By Junko Yoshida

What’s at stake:
In the computing industry, artificial Intelligence has created an unprecedented industry-wide AI wave, on which everyone in the electronics sector wants to surf. This includes photonics. But, is photonic computing mature enough to convince hyperscalers to try it for at least part of their AI processing?

Photonics developers now have a fresh pitch to the AI community: By using photons for AI processing instead of solely depending on electrons, the computing industry can substantially reduce AI-induced energy gluttony at data centers.

Optical computing itself isn’t new. But what’s different now is that many photonic processor startups with optical computing products are scrambling to exploit the fast-growing demand for energy-efficient AI computing.

Michael Förtsch, CEO of Q.ANT, a Stuttgart, Germany-based photonic computing startup, told the Ojo-Yoshida Report: “If you’re not changing the way we compute, the energy supply for data centers is going to be the bottleneck for progress on the next big step in AI.”

By Peter Clarke

What’s at stake:
An Israeli-Swiss startup called RAAAM has gone back to the drawing board to design a CMOS-compatible memory that sits between SRAM and DRAM and that could offer a cost advantage for numerous applications, including artificial intelligence. If RAAAM can persuade a few foundries and fabless companies to adopt its approach, it could become an industry-changer.

RAAAM Memory Technologies Ltd., headquartered in Pekah Tikva, Israel, and with an R&D base in Lausanne, Switzerland, is a 2021 startup that claims to have developed a better way to do embedded memory. The firm’s technology is called Gain-Cell RAM or GCRAM.

RAAAM claims it can achieve single-cycle SRAM read and write performance while providing up to a 50 percent area reduction and up to 10x power reduction compared with high-density SRAM. And GCRAM is fully compatible with standard CMOS fabrication flows, requiring no additional process steps.

By Peter Clarke

What’s at stake:
Right now, AI/ML is most powerful driver of technology but there are already signs that its runaway success is unsustainable on energy consumption grounds. Can a novel technology from a startup open up new frontiers in the artificial intelligence and machine learning (AI/ML) sector and ultimately impact the leaders of the semiconductor industry?

AI has a problem: the energy it consumes. A UK startup called Literal Labs may have a radical solution.

The company reckons a mathematical curiosity called the Tsetlin machine could provide an approach to many AI applications that is up to 1,000-times faster than GPU-based training and up to 10,000-times more energy efficient than today’s neural networks. If such efficiencies can be deployed while fitting into the established technology ecosystem, it could disrupt market leaders and enable AI at the edge, which has largely been stalled up until this point.

By Junko Yoshida

What’s at stake:
Before founding Oculi, Charbel Rizk was a designer of autonomous systems frustrated with computer vision systems available on the market. Traditional sensors, often built for human consumption, produce massive volumes of data, which result in systems needing more bandwidth and suffering from increased latency. Can Rizk convince other systems designers to embrace Oculi’s new vision architecture originally developed to fulfill Rizk’s own wish list?

Oculi, a Baltimore, Maryland startup, offshoot of a Johns Hopkins University research team, has developed a vision technology architecture in which sensing and processing both reside at the pixel level. The company calls it Sensing and Processing Unit (SPU).

Charbel Rizk, Oculi’s founder and CEO said in a recent interview with the Ojo-Yoshida Report, “My claim to the world is that we will always enable the lowest power, bandwidth, latency and ultimately cost computer vision solution with privacy.”

This is big talk among the many players in sensing and processing, all of them pursuing ultimate edge AI solutions in a broad range of embedded systems.

By Junko Yoshida

What’s at stake
Chiplets present a set of multi-layered, multi-faceted, multi-dimensional technology and business problems with no one-size-fits-all answer. Numerous startups are proposing various solutions to tackle the complexity of die-to-die interconnects.

For every player in the semiconductor supply chain – from chip designers and EDA tool vendors to semiconductor foundries, OSAT companies and the Babel of technology startups – the toughest challenge, arguably, boils down to how to connect chiplets.

Fortunately, startups such as Eliyan, Blue Cheetah and YorChip are poised to tackle the issues, each in its own way.

By Junko Yoshida

What’s at stake?
With chiplets poised to disaggregate SoCs into tiny dies, companies have begun to generate new ideas, tools and “chiplet platforms” designed to put back together these small dies (chiplets) – horizontally or vertically – in an advanced form of system-in-package. It is almost ironic that the chiplet originally conceived to disaggregate SoCs in Lego-like blocks appears to be getting back into an integration race again.

DreamBig Semiconductor, based in San Jose, Calif., is one of the startups at the gate. It came to CES last week and unveiled an “Open Chiplet Platform” called MARS.

By Junko Yoshida

What’s at stake: 
The chip industry’s trajectory changed forever when a wave of EDA companies sprang up in the 1980s and offered commercial tools to accelerate complex IC design. Traditionally reliant on internal expertise to improve process engineering at fabs, chipmakers might now be ready to embrace commercial tools developed for semiconductor production. 

Manufacturing is the new black in the chip industry.

By Junko Yoshida

What’s at stake:
The human brain is known for its efficiency, storing and processing information as sparse representations. At any moment, only a small fraction of neurons are active. Neuroscientists believe they can map brain-inspired logic into algorithms, data structures and architecture running AI models so that they offer a recipe to put “AI power hog” on a diet. Undetermined is whether data scientists are willing to alter their brute-force compute-based AI practice, which appears increasingly unsustainable.

It’s not often you meet a mostly self-funded startup, much less one that, for 18 years, has played a long game in neuroscience research, before suddenly unveiling a commercial software product.

More significantly, the startup’s new product poses a direct challenge to the red-hot Large Language Model (LLM) AI market.

By Junko Yoshida

What’s at stake:
Digital Twin is not a new concept. But by combining it with a new generation of AI models, MeetKai, a startup, is committing resources to creating effective interactive tools – to be used by factories, for example – for optimizing the production process and training engineers remotely and virtually.

Meet MeetKai, an AI-Metaverse tool company. Also, meet James Kaplan, MeetKai’s CEO, who dropped out of Harvey Mudd College in 2014 to run a private equity fund for Michael Milken, once a junk-bond villain.

Kaplan co-founded MeetKai in 2018 in L.A. with Weili Dai, Marvell Technology’s co-founder.