Juniper funds a firm that develops AI chips; acknowledges and describes technical cooperation

Recogni Inc. is a startup that has an artificial intelligence chip that is said to be much more efficient than the competition. Juniper Networks Inc. announced on Tuesday that it had invested in the company.

The firms withheld the amount of the investment. The funding came from Juniper as part of a Series C investment that Recogni first revealed in February. With the investment from Mayfield, the venture capital arm of the BMW Group, and other institutional backers, the chipmaker claimed to have raised $102 million at the time.

Recogni’s main AI chip was introduced approximately six months following the Series C financing. The processor, known as Pareto, claims to be able to operate neural networks with less power than some rival graphics cards. Additionally, because it is smaller, each server can have additional processors installed to increase processing rates.

Matrix multiplications, the computations AI models do to handle data, are the source of Pareto’s efficiency. Similar to a spreadsheet, a matrix is a group of integers arranged in rows and columns. The mathematical process of multiplying the numbers in one matrix by the numbers from another is known as matrix multiplication.

Numerous circuits designed to do matrix multiplications are commonly found in AI chips. Recogni claims that scaling such circuits is challenging. According to the corporation, a processor’s size and power consumption must be greatly increased in order to conduct more matrix multiplications, which drives up expenses.

Acknowledge claims to have discovered a solution. The company’s Pareto chip uses additions, which are far easier to execute, rather than matrix multiplications to accomplish AI inference. As a consequence, hardware efficiency rises.

Recogni has given Pareto a number of unique enhancements to make its computing methodology feasible for massive AI workloads.

Initially, the business eliminated the requirement for lookup tables, which are data structures that let programs get information more quickly than they otherwise could. In order to do the types of additions that Recogni’s chip does, such structures are typically required. The business was able to eliminate the related computing expense by phasing out lookup tables.

Recogni also eliminated the requirement that clients employ quantization-aware training, according to EE Times. This method of AI training reduces the size of a neural network’s parameters, or the configuration settings that dictate how the network interprets data. This approach can be difficult and time-consuming to implement. Recogni simplified the adoption process for its chip by avoiding the requirement for customers to implement quantization-aware training.

The company described intentions to market Pareto as a component of a customized rack-scale data center system when it made its debut in August. Numerous servers can be stored on a rack, which is a big closet. Recogni also disclosed at the time that it was preparing to reveal in the next months a “technology partnership that will make the power of Pareto more widely available.”

According to the firms’ disclosure today, that agreement is with Juniper. The manufacturer of network equipment will assist Recogni in developing an AI inference system that can be mounted in server racks, according to Reuters. Since AI processors are frequently used in multichip clusters, Juniper’s network equipment is required to connect them.

Recogni hopes that cloud providers, corporations, and hyperscalers will employ its gear. A new Pareto chip is being developed by the business concurrently with its efforts to create a rack-based AI inference system. It is anticipated that the next-generation CPU will go into production in 2026.