Advanced Computing in the Age of AI | Thursday, March 28, 2024

IBM Commits $3 Billion To Future Chip, Systems Research 

If the rumors are right, and they usually are on these sorts of things, then IBM is looking to get out of the chip manufacturing business and has been shopping around its two foundries in Vermont and New York. But that does not mean Big Blue wants to get out of doing basic research in chip design and manufacturing. In fact, IBM thinks chip and system design research is fundamental to its future and that is why it is putting up $3 billion over the next five years on a wide variety of technologies that could end up in systems someday.

IBM spends about $6 billion a year on research and development, with the exploratory work done by its IBM Research unit and the productization of that fundamental science being performed by the labs in its Software Group and Systems and Technology Group. IBM is by far the largest patent holder in the world, and the company has amassed over 2,500 patents and patent applications that cover future chip technologies that could be commercialized to drive transistor sizes down to 7 nanometers and below. This is an important dimension because many semiconductor industry players think the current silicon wafer methodologies and CMOS technologies used to etch chips will not scale easily below 7 nanometers. IBM, knowing this and looking a decade or so out to the future, is investing in the fundamental research that it believes will be necessary to keep us on a Moore's Law curve even if we are not necessarily on silicon wafers.

This $3 billion in funds is for research, not for development, according to an IBM spokesperson, and it is not incremental increase in spending above and beyond the current levels but rather a concentration of spending in research projects that IBM is betting will be necessary to build the systems of the future. Averaged out over the next five years, this semiconductor and system research will represent around 10 percent of total research and development spending. In its annual report this year, Ginni Rometty, IBM"s CEO and chairman, was clear, saying that "IBM will remain a leader in high-performance and high-end systems, storage, and cognitive computing, and we will continue to invest in R&D for advanced semiconductor technology." This re-allocation of funds, which is presumably larger than what was previously allocated for such projects (IBM would not say), is IBM making good on that commitment.

Note that IBM did not say it remained committed to being a chip foundry, for its own devices or those of its foundry customers. IBM is currently making Power8 processors using its own 22 nanometer processes in its East Fishkill, New York foundry. This process mixes various technologies, such as copper wiring, strained silicon, silicon-on insulator, and high-k gates, to make the transistors on the Power8 processors. The Power8 chips have fifteen layers of metal on top of the silicon wafer and cram 4.2 billion transistors on a twelve-core die with 96 MB of on-chip eDRAM L3 cache; the chips can run at between 2.5 GHz and 5.5 GHz, depending on the voltage applied.

IBM is not worried about what it needs to the future Power8+ and Power9 processors, it looks like, but what happens after that. If history is any guide, then IBM will create Power8+ chips using a 14 nanometer process and Power9 chips using a 10 nanometer process. This is something that IBM will not, of course, confirm, particularly given the fact that it might be selling off its foundry operations before the year ends if the rumors are right.

But Supratik Guha, director of physical sciences at IBM Research, can talk about the challenges of shrinking circuits and where big issues will crop up and require a lot of fundamental science to be done to solve them.

"We are at 22 nanometers now, and I think that everybody more or less agrees that silicon will run out of steam in a few technology nodes," Guha explains to EnterpriseTech. "We think it will be around 7 nanometers. I believe that the path to 14 nanometers is quite clear, and the path to 10 nanometers is also reasonably clear. I think the point of inflection will come around 7 nanometers, where we will have to think of a replacement for silicon, and at IBM we have been looking at this for quite some time."

Guha says that IBM believes that there are several possible candidates as the material foundation for chip manufacturing in the post-silicon era. These include graphene, carbon nanotubes, piezotronics, and III-V compounds, so called after their columns in the periodic table. (Column III starts with boron and has aluminum, gallium, indium, and titanium while Column V starts with nitrogen and includes phosphorus, arsenic, and antimony.) Carbon nanotube transistors and III-V compound semiconductors have moved out of the fundamental research phase but are not proven technologies yet and are far from being commercialized.

"Carbon nanotube transistors, at those types of dimensions, outperform any other materials, and this has been experimentally demonstrated," says Guha. "The question now is how do you make a technology out of it, how do you make transistors without defects, how do you make it a drop-in replacement for silicon."

IBM has also done research in conjunction with the US Defense Advanced Research Projects Agency, the University of Wisconsin, and Penn State on piezotronics, which create very low-power switches that are based on mechanical, rather than semiconductor, properties of a material as it responds to electric current. With piezoelectric materials, if you squeeze them, they generate electricity and conversely if you apply current to them they deform. These piezoelectronic materials are also notoriously hard to manufacturer. (You can learn more about piezoelectronic devices in this electrical engineering presentation from Columbia University, and you can see more about the piezoelectronic transistor in this presentation from principle investigator Dennis Newns from IBM Research.)

The reason IBM is investing in these chip technologies today is to have them ready for when silicon does run out of gas. Moving from research to commercialization takes time, and Guha knows from personal experience having works on the high-k gate technology that is used in the Power8 chips and has been adopted, in one form or another, in the chips used in about half of the smartphones sold these days. High-k metal gate got its start at IBM Research in 1997, and it replaced the silicon dioxide used to gate transistors with other materials that had a high dielectric constant (that's where the k comes from). At small transistor sizes, silicon dioxide leaks current, high k materials based on hafnium do not. IBM started working with universities at that time to see how different oxides might perform as gates, and by the early 2000s, IBM Research was pretty confident that it might be a technology that could be used. It then tapped the Microelectronics Division to start making experimental circuits based on high-k gating, and by 2004 it showed that the technology actually worked. Several years later, high-k gates were part of the chip making process in East Fishkill. That took on the order of eleven years to move from research to product.

The $3 billion that IBM is putting towards research is not just going to chip manufacturing, but also to fund work in silicon photonics, an area that IBM has been working on for the past dozen years already. The idea is to use light coming right off the chips to link components in a system together; light signals consume less energy than pushing electronics through copper wire. IBM will also invest in quantum computing – which Guha said his peers at IBM Research still reckoned was in the early research phase and nowhere near being commercialized, despite the noise being made by D-Wave and Google. Quantum computers use the superposition property of quantum mechanics to create qubits, which have several states at the same time, rather than the electronic binary bits of our computers, which can either be on or off, a 1 or a 0.

IBM is also allocating funds for neurosynaptic computing, which involves emulating a brain, more or less. The goal, says IBM, is to build a system with ten billion artificial neurons with a hundred trillion synapses that only occupies two liters of space and burns only one kilowatt of power.

EnterpriseAI