Coolidge is Kalray's third generation of massively parallel processor array (MPPA) architecture and is designed to support machine learning applications in addition to high-performance computing.
Kalray first mentioned Coolidge back in May 2017 (see Kalray's Coolidge processor adds deep learning acceleration) and subsequently conducted an IPO of shares in the company to help fund its development (see Loss-making Kalray raises funds via IPO).
Coolidge includes 80, 64bit VLIW processor cores plus 80 coprocessors designed to boost performance for artificial intelligence and compute intensive applications, to reach up to 4 Tera FLOPS and 25 TOPS – about 25x the performance of the previous generation MPPA. The total chip consumes less than 20W.
The chip is provided with high-speed interfaces such as 100Gbit Ethernet and PCIe Gen4, while specialized interfaces such as CAN assure seamless compatibility with targeted applications such as autonomous cars.
Kalray said Coolidge will address data center and automotive markets. Embedded applications include intelligent cars and autonomous vehicles, advanced driver assistance systems (ADAS), aerospace and defense, medical equipment robots, data center, networking, high performance computing, high performance storage, as well as, machine learning and artificial intelligence applications.
"Since the IPO of Kalray in June 2018, it has been one of the main focuses of the company to bring to life this new product," said Eric Baissus, CEO of Kalray, in a statement. "Numerous customers have identified the potential of our breakthrough solution and are eager to get the first samples of Coolidge."
Coolidge includes security and safety mechanisms defined in collaboration with Kalray's partners in the aerospace, defense and automotive fields.
Coolidge will be delivered to Kalray customers and partners with an upgraded version of Kalray Software Development Kit, AccessCore including Kalray Artificial Intelligence code generation tool (KaNN) as well as reference boards.
Related links and articles: