June 11, 2003 9:00 PM PDT
AMD overhauls transistors, chips
The Sunnyvale, Calif.-based chipmaker is examining how to incorporate a wide variety of cutting-edge concepts--strained silicon, multigate transistors, replacing silicon with metal in key transistor components--to boost the performance of chips that will hit the market in the second half of the decade.
The race to change the basic structure of microprocessors is a matter of survival, said Craig Sander, vice president of process technology at AMD. Chips are increasingly getting smaller and running at faster speeds, but many of the materials and structures used to make processors these days can't be pushed much more without unleashing unintended consequences.
The gate oxide, for example, one of the crucial components of a transistor, is only about five or six atomic layers thick on current chips, Sander said. Further thinning, without creative changes, will cause electricity to leak, leading to lower battery life, excessive power consumption and potentially dangerous levels of heat inside computers.
"More than ever, we are up against power constraints," Sander said. Without manufacturing improvements, semiconductor companies will be "out of business."
Concepts being presented by researchers now will likely start to appear in commercial chips with the dawn of 45-nanometer manufacturing, due in 2007, or in the 32-nanometer generation of chips due in 2009, he said.
AMD's research was outlined in two papers delivered at the Very Large Scale Integration (VLSI) Technology Symposium taking place in Kyoto, Japan this week, one of the premier scientific conferences on the annual semiconductor calendar.
In its first paper, AMD described how it achieved a 30 percent performance improvement in transistor speed by incorporating two changes into experimental transistors: a nickel silicide transistor gate and the addition of a fully depleted silicon-on-insulator layer.
Currently, transistor gates, which let electricity pass from one end of a transistor to another, are made of polysilicon, a crystalline form of silicon. Underneath the gate is an insulating layer called the gate oxide. Thin gate oxides improve performance; however, thin oxides also leak electricity and create traps, which can dissipate performance inadvertently.
Nickel gates, by contrast, improve the flow of electrons, cut down on leakage and can be thicker, making them potentially easier to use for future manufacturing.
"You can have a thicker gate oxide physically but one that will act like a thinner one," Sander said. Current gate oxides are about 16 to 17 angstroms thick, he said, and experimental ones have gate oxides about 8 to 10 angstroms thick (an angstrom is a tenth of a nanometer). A metal gate can make a thicker gate act like one that measures only 4 to 5 angstroms thick, which would be extremely difficult to physically achieve with current materials: the distance between silicon atoms is just over three angstroms.
The silicon-on-insulator layer similarly prevents leakage by removing loose electrical carriers from the chip.
In the second paper, AMD showed how it experimentally boosted performance 20 percent to 25 percent on strained silicon transistors by using the nickel gate. Strained silicon involves physically stretching the silicon atoms apart from one another on select chip layers so electrons can travel more freely and rapidly. Stretching is accomplished by inserting germanium atoms, the melons of the chip world, into layers of apricot-sized silicon atoms.
"This was more than we expected," Sander said. The surprise results may be in part, he theorized, on additional stretching caused by the nickel.
AmberWave Systems, a Cambridge, Mass.-based start-up, worked in part with AMD on implementing strained silicon, but AMD also will begin to work with IBM in this regard as part of a far-reaching manufacturing research project, Sander said. Around 40 AMD researchers are already working out of IBM's labs.
Along with this research, AMD is examining multigate transistors. More gates mean a larger and smoother, flow of current. The company is looking at "finfet," or two-gate, transistors, as well as triple-gate transistors, like Intel's. It is too soon to tell which will work.
"There are trade-offs from a manufacturing viewpoint," Sander noted. Finfets are slightly more difficult to manufacture, but they feature a wide surface area for conducting electricity.
Although multigate transistors and many of the other technologies are inevitable, it's still impossible to tell exactly how these advances will be implemented.
"What we don't know yet is the exact sequence of how these things will play out," Sander said.