According to IBM Research’s blog, a change they made 20 years ago is still helping us today. (One big wire change in ’97 still helping chips achieve tiny scale). 20 years ago is a long time in technology. In 1997 we were talking about CPU speeds of 233Mhz on 180nm chips. IBM’s Deep Blue was operating at a mere 200 million chess positions a second. But we were about to approach one of the first limitations that could stop chip performance continuing to match Moore’s Law. And to prepare for its introduction without delay, IBM Research started looking at one of the alternatives. Copper.
The limitation was Aluminium’s performance as an interconnect within the chips. Aluminium was approaching some challenges as chip process nodes became smaller. To go beyond the next few nodes, you needed a new metal. Copper is a more efficient conductor of electricity and can handle the higher current loads that smaller more powerful chips would require. But copper came with challenges that needed overcoming.
Using a new metal presented challenges including how to make sure it could be placed on the chips, that it remained in the correct place, and that wouldn’t affect the rest of the chip infrastructure. Copper is of particular concern, as unlike Aluminium, it creates deep-level traps in the silicon. These traps capture electrons that flow through the semiconductor, reducing the silicon’s conductivity. The team had produced some test connections and written a research paper in 1993. This process, however, wasn’t a cost-effective for manufacturing. This needed fixing, as IBM wanted profitable chip manufacturing.
IBM Research’s team, led by Dan Edelstein, delivered a process that allowed copper deposition safely on the chip, which prevented it poisoning the surrounding semiconductor and would provide the benefits of higher conductivity than aluminium. Adding silicon-dioxide at the same time as the copper to provide an insulation layer was vital. By 1995, the process was passing and surpassing product quality tests. The manufacturing teams now could consider the costly transformation of production processes. In 1998, production PowerPC chips using copper interconnects delivered 400Mhz performance. This increased speed by 33% over the previous 333Mhz aluminium device on the same process node. This performance increase would continue for many years.
But we are again approaching some of the limits of copper at the current process nodes, and IBM Research is looking at other metals. In future metals like cobalt, nickel and ruthenium may provide chip interconnects.