Highlights:

  • Recogni’s Pareto chip boosts hardware efficiency by performing AI inference using simpler additions instead of matrix multiplications.
  • AI chips usually feature numerous circuits optimized for matrix multiplications, but Recogni notes that these circuits are challenging to scale.

Recently, Juniper Networks Inc. partnered with Recogni Inc., a startup known for its artificial intelligence chip, which is claimed to be much more efficient than competing technologies.

The companies did not disclose the investment amount. Juniper contributed to Recogni’s Series C funding round, which was initially announced in February. At that time, the chipmaker revealed it had raised USD 102 million from investors including Mayfield, BMW Group’s venture capital arm, and other institutional supporters.

About six months after the Series C round, Recogni introduced its flagship AI chip, Pareto. This processor is designed to run neural networks with lower power consumption than some competing graphics cards. Additionally, its smaller size allows for more chips to be installed in each server, enhancing processing speeds.

Pareto’s efficiency stems from its ability to perform matrix multiplications, the core calculations AI models use to process data. A matrix is a set of numbers arranged in rows and columns, similar to a spreadsheet. Matrix multiplication is a mathematical operation that multiplies the numbers in one matrix with those in another.

AI chips usually contain numerous circuits designed to optimize matrix multiplications. Recogni explains that scaling these circuits is challenging. The company states that increasing the number of matrix multiplications a processor can handle requires a substantial increase in both size and power consumption, which in turn drives up costs.

Recogni says it has found a solution. Instead of relying on matrix multiplications, the company’s Pareto chip performs AI inference using additions, which are simpler to execute. This approach leads to improved hardware efficiency.

To ensure its computing approach is suitable for large-scale AI workloads, Recogni has incorporated several specialized optimizations into the Pareto chip.

First, the company eliminated the need for lookup tables, which are data structures that allow applications to retrieve information more quickly. These tables are typically required for the additions performed by Recogni’s chip. By removing them, the company was able to reduce the associated computing overhead.

According to EE Times, Recogni also eliminated the need for customers to use quantization-aware training, an AI training technique that reduces the size of a neural network’s parameters, which control how it processes data. This method can be complex and time-consuming. By removing this requirement, Recogni made its chip easier for customers to adopt.

When the company launched Pareto in August, it outlined plans to sell the chip as part of a custom rack-scale data center system. A rack is a large enclosure that can house dozens of servers. At that time, Recogni also shared that it was preparing to announce a “technology partnership” aimed at making Pareto’s capabilities more widely accessible within the coming months.

The partnership, as disclosed by the companies today, is with Juniper. According to Reuters, the network equipment provider will assist Recogni in developing an AI inference system designed for installation in server racks. Since AI processors are typically deployed in multichip clusters, they need to be connected using the type of network equipment that Juniper supplies.

Recogni anticipates its hardware will be used by hyperscalers, cloud providers, and enterprises. Alongside its work on a rack-based AI inference system, the company is developing a new version of the Pareto chip, with the next-generation processor expected to begin production in 2026.