This month issue An Independent Journal Dedicated to the Advancement of Chip - Scale Electronics September - October 2000 Email the editor

Harvey Miller Contributing Editor

In 1994, the 100 MHz Pentium was on its way to dominating the microprocessor market. Intel's revenues in 1993, when the 66 MHz Pentium was introduced to replace the 50 MHz 486, were less than $9 billion. This year, Intel's annual revenue rate is $32 billion.

Credit the 20% annual growth rate to a grand strategy that includes:

• Proprietary microprocessor names instead of clonable numbers

• Expansion into manufacture of chip sets, packaged with microprocessors in cartridges

• Motherboards

This strategy has translated into market leverage and more control of a fragmented personal computer market (accounting for $27 billion of Intel's revenues). It's been a great seven-year ride, and Pentiums were the main vehicle.

Fast-forward to May 23, 2000, Las Vegas, Nevada.

At the 50th Electronic Components & Technology Conference, an important paper ("High Performance Package Designs for a 1 GHz Microprocessor") by Intel engineers described the Pentium package evolution from 800 MHz to 1 GHz.

Most observers consider 1 GHz beyond the needs of personal computer users today, but servers and workstations are another matter. Intel's penetration of these $100 billion+ markets is under 50% compared to its domination of the $200 billion+ personal computer markets.

Server markets require ever-higher speeds and 64-bit buses from their microprocessors, and Intel has two answers almost ready to go: the 64-bit Itanium and the Pentium 4 coming out at 1.5 GHz. So the 1 GHz Pentium should provide some hint about Intel's future packaging directions.

As of July 1, 1 GHz microprocessors are rare. Neither Sun Microsystems UltraSPARC IV nor IBM Microelec-tronics' POWER 4 has yet emerged-although maybe they will by the time you read this in September.

AMD has announced a 1 GHz Athlon in a CBGA, but we haven't even heard of any plans for a 64-bit AMD microprocessor and question whether AMD could afford its development without an alliance.

Success at 1 GHz and above depends on the package, even more than the silicon. That's why the Intel work deserves our respect. (So does IBM's HyperBGA package using PTFE and C4 technology.)

Two Packages

Intel's evolution to 1 GHz follows two established models: (1) a PBGA that Intel designates OLGA (organic land grid array) single-edge connector cartridge (SECC2) attachment and (2) an FCPGA to be plugged into a Socket 377. They have two features in common: flip-chip die and laminate substrate with blind and buried microvias.

The OLGA package was designed for the processor first; consequently, I/O signals were placed on the die in locations enabling straight routing out to the package lands and then across the SECC2 substrate to the edge connector. The FCPGA had to plug into a socket designed for a previous generation microprocessor, which created many compatibility issues that had to be resolved.

The FCPGA is the less-efficient package because of high inductance associated with socket and package pins. Its availability was mandated by the need for backward compatibility, but we probably won't see many PGA packages in the GHz domain after this one.

The FCPGA required 12 decoupling capacitors and a six-layer laminated PC board (which includes four build-up layers on a FR5 laminate structure.) A fan was mounted atop the package. The leadcount was limited to 370 due to socket limitations.

The BGA package used a four-layer BT laminate structure with tighter design rules (not cited in the ECTC paper by Hasan, Sarangi, et al.) The BGA structure permitted 495 solder ball sites for ground/voltage plus signal. The entire cartridge is passively heat-sinked.

The authors point out that a host of package design items that impact package cost, directly or indirectly, exist. These include I/O timings, signal trace impedance, number of routing layers, power delivery, thermal performances and silicon and system level interface.