A paper published in Nature Electronics argues that resistive memory-based circuits could approach the accuracy of digital processors in specific tasks, with potential gains in speed and efficiency, particularly in the field of AI. However, the results still need to be validated on larger systems by Andrea Monti – Originally published in MIT Technology Review Italia
Research published on 23 December 2024 in Nature Electronics documents a theoretical result that, if confirmed and generalised, could have a significant impact on the global competition for supremacy in supercomputing and the development of hardware infrastructure dedicated to AI.
Specifically, the article, the result of collaboration between scientists from the Institute for Artificial Intelligence and the School of Integrated Circuits at Peking University and the Advanced Innovation Centre for Integrated Circuits at Tsinghua University, describes the possibility of achieving levels of precision comparable to those of digital processors in the analogue computation of matrix systems.
Analogue computing is a way of performing calculations in which numbers are represented by physical quantities (e.g. voltage, current, electric charge or resistance) instead of sequences of 0s and 1s, and operations are then performed by measuring the variations in these physical quantities.
The advantages, in some specific areas such as signal analysis, are greater speed and lower energy consumption, but at the cost of less accurate results. Because of this limitation, analogue computing is of little use in areas such as AI, which would benefit greatly if this obstacle could be overcome. This is what Chinese scientists claim to have done.
To understand the nature and consequences of this discovery, MIT Technology Review Italia met with one of the authors of the article, Dr. Zhong Sun, who conducts his research at Peking University.
In which field of research do you carry out your academic work and what links do you have with Italy?
I am currently an assistant professor at Peking University. I returned to China to work at Peking University in 2020. Before that, from 2016 to 2020, I was a postdoctoral researcher at the Politecnico di Milano.
My research activities focus on RRAM — random access resistive memory. During my PhD, I worked on the physical functioning of RRAM devices, then when I went to the Politecnico di Milano, I worked on in-memory computing using RRAM devices with circuits designed to perform Boolean operations and matrix calculations.
In 2019, we designed what, to our knowledge, was the first circuit capable of reliably solving matrix equations in a single pass. This research forms the basis of the work we have just published in Nature Electronics.
If you had to summarise the “before and after” of this research, what is new about it?
We have designed what is probably the first circuit architecture to solve linear algebra operations involving matrices using analogue computation in a highly reliable manner.
Essentially, the process consists of “incorporating” the equations into a physical system by repeating the calculation using an algorithm that significantly increases the accuracy of the results compared to previous approaches. Previously, the main limitation of analogue computing systems was their low precision, which made them unsuitable for applications that required higher precision.
We believe we have solved this problem because our tests have shown, in a very simplified way for now, that it is possible to achieve precision comparable to that of digital computing.
Have you carried out performance benchmarks, comparing your analogue chip with a digital chip, at least at the proof-of-concept level?
We have carried out a comparison based on assumptions, so the results do not come from more extensive experimental measurements. However, the results appear consistent, albeit within the limits of the current restrictions and pending experimental validation on larger integrated arrays.
How did you design the benchmarks?
The reasoning started from the observation that, specifically in the analogue computation phase, the time taken by the circuit has, we can say, an approximately constant latency, so the time required to perform the calculations is comparable for matrices of different sizes. Starting from this assumption, we measured the time required to perform the calculations on smaller matrices and used this result to extrapolate the throughput for solving 128×128 matrix equations. Based on this hypothesis, we calculated that, in some applications, the resulting throughput and energy efficiency of our chip could improve by 2 to 3 orders of magnitude compared to equivalent digital processors.
Of course, at the system level, these results are not entirely independent of other variables, such as matrix size, but if confirmed, they would make our chip a competitor to “pure” digital processors — i.e., those not integrated with ours.
What results did you obtain?
The prototype uses an RRAM chip on a Printed Circuit Board (PCB) with off-chip peripheral circuits. Given the size of the prototype array and the available devices, we have demonstrated the possibility of solving 16×16 matrix equations with hardware, achieving 24-bit fixed-point precision, which is equivalent to 32-bit floating-point.
Speaking of the engineering side of this chip, how much do you think it can be miniaturised?
The chip is very small because the most significant advantage of RRAM technology is that it can be built in very thin slices. The density can be very high because a crossbar-array architecture is used, although we will have to wait for the final integration into a complete system to determine the actual size.
Do you think this chip can actually be turned into a product or integrated into large-scale industrial products?
Absolutely, in fact, we are moving in that direction. As I mentioned in our article, only RRAM is present on the chip, while the other components are installed on the PCB. So the next step will be to integrate all the components, including our arrays, operational amplifiers, digital-to-analogue (DAC) and analogue-to-digital (ADC) converters, onto the chip. We expect the result to be a small, low-power, stand-alone processor.
Do you think it can be integrated into all types of hardware or will it only have a few specific applications?
It is not a general-purpose processor because it will play a key role in scenarios involving complex matrix calculations, such as AI, scientific supercomputing, weather forecasting, signal processing and 6G wireless communications.
Do you expect analogue processors to replace digital ones?
This chip is different from digital processors. It is designed as a coprocessor/accelerator integrated into a conventional system. For this reason, we will include a DAC in the front end and an ADC in the back end. In this way, digital inputs are converted to analogue, processed quickly and efficiently, and then returned to the output in digital form.
Does changing the physical operation of the processor also require the development of a new programming framework, or can you rely on traditional tools?
I don’t think it’s necessary to change high-level languages, but we do need a dedicated compiler and an interface to the Instruction Set Architecture (ISA).
What is the relationship between analogue computing and quantum computing?
Unlike quantum or photonic approaches, we use conventional electronic circuits compatible with the widely used Complementary Metal-Oxide-Semiconductor (CMOS). This means that our analogue chip could be produced using existing industrial processes and reliable technologies. This means that it is possible to turn to factories with the capacity to produce RRAM, without having to rely on the special technologies used to build quantum computers.
Furthermore, quantum computing addresses classes of problems that are partly different from those requiring matrixcomputation, which is one of the most important tools in the development of AI today.
Finally, our approach may be easier to implement, facilitating the production of the related hardware.
What message do you think is right to convey to the public about this technology?
The key message is that analogue computing deserves a chance. Because of its low precision, no one has ever given it a chance to compete with digital chips. But now that we may be able to overcome this limitation, it is important to reconsider the role of analogue computing in modern applications.