Startup Gyrfalcon is rapidly emerging with a corollary chip for deep neural networks (DNNs), but it is also facing increasing crowding in the Artificial Intelligence (AI) chip market after the company first won One year after funding, the AI Altiplates chip, which has been available recently, is demonstrated, claiming multiple design orders in the areas of smartphones, security cameras and industrial automation equipment.
Data centers typically train in deep neural networks and use a range of servers to perform inferences about them, and accelerators are being used by more and more clients and embedded systems today, from cars to mobile phones.
Companies such as Apple, Google and Huawei have introduced in-built SoC smartphones with inference modules, and Google and Microsoft have also built inferences accelerators for their data centers.
Of the many vendors, Movidius, now part of Intel Corp., has already begun shipping inference chips, while Cadence, Ceva and Synopsys supply IP blocks, In addition, at least a dozen AI startups are still in stealth mode, including Groq, a co-founded by a group of former Google chip developers, that is working to build inference chips.
Linley Gwennap, head of market watcher The Linley Group, said: "Much of the movement about AI inference chips comes primarily from IP providers; for the most part it will be an integrated chip market."
Gyrfalcon has introduced its Lightspeeur SPR2801 processor-available chip for Convolutional Neural Networks (CNN) since autumn last Fall.There are TSMC's 28nm process and 7x7-mm package, including tens of thousands of dedicated cores and embedded static Random Access Memory (SRAM) delivers 9.3 TOPS / W efficiency without the need for an external dynamic random access memory (DRAM) These are mainly arithmetic logic units (ALUs) with associated memory and logic for controlling data flow element.
The first generation of the Gyrfalcon Lightspeer 2801, with APiM technology, provides clouds and edges with parallelism of several orders of magnitude more power than the rest of the market for deep training and reasoning tasks. The accelerator card works with M.2 and PCIe Interface, support TensorFlow, Caffe and other open source environment.
The new chip concept came from the idea put forward by its co-founder and chief scientist Lin Yang in his 1988 doctoral dissertation when he conducted a CNN study at UC Berkeley in California The chip is designed to accelerate CNN, which is one of the most widely used neural network applications.
Frank Lin, co-founder and president of Gyrfalcon, said the top five smartphone makers will import the chip in a future flagship handset, a major insurer in China that will be designed into its surveillance cameras and a Shanghai-based Steelmakers will also use the chip in their inspection system.
Yang Lin said: 'We have achieved rapid growth after securing the first tranche in February last year. We verified the FPGA in April and completed the chip manufacturing at TSMC in mid-September.'
Developed in partnership with Socionext and QuickLogic
Gyrfalcon announced in December last year that its chips will be used in a video analytics platform from Socionext in conjunction with its parallel processors, a product developed in collaboration with Sociétéxt and QuickLogic to be launched in March and Turkey Security Monitor System manufacturer Bilkon Ltd. will be its first customer.
QuickLogic is evaluating Gyrfalcon chips with its next-generation sensor hub for the possibility of detecting wake-up words in a variety of 'hearable' products. "Timothy Saxe, QuickLogic's chief technology officer, said: 'At least a dozen inference chips Of the companies evaluated by QuickLogic, they are closer to our market than other companies. '
Saxe said AIAs such as Amazon's 'Alexa will make people want a broader vocabulary for more natural language processing, so we must speed things up through alliances.'
Gyrfalcon plans to launch the second product offering higher performance in June of this year and release a new series of chips by the end of this year and will also explore the possibility for a large number of customers to create customized chips as low as 50mW.
The startup, which designed a PCI Express card for server-on-chip integration, will follow Movidius to launch a USB dongle with the new chip in April to attract broader CNN software development Business
'We do not want this task to be done by only a Ph.D. - this USB dongle supports several pre-loaded neural network models and is very easy to use.' "Yang also assisted China Mobile before joining Gyrfalcon, Promote its international business.
However, he did not reveal how much money Gyrfalcon has received so far from the two rounds of investments made by China, South Korea, Japan and the United States. He said only: 'By 2018, we have raised enough funds to support the overall operation, Design orders and potential orders, expected to be profitable this year, and even hope to breakeven.
So far, the company has 41 employees, most of whom are in Silicon Valley headquarters.
Compile: Susan Hong