For over 50 years, the exponential shrinking of circuit components on chips predicted by Gordon Moore has allowed all sorts of modern wonders, from personal computers and mobile phones to social media and smart cars.
Gurtej Sandhu, a senior fellow and vice president at Micron Technology in Boise, Idaho, is behind key innovations that have driven Moore’s Law forward all these years. And now as the electronics scaling law declines, he’s looking at advanced technologies that can keep accelerating computing. An IEEE Fellow, Sandhu won the 2018 IEEE Andrew S. Grove Award for his contributions to silicon CMOS process technology that have enabled extreme scaling of dynamic random access memory (DRAM) and NAND flash memory. He also holds the seventh-highest number of U.S. patents.
Sandhu spurred the development of atomic layer deposition for high-k dielectric films to make DRAM devices. He also invented a semiconductor patterning process called pitch doubling, which drastically shrunk NAND flash memory. His chemical vapor deposition process for metal barrier layers is still used to make DRAM and NAND chips.
Sandhu spoke to us about memory technology developments at Micron and why there’s no better time to be a computer engineer.
IEEE Spectrum: What are some key innovations from Micron Technology that have kept Moore’s Law going all these years?
Sandhu: A lot of the innovations at Micron were in processes and techniques to enable new materials and complex structures. And many of them were quietly adopted in the memory space long before anybody else in the industry had heard of them.
For DRAM, for example, Micron had 3D capacitors. We were working with high-k dielectrics long before they were adopted in logic. We enabled pitch-doubling patterning in 2005, long before others started talking about it. We were using 3D transistors in the early 2000s, long before the 3D FinFET was adopted in logic.
The latest innovation is a technology called 3D XPoint Memory. This is a transistor-less architecture where memory cells sit at the intersection of word lines and bit lines. The memory layers can be stacked. Each cell holds one bit of data, and each can be written or read individually by changing the voltage sent to it. The data can be written and read in small packets, and it’s faster than other nonvolatile memory available today.
IEEE Spectrum: Do you see the decline of Moore’s Law as a positive thing or a negative?
Sandhu: Moore’s Law is dead in some cases. For example, in NAND, planar scaling died a few years ago, so we went to 3D NAND. Moore’s Law’s decline was inevitable. The question is, what do we do next? It opens up new opportunities for innovation.
I think opportunities are even more for a young engineer now than when I started because there are more diverse vectors one can be thinking about to push innovation forward.
IEEE Spectrum: What are some innovations you are spearheading at Micron to continue the pace of electronics development?
Sandhu: We have to look at system-level solutions. People think of DRAM and NAND as bottoms-up innovations—that they were invented and then we built all these systems around it. Nothing could be further from the truth. DRAM succeeded because Intel decided they needed low-cost memory and they made system-level changes to enable DRAM. NAND is the same story. Steve Jobs was looking for cheap solid-state memory which spurred the adoption of NAND Flash.
The industry may even have to look at novel techniques like bringing processing into memory. Everyone talks about the proximity of memory and computing. Some of the computing may actually need to go into memory to ensure innovations.
IEEE Spectrum: Does that mean an overhaul of computer architecture?
Sandhu: We’ve been computing the same way for last 50 years, and it’s not sustainable. There’s a bottleneck of getting data in and out of the processor. It’s only going to get incrementally better. If you’re looking for 10x to 100x or more improvements, we will have to change how we compute.
Instead of being compute-centric we need to be memory-centric. Right now data sits in memory, you move it into the processor for operations, and then send it back. Even a simple thing like A + B takes 25 cycles of the clock. But instead of moving data, you can just send instructions into the memory and say do A + B and store it. This is much faster, plus it requires very little energy to move data locally over nanometer scales.
The issue is, you have to change coding and software and the way you think about computing. Just a memory company or processor company can’t do it. Everybody has to participate at some level to enable this.
IEEE Spectrum: Are people in the industry considering this seriously?
Sandhu: It’s been discussed for a while, and now I see a slightly positive trend. When I talk about it in an open forum no one stands up and says it’s a bad idea. People know this is the way it has to be, but it’s very complex to change the way we do things.
New apps being designed will likely present opportunities. With AI, for example, even car companies have to design chips for autonomous driving. With IoT and edge computing, everything from software and firmware to coding has to be redone anyway. When you start from scratch, there’s a possibility to implement completely new system architecture.
IEEE Spectrum: There is a more general trend in computing of hardware specialization. AI accelerator chips are one example. How does that influence the memory and storage technology Micron is researching?
Sandhu: When I started in DRAM, we just made chips and sold them and hoped somebody would use them. The first killer app was the PC from Dell and Compaq. In the last 10 years, we have huge markets in enterprise and mobile. Automotive apps are now coming. We’re already optimizing solutions. You can consider that as hardware optimization for different market segments. For AI, clearly chips are going to be special-purpose ASICs (application-specific integrated circuits). Memory is going to be a huge component of that. Diversification of the market has already happened. It’s going to happen even more.
IEEE Spectrum: Micron is funding research on storing digital data in DNA sequences. What’s the promise of DNA data storage?
Sandhu: It started with scientific curiosity. Just to know if the physics is even right. We found that the physics really works. Then that led to thinking about what it can do in terms of practical applications. Data explosion is driving the explosion in the DRAM and NAND memory market. Wait until AI kicks in. We’re just at the beginning of an exponential curve of data explosion.
Tape drives are used for archival memory, but they have to be rewritten every 10 years. It’s not long-term data. DNA is an interesting medium. Nature has used this for millions of years, and apparently it works, right? Our bodies are made from DNA memory from generations past.
In principle you could have a teaspoon of DNA and the entire world’s info as it exists today could be fit into it. But that’s easier said than done. There are a lot of challenges before you’ll see that happening.
IEEE Spectrum: What are those challenges? When do you see DNA data storage becoming a viable technology?
Sandhu: So far the only thing I’m convinced about is the physics. Everything else is a challenge. How you input data, read it and manage it, and so forth. The Human Genome Project has made millions of times improvement in mapping DNA. That’s the right trajectory. But for the app we are talking about there is another million, if not more, times more improvement needed before it becomes practical. It’s really a pretty tall order before we can get to the point where we can use DNA as a medium to store and retrieve data.
IEEE Spectrum: What does the future hold for memory technology?
Sandhu: We take technology for granted because we don’t remember what we didn’t have 10 years ago. Young kids today have never seen a check. They all do banking from their phones. This phone wasn’t even possible 10 years ago before NAND technology became mainstream. Just imagine what will happen in the next 10 years. It’s an exponential curve.
Memory technology is going to change our lives. It really needs a lot of smart people. We need to energize young kids to get interested in this space.
This post was updated on 22 October 2018.