A Programmable Analog Computer on a Chip John Milios¹, Nicolas Clauvelin¹ Abstract --- With the advent of digital computing, analog computers became extinct. Yet during their stint with the space program and many other areas they exhibited several traits that digital computers still cannot match. Their main advantage was high processing speed in continuous time and and reliable crashless operation – albeit at the expense of accuracy (analogous to that of approximate computing). Advances in analog CMOS technology and years of research at Columbia University have made it possible today to reincarnate this technology, adapted to our digital world. We will present the first fully digitally programmable analog computer imple- mented in a 4x4 mm2 chip and capable of solving non-linear differential equations. This first-of-its-kind chip can interface directly to analog sensors and devices in the embedded world and avail its predictions either directly to analog controls or digitally – with orders of magnitude higher speed and energy. Keywords --- analog computing; accelerators; IoT; complex IC; embedded scientific computing Advances in processor technology thus far have have reached a plateau imposed by power and supported the growing demands of embedded heat constrains in CMOS technology. High Perfor- computing. Moore’s law enabled embedded ap- mance Computing (HPC) and, to certain extent, plications of increasing complexity, which in turn desktop computing are responding to this reality opened up new markets. The relatively recent by employing hardware accelerators and dedicat- emergence of the Internet of Things (IoT) is rap- ed hardware like GPUs and FPGAs, as well as ex- idly driving the demand for embedded processing ploring software alternatives such as approximate by shifting processing away from the cloud to the computing[1]. This article has been ac- cepted for publication at the edge, and towards advanced intelligence technolo- In many sensory and control applications in em- “Embedded World Confer- gies like cognitive embedded systems, deep learn- bedded computing, devices are battery-powered, ence 2019” proceedings. ing, edge analytics, embedded vision and speech, imposing additional constrains to already existing www.embedded-world.eu machine learning and others. ones. It is in this context – the pursuit of faster and ¹Sendyne Corp., 250 West Thus far, the semiconductor industry is respond- less power-consuming architectures – that we are Broadway, New York, NY 10013 ing to this increasing processing demand by de- presenting in this paper the idea of modern analog Correspondence to J.M. ploying multicore embedded processors, taking a computing and the first CMOS programmable an- and N.C. similar path to the one it has followed for desktop alog computer chip in the industry. [email protected] [email protected] computing. At the same time, processor speeds EMBEDDED WORLD CONFERENCE 2019 PREPRINT COPY | 1 © 2019 Sendyne Corp. Computing with an analog computer his PhD at MIT and, as a research assistant at What if transistors, instead of just representing the Electrical Engineering department, he was ones and zeros, were capable of representing any operating Dr. V. Bush’s “mechanical differential value between one and zero? It is a question that analyzer” for obtaining numerical solutions to or- many engineering and computer science students dinary differential equations. Shannon attributed have wondered about at some point during their the original idea of a mechanical computer to Lord studies. One answer is that you could reduce the Kelvin but he noted that “at the time it was impos- number of lines required to transmit information, sible, due to mechanical difficulties, to construct as you would theoretically need just one line to a machine of the type he contemplated”. Fig. 1 communicate an infinite number of values. What shows the basic units of the GPAC as conceived by if, in addition, there was no clock, and informa- Shannon. tion was propagating and processed as soon as it Operating the mechanical computers of Shan- was available? It would seem that every piece of non’s era was a tedious task. It involved a lot of information waiting for the next clock pulse would mechanical skills and willingness to be covered in not have to wait anymore and certain operations oil. What was the motivation of these pioneers and would happen faster. why differential equations are so important to sci- In 1941 Claude Shannon, conceived and intro- entists and engineers? Richard Feynman said that duced the model of a “General Purpose Analog “There is only one precise way of representing the Computer” (GPAC)[2]. Shannon was pursuing (physics) laws, and that is by means of differen- tial equations”. Solving the equations describing a physical system is a powerful tool for control, pre- diction and analytics. Most of the complex