Venture-backed chip startup Blaize Inc. has introduced three new hardware modules for running artificial intelligence models at so-called edge locations such as factories.

The modules, announced on Thursday, are based on the Graph Streaming Processor that the startup first previewed last November. That’s the month Blaize exited stealth mode with $87 million in funding from investors that included Samsung Electronics Co. Ltd.’s Catalyst Fund and Daimler AG.

Blaize’s Graph Streaming Processor or GSP is a chip designed specifically to run artificial intelligence models. It performs this task up to 60 times more efficiently than traditional central processing units and graphics cards, according to the startup. The chip can manage as many as 16 trillion operations per second while consuming about seven watts of power, which is on par with the electricity consumption of a smart light bulb.

The GSP’s efficiency is the result of a highly specialized chip architecture. At the software level, neural networks can be implemented in the form of a data structure known as a graph, and Blaize has equipped the GSP with optimizations specifically designed to run such graphs. One of the main optimizations is a feature that saves electricity by lowering the number of times the chip has to move data to and from memory while processing it, an operation traditional processors have to perform often.

Blaize’s new modules make it easier for enterprises to deploy the GSP. The first module, the Pathfinder P1600, can be integrated with connected devices such as industrial sensors to process the data they generate. A manufacturer, for example, could link the P1600 to a set of cameras above its factory floor and use computer vision models to spot defective components on the production line.

The two other modules Blaize debuted are the Xplorer X1600E and X1600P. They’re designed to add more AI processing power to desktop computers and servers, respectively, at edge locations such as retail stores. The X1600P, the chip for servers, provides up to four GSP chips for maximum performance equivalent to 64 trillion operations per second.

“Today’s edge solutions are either too small to compute the load or too costly and too hard to productize,” said Blaize Chief Executive Officer Dinakar Munagala. “Blaize AI edge computing products overcome these limitations of power, complexity and cost.”

Blaize is currently sampling the modules to early customers. The startup plans to start full-scale production next quarter. It says early adopters are harnessing GSP chips for multiple real-world applications including monitoring industrial robots in factories and analyzing sensory data from vehicles.

Photo: Blaize