Renesas Electronics Company (TSE: 6723), a premier provider of superior semiconductor options, right now introduced the event of embedded processor expertise that allows larger speeds and decrease energy consumption in microprocessor models (MPUs) that understand superior imaginative and prescient AI. The newly developed applied sciences are as follows: (1) A dynamically reconfigurable processor (DRP)-based AI accelerator that effectively processes light-weight AI fashions and (2) Heterogeneous structure expertise that allows real-time processing by cooperatively working processor IPs, such because the CPU. Renesas produced a prototype of an embedded AI-MPU with these applied sciences and confirmed its high-speed and low-power-consumption operation. It achieved as much as 16 instances quicker processing (130 TOPS) than earlier than the introduction of those new applied sciences, and world-class energy effectivity (as much as 23.9 TOPS/W at 0.8 V provide).

Amid the current unfold of robots into factories, logistics, medical companies, and shops, there’s a rising want for programs that may autonomously run in real-time by detecting environment utilizing superior imaginative and prescient AI. Since there are extreme restrictions on warmth era, significantly for embedded units, larger efficiency and decrease energy consumption are required in AI chips. Renesas developed new applied sciences to fulfill these necessities and offered these achievements on February 21, on the Worldwide Strong-State Circuits Convention 2024 (ISSCC 2024), held between February 18 and 22, 2024 in San Francisco.

The applied sciences developed by Renesas are as follows:

(1)  An AI accelerator that effectively processes light-weight AI fashions

As a typical expertise for enhancing AI processing effectivity, pruning is out there to omit calculations that don’t considerably have an effect on recognition accuracy. Nevertheless, it’s common that calculations that don’t have an effect on recognition accuracy randomly exist in AI fashions. This causes a distinction between the parallelism of {hardware} processing and the randomness of pruning, which makes processing inefficient.

To unravel this situation, Renesas optimized its distinctive DRP-based AI accelerator (DRP-AI) for pruning. By analyzing how pruning sample traits and a pruning methodology are associated to recognition accuracy in typical picture recognition AI fashions (CNN fashions), we recognized the {hardware} construction of an AI accelerator that may obtain each excessive recognition accuracy and an environment friendly pruning fee and utilized it to the DRP-AI design. As well as, software program was developed to cut back the burden of AI fashions optimized for this DRP-AI. This software program converts the random pruning mannequin configuration into extremely environment friendly parallel computing, leading to higher-speed AI processing. Specifically, Renesas’ extremely versatile pruning assist expertise (versatile N: M pruning expertise), which might dynamically change the variety of cycles in response to adjustments within the native pruning fee in AI fashions, permits for effective management of the pruning fee in line with the ability consumption, working velocity, and recognition accuracy required by customers.

This expertise reduces the variety of AI mannequin processing cycles to as little as one-sixteenth of pruning incompatible fashions and consumes lower than one-eighth of the ability.

(2)  Heterogeneous structure expertise that allows real-time processing for robotic management

Robotic functions require superior imaginative and prescient AI processing for recognition of the encompassing atmosphere. In the meantime, robotic movement judgment and management require detailed situation programming in response to adjustments within the surrounding atmosphere, so CPU-based software program processing is extra appropriate than AI-based processing. The problem has been that CPUs with present embedded processors aren’t absolutely able to controlling robots in real-time. That’s the reason Renesas launched a dynamically reconfigurable processor (DRP), which handles complicated processing, along with the CPU and AI accelerator (DRP-AI). This led to the event of heterogeneous structure expertise that allows larger speeds and decrease energy consumption in AI-MPUs by distributing and parallelizing processes appropriately.

A DRP runs an software whereas dynamically altering the circuit connection configuration between the arithmetic models contained in the chip for every operation clock in line with the processing particulars. Since solely the required arithmetic circuits function even for complicated processing, decrease energy consumption and better speeds are attainable. For instance, SLAM (Concurrently Localization and Mapping), one of many typical robotic functions, is a fancy configuration that requires a number of programming processes for robotic place recognition in parallel with atmosphere recognition by imaginative and prescient AI processing. Renesas demonstrated working this SLAM by means of instantaneous program switching with the DRP and parallel operation of the AI accelerator and CPU. This resulted in about 17 instances quicker operation speeds and about 12 instances larger working energy effectivity than the embedded CPU alone.

Operation Verification

Renesas created a prototype of a take a look at chip with these applied sciences and confirmed that it achieved the world-class, highest energy effectivity of 23.9 TOPS per watt at a traditional energy voltage of 0.8 V for the AI accelerator and working energy effectivity of 10 TOPS per watt for main AI fashions. It additionally proved that AI processing is feasible and not using a fan or warmth sink.

Using these outcomes helps clear up warmth era resulting from elevated energy consumption, which has been one of many challenges related to the implementation of AI chips in a wide range of embedded units similar to service robots and automatic guided automobiles. Considerably decreasing warmth era will contribute to the unfold of automation into varied industries, such because the robotics and sensible expertise markets. These applied sciences shall be utilized to Renesas’ RZ/V collection—MPUs for imaginative and prescient AI functions.