19-06-2021 | By Robin Mitchell
AI chip company Mythic has recently announced the release of their latest AI chip, the M1076. What hardware challenges does AI face, how do Mythic products overcome this, and what features does the M1076 offer?
AI has the ability to dramatically change any application thanks to its ability to not only infer from incomplete data but also learn from its mistakes. Just because an AI system today is incapable of differentiating between two brands of similarly looking cars does not mean that it is useless. Over time, the AI can train itself to better identify cars, and with enough data, would be able to identify different cars that may even look identical to human eyes.
While AI may have wide-reaching applications, actually running an AI system using off-the-shelf hardware is highly inefficient. This inefficiency comes from how AI neural nets work, and how different they work from a standard computer program.
A typical computer program that uses if statements and loops have each instruction executed one after the other. An AI, however, consists of many nodes that are connected to each other which carry weights. When signals propagate through a neural net, they are affected by the connections between nodes (hence the weighted aspect), and this changed how a node behaves.
When expressed in mathematical terms, a neural net is ironically very simple: multiply and accumulate tasks. However, unlike a computer program, a neural net can have millions of these trivial multiply-accumulate tasks running simultaneously. This means that typical CPUs are very poor at running AI.
GPUs are an alternative platform for running AI systems as they consist of many multiply-accumulate units all connected together in parallel. However, GPUS are still not as efficient as they can be for executing AI with large parts of the silicon space being dedicated to graphical tasks unrelated to AI. As such, current AI systems are nowhere near as hardware efficient as they could be.
One company that is developing AI processors is Mythic, and their approach to creating dedicated AI hardware is to go analogue instead of digital.
The first computers used by scientists were analogue computers, and these computers used analogue components instead of digital systems to solve equations and produce results. For example, the suspension on a car can be thought of as a second-order differential equation, and the goal of suspension is to remove unwanted jerks and dampen bumps. A second-order differential circuit can be constructed to solve this, and parameters that represent the springs and weight of the car can be used in the circuit. Researchers can see how changing these values affects the output of the circuit which represents how the car reacts to bumps.
While analogue computers are extremely bulky and complex, it turns out that they are highly ideal for AI systems. Simply put, a neural network has weighted nodes connected to summing and comparative units, and this is something that analogue circuits are very good at doing.
As such, Mythic (who develop AI hardware), are using this approach to create programmable neural nets. While their designs integrate digital systems for taking data in, configuring the AI engine, and providing outputs, they use analogue computing to execute the weighted nets. An array of DACs feed tunable resistors (which are programmed by SRAM), and the resulting current flow through the resistors is converted into a current that is then read by an ADC. Just like that, the network can evaluate a weighted node in a fraction of the time needed by digital systems.
Recently, Mythic announced their latest AI chip, the M1706, which utilises their analogue computing technology to accelerate AI execution. According to Mythic, their new AI processor uses 10 times less energy than those which use GPUs commonly found in System-on-Chips, and this makes the M1076 potentially useable in edge IoT applications.
The M1076 can provide up to 25 trillion operations per second while consuming 3W and can be scaled up for tackling larger applications. For example, Mythic has also released a PCIe card using 16 M1076 devices providing a total power consumption of 75W and a total computation operation of 400TOPS.
Applications using the M1076 not only benefit from a lower power consumption at a higher computational rate, they can also reduce their external component count thanks to the internal memory used by the M1076. The M1076 also supports integer operations between 4 and 16 bits depending on the needed resolution, supports PCIe with up to 2GB/s, and can store up to 80 million weighted parameters.