differ- u ential equations do not have an analytical solution k k + u+ v v that can be derived by pencil and paper, and that onn n n dder n was the motivation for creating machines that can solve them. Digital computers are solving them u u with numerical methods that involve a large num- v X uv v ∫ w=∫udv ber of steps, tries, calculations and repeated loops. per n n neror n Analog computers in contrast solve them by creat- ing a replica of the studied physical system. Solving an ordinary differential equation with an analog computer involves the following steps: Figure 1: Circuit representation of Shannon’s GPAC basic units • Connecting the basic units in such a way that they replicate the target system to be solved • Imposing initial conditions d 2 y dy + a + a y = x • Releasing the system and observe the out- dt 2 1 dt 0 dy(0) put evolving in time dt y(0) Fig 2. illustrates how a set of basic units can be connected to solve the equation: -x-x + y ∫ ∫ 2 d y dy 2 +a10 += ay x × dt dt Observing Fig. 2, it is noteworthy that solving the × equation does not involve any memory accesses and does not require a clock input. At any input a a 1 0 change, the solution appears at the output delayed only by the time required for signal propagation. Figure 2: The system to be solved is replicated by interconnected units of an analog computer These were the ideas that led scientists and engi- EMBEDDED WORLD CONFERENCE 2019 PREPRINT COPY | 2 © 2019 Sendyne Corp. neers to build mechanical and later electrical ana- V IN + log computers for solving mathematical problems. UP - DATA Analog computers dominated in the 50s and 60s, DN TRIGGER employed successfully by NASA in the Mercury, C + C C C C Gemini and Apollo programs. They became ob- - solete with the introduction of digital computers in the late 60s and 70s. Digital CMOS technology V TIG then experienced a tremendous growth, resulting in shrinking geometries and increasing switching C V TW S C speeds, as epitomized by “Moore’s Law”. C Analog CMOS technology also benefited from Figure 4: The Continuous Time ADC utilized in the this growth. Analog devices used in communi- “Nonlinear” units and the interface to a digital proces- sor operates without a clock. Its digital output changes cations, consumer and computing devices also when a change occurs in its analog input became smaller, faster and more accurate. These advancements in analog electronics lead a group It was decided that information would be repre- of researchers at Columbia University to revis- sented by currents instead of voltages. Because of it analog computer technology. As the develop- Kirchhoff’s law, currents are inherently capable of ment of the first generation of electronics lead to the transition from mechanical to electri- cal analog computers, Professor Tsividis’ team k ∫ at Columbia undertook the task and designed the first integrated CMOS general purpose pro- grammable analog computer with 21st centu- Conn no neror ry semiconductor technology . The innovative 4x4 mm2 chip they created [3], which is named Apollo by our company, can work in conjunction F( ) with a digital computer, consuming very little power, and it can solve computational problem at onner per non dderSror a fraction of the time it would take a digital com- Figure 5: The GPAC basic unit set was enhanced by in- puter. clusion of a Nonlinear function and Fanout units. The adder/subtractor does not require a dedicated circuit as current signals are added or subtracted in nodes fol- The path for designing the Apollo analog lowing Kirchhoff’s law computer chip The first step was to define the basic units of com- implementing the adder/subtractor unit. Current putation. The GPAC basic units were suitable signals are carried in differential form to minimize for solving linear ordinary differential equations. effects of common mode noise. As most of the physical problems are nonlinear, One of the most tedious tasks required by ana- a nonlinear unit was designed to complement log computers during the 60’s, configuring of the GPAC’s basic unit set, using a clockless RAM- basic unit, i.e. configuring the interconnections based storage of nonlinear functions. among circuits, was implemented using a switch Data Data C C no np C rer RAM rer C no op t t t t Figure 3: The nonlinear function unit is implemented utilizing Continuous Time ADCs, DACs and RAM circuits [3] EMBEDDED WORLD CONFERENCE 2019 PREPRINT COPY | 3 © 2019 Sendyne Corp. fabric under digital control. This is similar to how digital world, it is an appropriate compromise in configurating is implemented in today’s FPGAs. the analog world, where accuracy is measured by Information exchange with the digital world was percentage points. The Apollo IC, which exhibits accomplished by specially designed “continuous an accuracy of 0.4 %, is the first analog IC of its time” clockless ADCs and DACs[3].