the other with 2 MB of memory.

A 933 MHz Pentium 3 Xeon chip introduced around the same time has less memory built in, but could be used in the less demanding role of a "front-end" server to a network, according to another Intel spokesman, Otto Pijpker, who works for the company in Hillsborough, Ore.

Flow Modeling

NEC's engineers were working on a tight deadline and, to accelerate development, they added computational fluid dynamics to their resources at hand.

The engineers expected to create a conventional elliptical pin fin design to radiate heat away from the processor chip. They ended up with anything but.

The new heat sink design, which uses an unusual arrangement of different-size pins, was an eventual outgrowth of NEC's effort to shorten its design cycle by tapping the powers of CFD analysis. For the software platform, the engineers selected Coolit from Daat Research Corp. in Hanover, N.H., a CFD software tool specifically tuned for thermal management of electronics.

Jason Welch, a mechanical engineer on the development team, said he received the code by e-mail.

The engineering department had a little more than two weeks to develop an initial design, so it tapped Coolit to create an interim solution. It tested off-the-shelf extrusion heat sinks and when it selected the best for the job, had it machined to the company's final specifications. Computers using this form of the heat sink were on the market for several months.

The machined extrusions were relatively expensive to manufacture, so there was intense pressure to phase them out quickly. The estimated cost was $14 to $20 each, depending on quantity. This approach bought time to develop the final custom die-cast design, Hamlin said.

To shorten the custom design process, engineers sought to quickly eliminate unsuitable design alternatives before they progressed to the prototype stage.

According to Hamlin, the project engineers ran hundreds of iterations in two months, and got hard prototypes for lab testing of only the most promising designs. He said this is the phase of development where most time was saved.

At first, the models were too big to handle comfortably. "The designs reached a million and half cells, and the calculations had to run overnight," Welch said.

The varied size of pins in this NEC version accounts for decreases in cooling capacity as air traverses the heat sink. Each pin is about 1/7-inch thick.

Technical support staff at Daat suggested that Welch work with models on a friendlier scale. The software company's representatives recommended that he break down the problem and create several models, rather than try to solve all the variables of a design iteration at once.

As the design grew more specific, grid selection became increasingly challenging. A fine mesh delivers a more accurate answer than a coarse one, but takes considerably longer to run. Significant time could be saved if fine meshes were used only when absolutely necessary.

Technical advisors suggested that Welch use coarse grids when determining the sensitivity of design changes on critical components. He was also told to use the same approach for design optimization when the main objective was to compare different designs to determine which is best.

Coarse grid analyses would show the correct relative performance of designs, and once he had zeroed in on an optimal design, he could use a fine grid model to develop a quantitative prediction of its performance.

As the engineering team members began to analyze airflow and heat dispersion in Welch's various elliptical pin designs, they observed that the cooling surface requirements were increasing as the air moved across the heat sink. If the requirements were increasing then, they reasoned, the heat sink cooling capacity should grow to match them. So was born the variable pin design.

In a conventional design, the pins present uniform surface areas to incoming air, but in the new design the surface area of the pins varies to allow high-speed air through the entire heat sink. The pins are arranged in four rows, in which the pins for the most part increase in width and height. The varied size compensates for decreases in cooling capacity as the air traverses the heat sink. The farther the air travels through the structure, the more it loses velocity and increases in temperature.

Tiny Pins, All in a Row

All the pins are about one-seventh of an inch thick. Welch said their shape resembles that of airplane wings.

The smallest pins rise 1.47 inches above the heat sink's base and extend along the axis of airflow for 0.45 inch. They are positioned at the leading edge of the heat sink, where airflow is maximum and cooling demands are minimum. Locating the smallest pins upfront reduces the probability of an air dam forming at the leading edge, and ensures air percolates by flowing predictably through the entire depth of the heat sink.

In the second row, the elliptical pins reach 1.57 inches high and 0.59 inch wide. The third row consists of pins 1.67 inches high and 1.62 inches wide. The third row stands at the hottest point of the chip.

The fourth row's pins are the same height as those in the third row, 1.67 inches, but they are not as wide, coming down to 1.28 inches.

The device can rest in the outstretched palm of a hand.

"The design is counter to everything we've ever seen in elliptical pin fin heat sinks," Welch said.

Hamlin added, "In addition to the cross-sectional changes, our pins are aligned in rows, unlike conventional designs where they are staggered like bowling pins."

Tooling for the custom heat sink began in August last year, with the first finished products produced in December. The heat sinks are made for NEC Computers by Kennedy Die Casting of Worcester, Mass., and have been undergoing adjustments and thermal improvements since they were introduced, Hamlin said.

The heat sink outperforms the interim solution by slicing another 15¡C off the processor case temperature. The proprietary die-cast heat sink costs $3 to $4 apiece, about 80 percent less than the machined off-the-shelf alternative.

home | features | news update | marketplace | departments | about ME | back issues | ASME | site search

© 2000 by The American Society of Mechanical Engineers