Quadric Merges AI, Application Code

Quadric’s Chimera DLA IP executes both neural-network code and application code. Ranging from 1 to 16 TOPS, it blends the functions of a CPU, a DSP, and a DLA.

Bryon Moyer

Like an organist who plays more than one set of keyboards with a single pair of hands, Quadric’s Chimera technology executes both neural networks and application code in a single binary file. The company calls it a general-purpose neural processing unit (GPNPU) by analogy to GPGPUs. The processor can act as a CPU, DSP, and deep-learning accelerator (DLA), switching between the three without implementing modes of operation.

Complete AI systems require more than just neural-network processing; inference is but part of a complete application. For example, some computer-vision applications implement pre- and postprocessing in addition to AI inference routines. Typical SoCs handle this task in standard CPUs for the application code and in a DLA for inference; control code ensures each component receives the right task. Quadric’s approach, introduced at the recent Linley Fall Processor Conference, provides a single processor for all application code along with software-development tools that blend the code.

Quadric began in 2017 as a DLA-chip company. After prototyping its first chip on a multiproject wafer, it decided to license intellectual property (IP) instead of selling chips, retiring the original Vortex brand and changing its part-numbering scheme. Chips based on that first design now appear in development boards. Quadric closed a $21 million Series B round in January, bringing its total funding to $38 million.

The company plans to offer Chimera versions that deliver 1, 4, or 16 trillion operations per second (TOPS), initially focusing on computer vision. The only component that changes between models is the size of the vector/matrix-processing array. Implemented in 7nm, the 4-TOPS version is expected to occupy 3.5mm2 and draw 925mW (typical). Early software access is planned for late November, with general RTL availability scheduled for 1Q23.