Flex Logix Technologies Inc., a semiconductor startup that counts Harvard University and DARPA among its customers, today debuted a new artificial intelligence chip for edge devices.

The Infer X1 is built to power industrial robots, network gateways and other systems that run outside an organization’s data centers. Flex Logix claims that the chip can perform no less than 8.5 trillion computing operations per second when running at peak speed. According to the startup, this makes the Infer X1 up to 10 times faster than competing products already on the market.

The key to this performance is Flex Logix’s custom interconnect architecture. That’s the part of a chip responsible for linking the different circuits with one another so they can exchange data and synchronize processing.

Flex Logix’s interconnect regularly reconfigures itself while an application is running, making optimizations on the fly to speed up the software. The Infer X1 harnesses this technology to optimize itself for each of the different types of computing operations that an AI model carries out as it ingests data.

Neural networks consist of layers, discrete groups of virtual neurons, that often process information in different ways. One layer might ingest data in the form of eight-bit integers, a common data format, and then pass it on to the next layer in the form of 16-bit integers. In this situation, the Infer X1 can readjust itself for each individual stage of processing to give the AI a performance boost.

Another notable feature is that the chip relies mostly on its internal cache to store the data being processed. As a result, the Infer X1 reduces the amount of external DRAM memory that hardware makers must include in their devices to support AI workloads. Lowering the number of components in a system simplifies product design, cuts production costs and means the device can make do with less power.

Flex Logix plans to start sampling Infer X1 to customers in the third quarter. It will be available as a coprocessor, a module designed to be incorporated into bigger, more sophisticated chips, and as a standalone PCIe card that companies can plug into existing systems.

Flex Logix is backed by $12.4 million in venture funding.

Photo: Marco Verch/Flickr