Close-up of IBM’s Artificial Intelligence Unit chip. Image: IBM
The hype shows that artificial intelligence (AI) is already everywhere, but in reality, the technology that drives it is still evolving. Many AI applications are powered by chips that are not designed for AI – instead, they are based on general-purpose CPUs and GPUs made for video games. That mismatch has led to a flurry of investment — from tech giants like IBM, Intel, and Google, as well as from startups and VCs — in the design of clearly engineered new chips. for AI workloads.
As technology improves, business investment will inevitably follow. According to Gartner, total revenue from AI chips reaches more than $34 billion in 2021 and is expected to grow to $86 billion by 2026. In addition, the research firm says, less than 3% of central servers data in 2020 includes workload accelerators, while more than 15% is expected in 2026.
IBM Research, for its part, only Launched Artificial Intelligence Unit (AIU)a prototype chip specifically for AI.
“We’re running out of computing power,” IBM said. AI models are growing exponentially, but the hardware to train these massive things and run them on servers in the cloud or on advanced devices such as smartphones and sensors have not advanced rapidly”.
Also: Can AI help solve education’s big data problems?
The AIU is the first complete system on chip (SoC) from the IBM AI Hardware Center for AI Research designed to run AI deep learning models for businesses.
IBM argues that the “workhorse of traditional computing,” also known as the CPU, was designed before deep learning was born. While CPUs are good for general-purpose applications, they are not so good at training and running deep learning models that require batch parallel AI operations.
Jeff Burns, director of AI Compute at IBM Research, told ZDNET: “There is no doubt in our minds that AI will become the fundamental driver of IT solutions for a long time to come. “It will be delivered across the computing landscape, across these complex enterprise IT infrastructures and solutions, in a very broad and pervasive way.”
It made the most sense for IBM to build complete solutions that are universal, says Burns, “so that we can integrate those capabilities into different computing platforms and support supports many of the enterprise’s AI requirements.”
The AIU is an application-specific integrated circuit (ASIC), but it can be programmed to run any type of deep learning task. This chip has 32 processing cores built using 5 nm technology and contains 23 billion transistors. The layout is simpler than the CPU, which is designed to send data directly from one set of computers to the next, making it more energy efficient. It is designed to be as easy to use as a graphics card and can be plugged into any computer or server with a PCIe slot.
To conserve energy and resources, AIU promotes approximate calculation, a technique that IBM developed to trade off computational precision for efficiency. Traditionally, calculations have relied on 64- and 32-bit floating-point arithmetic, to provide a level of precision useful for financial, scientific computing, and other applications where granular precision is required. However, that level of precision is not strictly necessary for the majority of AI applications.
“If you think about plotting the trajectory of a self-driving car, there’s no precise position in the lane that the car needs to be in,” explains Burns. “There’s a bunch of places in the lane.”
Neural networks are fundamentally imprecise – they produce a probabilistic output. For example, a computer vision program can tell you with 98% certainty that you are looking at a picture of a cat. Even so, the original neural networks are still trained with high precision arithmetic, which consumes considerable energy and time.
Also: I tested an AI art generator and this is what I learned
The AIU’s approximation technique allows it to reduce from 32-bit floating-point arithmetic to a bit format containing a quarter of the information.
To make sure this chip is really popular, IBM has focused on more than just hardware improvements. IBM research has focused on platform model, with a team of between 400 and 500 people working on them. In contrast to AI models that are built for a specific task, foundational models are trained on a wide set of unlabelled data, creating a resource that resembles a database. huge. Then, when you need a model for a particular task, you can retrain the underlying model using a relatively small amount of labeled data.
Using this approach, IBM intends to address different verticals and different AI use cases. There are several areas where the company is building foundational models – those use cases include areas like chemistry and time series data. Time series data, simply data collected at regular intervals, is important for industrial companies that need to observe how their equipment is performing. After building foundational models for several key areas, IBM can develop more specific, vertical services. The team has also ensured the software for the AIU is fully compatible with the IBM-owned Red Hat software stack.
