The graphics processing unit (GPU) is the chip on which most AI models run and is energy-intensive. Goldman Sachs predicts that AI will increase electricity demand by 160% by 2030 as a result of accelerated adoption of GPUs in data centers.
This trend is not sustainable, argues analog and memory circuit designer Vishal Sarin. After working in the chip industry for more than a decade, Sarin launched Sagence AI (previously named Analog Inference) to design energy-efficient alternatives to GPUs.
“Applications that have the potential to truly popularize practical AI computing are limited because the devices and systems processing the data cannot achieve the required performance,” Sarin said. “Our mission is to push the boundaries of performance and economy in an environmentally responsible way.”
Sagence develops chips and systems to run AI models and the software to program these chips. There's no shortage of companies creating custom AI hardware, but Sagence is somewhat unique in that its chips are analog rather than digital.
Most chips, including GPUs, store information digitally as binary strings of ones and zeros. In contrast, analog chips can represent data using a variety of values.
Analog chips are not a new concept. They had a heyday from about 1935 to 1980, when, among other engineering feats, they helped model the North American power grid. However, the drawbacks of digital chips have made analog attractive again.
First, digital chips require hundreds of components to perform a particular calculation, whereas analog chips can accomplish this with just a few modules. Digital chips also typically have to shuttle data between memory and processor, creating bottlenecks.
“All of the major legacy suppliers of AI silicon use this outdated architectural approach, which is hindering progress in AI adoption,” Sarin said.
Analog chips like Sagence, which are “in-memory” chips, don't transfer data from memory to the processor, so they can potentially complete tasks faster. Analog chips can also have higher data density than digital chips because they can store data using different values.
Vishal Sarin, Co-founder and CEO, Sagence Image credit: Sagence
However, analog technology also has drawbacks. For example, analog chips require more precise manufacturing and can be difficult to achieve high accuracy. They also tend to be difficult to program.
But Sarin sees Sagence's chips complementing, rather than replacing, digital chips, for example to speed up specialized applications in servers and mobile devices.
“Sagence products are designed to provide high performance for AI applications while solving the power, cost, and latency issues inherent in GPU hardware,” he said.
Sagence, which plans to bring its chips to market in 2025, aims to compete with other AI analog chip ventures such as EnCharge and Mythic and has signed up “multiple” customers, Sarin said. “We are currently in the process of packaging our core technology into system-level products to ensure they 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 total of $58 million in its first six years.
Now, the startup plans to raise funding again to expand its team of 75 people.
“Our cost structure is advantageous because we're not chasing performance goals by moving to the latest thing.” [manufacturing processes] For our tip,” Sarin said. “That’s a big factor for us.”
Timing may work in Sijens' favor. Funding for semiconductor startups appears to be recovering after a lackluster 2023, according to Crunchbase. From January to July, venture capital-backed semiconductor startups raised nearly $5.3 billion. This number is significantly higher than last year, when these companies raised less than $8.8 billion. total.
Chip manufacturing is thus an expensive project, made even more difficult by international sanctions and tariffs promised by the incoming Trump administration. Getting customers “locked” into an ecosystem like Nvidia's requires an extra climb. Last year, AI chip maker Graphcore raised nearly $700 million and was once valued at nearly $3 billion, but filed for bankruptcy after struggling to gain a firm foothold in the market.
To have a chance of success, Sagens must prove that its chips actually consume dramatically less power and be more efficient than alternatives, and raise enough venture funding to manufacture them at scale. need to.