““First of all, IC technology has made progress in the past 50 years, making many things that were impossible or impractical now very feasible.” “Secondly, we found some new problems that are worth solving, and the number of these problems is solved. The solution is not enough to solve-specifically, this requires higher performance, while also significantly reducing power consumption.”
Claude Shannon (Claude Shannon) is widely regarded as the “originator of information theory.” During his studies at the Massachusetts Institute of Technology (MIT), he used the differential analyzer developed 10 years ago. The differential analyzer was essentially the first general-purpose analog computer, and Shannon’s experience with this machine had a profound impact on later works such as the mathematical theory of communication.
Around the same time, Shannon’s contemporaries made great strides in digital computing systems. In the next 20 years, these systems will be fully realized, and the digital signal processing revolution in the 1970s and 1980s Reached its peak.
But now, about 80 years after Shannon introduced the differential analyzer, analog computing, and even analog signal processing, seems to be making a comeback.
Gene Frantz, Vice President of Octavo Systems Engineering, explained: There are two important reasons for this.
“First of all, IC technology has made progress in the past 50 years, making many things that were impossible or impractical now very feasible.” “Secondly, we found some new problems that are worth solving, and the number of these problems is solved. The solution is not enough to solve-specifically, this requires higher performance, while also significantly reducing power consumption.”
As Moore’s Law draws to a close, more and more application fields will feel the demand for low power consumption and high performance. From mixed signal processing (MSSP) for neural network workloads to dynamic system simulation using differential equations, it has aroused new interest in analog computing.
Back to the basics of physical simulation
In order to illustrate the most basic advantages of analog calculation, consider processing an analog signal described by a set of differential equations. Since continuous time does not exist in the digital computing paradigm, the digital computer must sample the input of each clock cycle to generate a sampled signal. This may lead to a lot of calculations, which will bring chain effects such as higher latency and greater power consumption.
Compare the massive parallelism of analog computers. The analog calculation circuit can be configured as a basic unit (adder/subtractor, multiplier, integrator, fan-out, nonlinear function, etc.) to solve the differential equation in question, and then continuously sample the entire input signal (Figure 1), and Instead of breaking the input into continuous tasks.
The analog computer contains an integrator, a multiplier, a function generator, and other circuit blocks. Continuous time circuits form blocks capable of generating arbitrary functions.
The analog computing chip executes differential equations much faster than the digital computing chip, and its power consumption is much lower. Although the disadvantage of analog computers is that they must scale the system linearly according to the size of the equation you want to solve, their parallelism means that their capabilities and performance will also be expanded.
We support these claims with limited benchmarks. Figure 2 compares the power consumption and time to solve the Van der Pol equation on Sendyne’s Apollo IC and 25 MHz Texas Instruments MSP430 MCU.
Comparison of power consumption and delay between Apollo analog computing IC and Texas Instruments MSP430 single-chip microcomputer when implementing Van der Pauer equation
The Sendyne Apollo integrated circuit is a 4x4mm2 general-purpose analog computer based on CMOS technology. The chip was originally developed by a team of researchers at Columbia University and contains 16 analog integrators, using a 1V circuit to generate output, and consumes only microjoules of energy. In addition, it also contains a dedicated ADC to minimize conversion costs.
John Milios, CEO of Sendyne, said: “If you are dealing with analog signals, you can skip the steps from analog to digital and back to analog. This is obviously an advantage.” There are some special adcs that basically do not do any conversion. Or consume any energy, unless the input signal changes, so you won’t have any significant energy loss. “
He added: “If you consider complex problems that require a lot of parallel operations and millions of times, you will see a very significant benefit.”
The fusion of neural network and analog signal processing
Outside the field of differential equations, analog-based arithmetic logic units (ALU) are also becoming more and more popular in the MSSP field.
Frantz explained: “Each time we reduce the size of the multiplier by half, for example, from a 32-bit multiplier to a 16-bit multiplier, the performance is increased by about an order of magnitude, and the same ratio of the multiplier each time, the power consumption is also reduced by one. Orders of magnitude. Therefore, going from a 32-bit digital multiplier to a 1-bit analog multiplier can improve performance by orders of magnitude while reducing power consumption by orders of magnitude.”
“At the same time, the number of transistors required for multiplication has been reduced from tens of thousands to tens.”
Neural network is an application area that originally used analog signal processing. Such as keyword recognition and certain types of image processing, you can trade lower simulation accuracy for the power and performance improvements it provides.
Dr. Jeremy Holleman, Chief Scientist of Syntiant: “People are increasingly realizing that the workload of machine learning is different from the applications designed by previous processors. It requires strong computing power, memory-centric, mainly deterministic, and can Tolerate a certain degree of accuracy. All these factors give play to the advantages of analog calculations.”
Syntiant is an artificial intelligence semiconductor start-up supported by Bosch, headquartered in Irvine, California. This company focuses on processing deep learning algorithms in resource-constrained systems such as wearables, headsets, remote controls, and sensors.
Marcellino Gemelli, Director of Business Development at Bosch Sensortec, said: “Our whole idea is to stay in the analog domain from the sensor front end to the neural network processing. You can also understand that the signal conditioning before the neural network processing occurs on the ADC.”
