TREND TWO:

Energy-efficient chips

The graphic processing units (GPUs) that power most AI systems are tweaked versions of designs originally developed for video gaming.

This has led innovators to conclude that chips designed for AI from the ground up could deliver improvements in efficiency. Already, the IEA reports that today’s AI-related computer chips use 99 per cent less power than a standard 2008 model when performing the same calculations. Today, innovators continue to develop new ways of producing efficient chips – not just for AI but for other types of electronics too.

The graphic processing units (GPUs) that power most AI systems are tweaked versions of designs originally developed for video gaming.

INNOVATION ONE:

3D chips for faster, more efficient electronics

As the world’s data demands continue to skyrocket, our information infrastructure must continue improving. Gallium nitride (GaN), an advanced semiconductor material, could play a big role in that undertaking, supporting the creation of high-speed communication systems and new electronics for cutting-edge data centres.

Up to this point, however, integrating GaN into electronics has proven too expensive and difficult to achieve using traditional equipment and processes. Now, researchers from the Massachusetts Institute of Technology (MIT) believe they have overcome this barrier.

In order to maximise its performance, a GaN chip needs to be connected to a silicon digital chip, or a complementary metal-oxide-semiconductor (CMOS) chip. Typically, this is done by soldering the connections between the GaN and CMOS chip. Having the GaN transistor as small as possible is essential to maximise its performance, but this soldering method often limits how small the transistor can be made. Alternatively, a GaN wafer can be connected to a silicon one, but using this much of the material is extremely expensive and wasteful.

Having the GaN transistor as small as possible is essential to maximise its performance, but this soldering method often limits how small the transistor can be made.”

Instead, the MIT team has created a novel and scalable production process that involves building several tiny transistors on the surface of a GaN chip and cutting these out. Then, only the necessary number of transistors are bonded onto a silicon chip using a low-temperature method, reducing the volume of GaN required without compromising the performance of the silicon or GaN.

With its new process, the scientists created a power amplifier for mobile phones, helping it to gain greater signal strength and efficiencies than what’s typically possible with silicon transistors. If the compact hybrid chips (which are less than half a square millimetre in area) were integrated into our everyday smartphones, it could boost call quality, wireless bandwidth, and connectivity, all while increasing battery life.

INNOVATION DATA:

Country: US

Development stage: Research

Contact: mit.edu

TAKEAWAYS:

Gallium nitride (GaN) is an advanced semiconductor material that could facilitate the creation of new, high-speed chips
An MIT team has created a novel and scalable production process for creating hybrid GaN chips, which it tested by creating a mobile phone power amplifier
If integrated into smartphones, the chips could boost call quality, wireless bandwidth, and connectivity, all while increasing battery life

INNOVATION TWO:

Can a new chip design shrink AI’s energy use?

ChatGPT and other large language models (LLMs) are finding new uses every day. But these platforms also use a tremendous amount of energy in their training and operation.

One reason for this energy usage is the fact that, as LLMs grow in speed and capability, they require more and more data, which leads to a massive volume of information being transmitted across data centres. As data centres transmit ever-larger volumes of information at higher speeds, maintaining signal integrity becomes more challenging, requiring more sophisticated and energy-intensive signal processing techniques.

To clean the data, conventional communication systems use blocks known as equalisers, which compensate for signal distortion by flattening the frequency response of the channel, ensuring that the received signal accurately reflects the transmitted signal. However, equalisers use a lot of energy, especially at high speeds.

Now, researchers at the Oregon State University College of Engineering have developed a communication chip based on AI principles that uses half the energy of traditional designs.

Engineering have developed a communication chip based on AI principles that uses half the energy of traditional designs.”

One of the researchers, Ramin Javadisafdar, explained to Springwise that the new chip leverages AI principles directly within the hardware, “allowing it to learn and identify common error patterns in data and correct them intelligently. In terms of functionality, it performs the same role as a traditional equaliser and maintains a similar physical footprint, but consumes roughly half the energy.”

The work was supported by the US Defense Advanced Research Projects Agency (DARPA), the Semiconductor Research Corporation, and the Centre for Ubiquitous Connectivity. Next, the researchers will work on the latest iteration of the technology: an adaptive, trainable chip capable of tuning itself to specific communication systems, further improving efficiency and performance.

INNOVATION DATA:

Country: US

Development stage: Research

Contact: engineering.oregonstate.edu

TAKEAWAYS:

Conventional data centre communication systems use equalisers to compensate for signal distortion
These equalisers use a lot of energy, especially at high speed
Oregon State University researchers have developed a communication chip based on AI principles that uses half the energy of traditional designs