Graphics processing units (GPUs), the chips that run most AI models, are power-hungry beasts. As a results of the increasing integration of GPUs into data centers, AI will result in a 160% increase in electricity demand by 2030, in accordance with Goldman Sachs Estimates.
The trend is unsustainable, argues Vishal Sarin, a designer of analog and memory circuits. After working within the chip industry for over a decade, Sarin founded Sagence AI (previously called it). Analogous conclusion) to develop energy-efficient alternatives to GPUs.
“The applications that would make practical AI computing truly ubiquitous are limited since the devices and systems that process the info cannot deliver the required performance,” Sarin said. “Our mission is to beat performance and economic limitations in an environmentally responsible manner.”
Sagence develops chips and systems to operate AI models in addition to the software to program these chips. While there is no such thing as a shortage of corporations developing custom AI hardware, Sagence is exclusive in that its chips are analog, not digital.
Most chips, including GPUs, store information digitally as binary strings of 1s and 0s. In contrast, analog chips can represent data using a spread of various values.
Analog chips are usually not a brand new concept. They experienced their heyday from around 1935 to 1980, when, amongst other things, they helped model the North American power grid. But the disadvantages of digital chips make analog chips attractive again.
Firstly, digital chips require Hundreds of components to perform specific calculations that analog chips can do with just just a few modules. Digital chips also typically need to move data forwards and backwards from memory to the processor, which creates bottlenecks.
“All of the leading legacy AI silicon vendors are using this old architectural approach, and that is obstructing progress in AI adoption,” Sarin said.
Analog chips like Sagence's, that are “in-memory” chips, don’t transfer data from memory to processors, potentially allowing them to finish tasks faster. And due to their ability to make use of a spread of values ​​to store data, analog chips can have higher data density than their digital counterparts.
However, analog technology also has its downsides. For example, it might be tougher to realize high precision with analog chips because they require more precise manufacturing. They also are inclined to be tougher to program.
But Sarin sees Sagence's chips as a complement – slightly than a substitute – to digital chips, for instance to hurry up specialized applications in servers and mobile devices.
“Sagence products are designed to eliminate the performance, cost and latency issues related to GPU hardware while delivering high performance for AI applications,” he said.
Sagence, which plans to launch its chips in 2025, is working with “multiple” customers to compete with other AI analog chip corporations corresponding to EnCharge and Mythic, Sarin said. “We are currently packaging our core technology into system-level products and ensuring we fit into existing infrastructure and deployment scenarios,” he added.
Sagence has secured investments from backers including Vinod Khosla, TDK Ventures, Cambium Capital, Blue Ivy Ventures, Aramco Ventures and New Science Ventures, raising a complete of $58 million within the six years since its inception.
Now the startup plans to lift capital again to expand its 75-person team.
“Our cost structure is favorable because we don’t pursue performance goals by migrating to the most recent (manufacturing processes) for our chips,” Sarin said. “That’s an enormous factor for us.”
The timing could work in Sagence's favor. Per CrunchbaseAfter a lackluster 2023, funding for semiconductor startups appears to be rebounding. From January to July, VC-backed chip startups raised nearly $5.3 billion – a number well above last yr, when these corporations raised lower than $8.8 billion overall.
Against this backdrop, chip manufacturing is an expensive undertaking – made even tougher by the international sanctions and tariffs promised by the brand new Trump administration. Acquiring customers “locked” into ecosystems like Nvidia’s is one other steep climb. Last yr, AI chipmaker Graphcore, which earned nearly $700 million and was once valued at nearly $3 billion, filed for bankruptcy after struggling to realize a foothold available in the market.
To have a likelihood of success, Sagence might want to prove that its chips actually use significantly less power and offer greater efficiency than alternatives – and lift enough enterprise capital to supply them at scale.
