Juniper Networks Inc. on Tuesday disclosed that it has invested in Recogni Inc., a startup with an artificial intelligence chip touted as significantly more efficient than the competition.

The companies didn’t divulge the size of the investment. Juniper provided the capital as part of a Series C round that Recogni originally announced in February. At the time, the chipmaker said that it had raised $102 million through the investment from Mayfield, BMW Group’s venture capital arm and other institutional backers.

About six months after the Series C round, Recogni debuted its flagship AI chip. Pareto, as the processor is called, promises to run neural networks using less power than some competing graphics cards. It’s also smaller, which means more chips can be installed in each server to boost processing speeds.

Pareto’s efficiency has to do with matrix multiplications, the calculations that AI models use to process data. A matrix is a collection of numbers organized into rows and columns like a spreadsheet. A matrix multiplication, in turn, is a mathematical operation that multiplies the numbers inside one matrix by numbers from another.

AI chips typically include a large number of circuits optimized to perform matrix multiplications. According to Recogni, such circuits are difficult to scale. The company says increasing the number of matrix multiplications a processor can perform requires significantly boosting both its size and power consumption, which raises costs.

Recogni claims to have found a workaround. The company’s Pareto chip performs AI inference not with matrix multiplications but rather using additions, which are much simpler to carry out. The result is an increase in hardware efficiency.

To make its computing approach practical for large-scale AI workloads, Recogni has equipped Pareto with several specialized optimizations. 

First, the company removed the need to use lookup tables, data structures that enable applications to retrieve information faster than would otherwise be possible. Such structures are usually necessary to carry out the kind of additions that Recogni’s chip performs. Phasing out lookup tables enabled the company to avoid the associated computing overhead. 

According to EE Times, Recogni also removed the need for customers to use quantization-aware training. This is an AI training approach that shrinks a neural network’s parameters, the configuration settings which determine how it processes data. Applying this method can be complicated and time-consuming. By sparing customers the need to implement quantization-aware training, Recogni made its chip simpler to adopt. 

When the company debuted Pareto in August, it detailed plans to sell the chip as part of custom rack-scale data center system. A rack is a large closet that can hold up to dozens of servers. At the time, Recogni also revealed that it was gearing up to announce a “technology partnership that will make the power of Pareto more widely available” within months.

That partnership is with Juniper, the companies disclosed today. According to Reuters, the network equipment maker will support Recogni’s efforts to build an AI inference system that can be installed in server racks. AI processors are often deployed in multichip clusters, which means they must be linked together using the kind of network gear that Juniper sells. 

Recogni envisions its hardware finding use among hyperscalers, cloud providers and enterprises. In parallel with its efforts to develop a rack-based AI inference system, the company is working on a new version of Pareto chip. The next-generation processor is expected to enter production in 2026. 

Image: Recogni