The architectural goal described by Gemelli is simple: try to keep the digital core dormant. Another artificial intelligence startup Aspinity in Pittsburgh, Pennsylvania has developed a reconfigurable chip called RAMP, which can accurately replicate digital signal processing tasks in an analog environment (Figure 3).
Aspinity’s Reconfigurable Analog Module Processor (RAMP) is located between the sensor and the digital system components to save a lot of energy.
Aspinity CEO Tom Doyle said: “All the sensory data is naturally analog, but we digitize all the data automatically, and then process it downstream of a digital core. But if you apply it to analog transistors, This can be done efficiently and accurately.”
“What we can do is to monitor very specific frequency changes accurately enough early in the signal chain,” “So after the sensor, Aspinity’s RAMP core can view all raw analog sensor data. When we detect sound or glass fragments, We will wake up the ADC and DSP to run the FFT to obtain all the data needed to determine what to do next.”
According to Doyle, applications such as glass break detection and voice activation use RAMP technology to save power by 10 times or more.
Are we back to the future? Not exactly
Although the potential of analog technology as an alternative or complementary processing technology is obvious, it is still lacking in the commercial market for a long time. First, there is limited information on how analog circuits respond to the effects of temperature and aging. Another issue that needs to be considered is that digital interfaces are now commonplace.
Gemelli explained: “In order to make full use of the advantages of low power consumption and small size molds, the signal will need to be more upstream at the analog level, which in turn requires sensor suppliers to implement architectural changes. In the analog domain, a major redesign of the sensor front end is required. “
What ultimately drives these redesigns is the wider use of analog computing technology. In this regard, it would be helpful to provide development tools that access analog circuits from a digital environment, and progress is being made in this area as the analog hardware goals are achieved.
Of course, I use “become available” to express the truest meaning. Sendyne’s chips are the product of academic research, while the first-generation products of companies such as Aspinity and Syntiant have just entered the market.
It is undeniable that our ability to increase speed and density in the digital domain is rapidly diminishing. At the same time, our demand for computing power is growing exponentially.
What does it take to go back to the future with simulation calculations?
Frantz said, “Someone needs to invest in this high-risk opportunity and create the first solution. I estimate that the cost of a new computer architecture is between $100 million and $1 billion.
“Indeed, the risks are great. There are very few people who can do this. But the rewards are huge.”
The analog computing system is very effective in calculating the differential equations related to the dynamic system. However, mapping these equations to analog circuits embodies its own complexity.
The biggest problem may be the lack of programmable simulation targets, which hinders the emergence of simulation computing tools that can automatically or at least simplify the reconfiguration of analog integrated circuits to meet new or different differential equations.
Thanks to the collaboration of Edwin Howard Armstrong, a professor of electrical engineering at Columbia University, and Sara Achour, a PhD candidate at the Massachusetts Institute of Technology (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL MIT), this situation is beginning to change.
After Tsividis’ team completed a general-purpose analog computer, Achour developed Legno, a compiler that successfully locates the device through the USB interface on the chip. In fact, it is believed that the Legno tool made by this graduate student is the first compiler to successfully target programmable analog hardware.
Legno accepts first-order differential equations written in Python, allowing users to define state variables, expressions and variables they wish to observe. The compiler uses these inputs to synthesize configurations for devices such as Sendyne Apollo and ensure that all physical constraints of the target are met.
Achour: “Suppose I throw a ball from 50m away. If I can’t drive 50 microamps on a wire, I can’t simulate it.””[Legno]The compiler will automatically reorder the system. Suppose we want to shrink it five times, and now you have mapped the position of the ball to between 0 and 2 microamperes, which can be represented by the device.
“Maybe I have defined a speed and a position (in a high-level language), and I want to observe this position. On output, the compiler will generate a set of configurations for each block on the simulated device, and it will also determine which ones are enabled Line, in addition, “it will ensure that all working ranges on the equipment will be taken seriously, taking into account manufacturing changes, it will comply with any frequency limits, and will try to amplify all signals to obtain a good signal-to-noise ratio (SNR).
The Legno compiler uses differential equations expressed in Python to properly configure analog circuits. Shown here is a dynamic system specification (DSS) for a damped oscillator.
“All these observable variables can be read through the pins on the chip, so[Legno] It will ensure that devices outside these connections are routed so that they can be measured, “Basically, it solves the problem of auto-scaling. You can rearrange the differential equations and boundary variables.
She metaphorically said: “It is basically doing “rich” work for senior electrical engineers.”
Of course, for embedded engineers, Legno is not the only available tool for outputting differential equations. It is a very good thing for future analog computing developers. Because there are many other tools that can output differential equations, it is not difficult to imagine that these equations can be directly mapped to the simulation calculation target through a specially compiled solution like Legno.
Programmable analog with fast forwarding
Legno is open source, it has been evaluated by members of the engineering community and copied Achour’s results. Obviously, most of these users do not have programmable analog chips, so Archour provides 10g oscilloscope data, which can be used for simulation.
This is the first iteration of the Legno compiler, and it is currently being finalized. MIT students want to continue to improve this tool.
Achour said: “I am very interested in developing it in the near future, because one thing I do with this compiler is to analyze all the analog blocks on the chip. So, I have these mathematical models to describe each multiplier. The behavior of the instance.”
“The interesting thing is that we can create random simulators so that those who cannot use programmable analog devices can still experiment on them.”