CMOS running out of gas, new effort looks for scalable replacement, ICCAD keynoter says
By Michael Santarini, Senior Editor - November 7, 2007
SAN JOSE -- The CMOS FET, the main fabric and process used for the production of most semiconductors today, will run out of gas by the year 2020, so multiple research projects -- directed by the Semiconductor Research Corporation’s (SRC) Nanoelectronics Research Initiative -- are diligently looking for new materials and semiconductor fabrics to allow semiconductor scaling to continue well into the second half of the century.
In a keynote here at ICCAD Jeff Welser, director of the SRC Nanoelectronics Research Initiative (NRI) for the Semiconductor Research Corporation and executive at IBM Almaden Research Center in San Jose, described how four main research centers under the NRI are trying to find a replacement for CMOS.
“NRI was create a few years ago and is looking at post-CMOS, essentially what comes next?” Welser asked.
Welser said that up to the mid-1980s, most semiconductors were created using bipolar processes, while CMOS, a new process at the time, was mainly being used for low-cost products. However, in the early 1980s, as semiconductor vendors needed higher performance and more transistors from silicon, bipolar began to show significant problems in scaling -- with the biggest problem being power consumption. Indeed, as designs demanded a greater number of features (thus more transistors and performance), bipolar processes started to consume way too much power and generate too much heat, which increased chip failures.
So in the mid-1980s, when it became apparent that bipolar processes would no longer scale well, semiconductor vendors started to move their designs into CMOS processes. CMOS has served the semiconductor industry fairly well as a replacement as it continued to scale well in terms of transistor counts (in line with Moore’s law) and performance. That is, it scaled well until a couple of years ago, when power density once again started to become primary issue. Around the 90-nm CMOS process node, transistors started to leak power when they were not in use, forcing vendors targeting high performance designs, initially MPUs and GPUs, to turn to multi-core architectures rather than staying with single processor architectures and increasing clock rates, which increased leakage, heat and thus cooling requirements. The issue has become progressively worse as CMOS processes have scaled down to 65-nm and 45-nm so even customers targeting bulk CMOS processes must contend with power management. Welser noted that as power becomes a larger issue, heat dissipation will also become a bigger issue, as was the case with bipolar processes.
But Welser said in the 1980s when it became evident bipolar would no longer scale physically and economically, the industry already had a backup process, namely CMOS, in place. Today, however, there isn’t an evident replacement for CMOS.
Welser said that nanowires, nanotubes and nanoarray crosspoint register diodes are a just a few of the more popular structures researchers are looking into to extend the life of CMOS. “What these all have in common in my mind is they are not certainly silicon FETs but they behave a lot like silicon FETs,” said Weiser. “What NRI is looking at is what’s beyond that.”
To do that, NRI has three research centers diligently working on it. The first, based in California, is the Western Institute of Nanoelectronics (WIN), which is researching gated spin wave devices, using spinning electrons to do logic, as well as functions such as reads and writes. The second center, based in New York, is the Institute for Nanoelectronic Discovery and Exploration (INDEX), which is also looking at spintronics and graphing and related issues. Welser said the third research center is the Southwest Academy for Nanoelectronics (SWAN) based in Austin, Texas, is also looking at new devices and even nanoscale magnetic device that uses structures he likened to magnetic dominos.
Welser said ultimately the CMOS replacement must meet fairly strict criteria. It must of course have a lower power density than CMOS, reduce process variability, reduce parasitics, and have great potential for future scaling. The CMOS replacement must be also compatible with CMOS, he said, noting that the industry must pick the replacement process between 2010 and 2015, so that equipment makers and manufacturers will have time to iron out issues in the fabric and develop manufacturing lines. Welser didn’t address the issue directly but designers using the new fabric would also seemingly need new tools and methodologies to design circuits or whatever they will be called on the new process.
Welser refused to speculate as to which fabric or process will ultimately be the winner in replacing CMOS. He was also emphatic that CMOS will not go away entirely in 2020, as there are bound to be uses for it for many years to come.
However, he pointed out that in the meantime, to maintain the scaling of CMOS up until 2020, the semiconductor industry will need to come up with better ways to stem transistor leakage and better ways to cool chips, which is bound to become a bigger issue as CMOS scales below 45-nm. But that’s an issue for different research effort, like the Gigascale Research Center, which is looking for ways to prolong the life of CMOS.