News from MarketWatch.com
《GyrfalconTechnology Enables High-Performance AI at the Edge with Lowest EnergyConsumption》
MILPITAS,Calif., Sept. 18, 2018/PRNewswire/
Patentedprocessing-in-memory solutions rely on standard technologies forcost-effective,scalable AI compute in a small footprint.
GyrfalconTechnology Inc. (GTI) todayemerged from relative stealth mode to become thefirst company withcommercially available hardware that delivershigh-performance, low-power AIprocessing at the edge, shipping to customersincluding LG, Fujitsu andSamsung. GTI's generally available patented MatrixProcessing Engine (MPE) isdesigned to enable high-performance AI processing forIoT and other edgeequipment within the smallest energy envelope at anindustry-leading ratio.
Stealthmode：Inbusiness, stealth mode is a company's temporary state of secretiveness, usuallyundertaken to avoid alerting competitors to a pending product launch or otherbusiness initiative.
UnlikeAI-enabling technologies offered bysome of the industry's cloud players, GTIhas taken an "edge-first"approach to its IoT technology. By enablingcore AI processing in the edgeequipment – vs. relaying data to be processed bydata center AI and back to theedge via the cloud – GTI's solution avoids thelatency in edge equipmentreceiving actionable insight and reduces the costinvolved in transmittingdata. In addition, the solutions increase privacy as inmany cases data neverleaves the edge equipment.
GTIprovides unparalleled edge AIcapabilities thanks to its patented MatrixProcessing Engine (MPE™), which runsAI solutions on an ASIC for domain-specificapplications. Coupled with atwo-dimensional matrix with AI processing in Memory(APiM™), MPE enables theindustry's best performance-to-power usage meaningequipment can operate athigh speeds to handle AI quickly, with low energyconsumption.
Thecompany's first-generation product, theLightspeeur® 2801S, is a 7mm x 7mm 28nmASIC that uses just 300mW of power todeliver a superior 9.3 TOPS/W that canprocess audio and video input. With itssmall size and the ability to combinebetween two and 32 chips on one board forheavy compute loads or separate taskhandling, the 2801S enables the broadestrange of use cases and equipmentdesigns, at a price point 10X lower thancompetitive hardware. Therefore, notonly does the high performancedifferentiate AI products, but companies can alsoachieve significantlyimproved margins. This is a breakthrough for companieswith an interest inmanufacturing mass-market products at scale, where thetechnology needs to bemade small and inexpensive without compromisingperformance, such as in a smarthome lock or camera.
"Balancingthe cost-performance-energyequation has been a challenge for developers lookingto bring AI-enabledequipment to market at scale," said Dr. Lin Yang, chiefscientist, GTI."The GTI founding team has been watching the industrystruggle with thischallenge for decades, and believe that our AI Processing inMemory and MatrixProcessing Engine provide an elegant solution to avoid havingto maketrade-offs. By deploying APiM and MPE on a standard, commoditized ASIC,GTI isenabling our customers to bring innovative, AI-enabled devices tothemasses."
"Weare paving the way for the nextwave of AI products to make it to market,"said Kimble Dong, CEO of GTI."We recognized that device makers werecompromising on essential designvariables in AI-enabled equipment and havesought to solve this over the pastfew decades. Our offering marries our"edge-first" approach withultra-fast AI data processing technology,low power consumption and a smallchip design to enable the best AI experienceand performance at a low cost,within any AI use case, physical fit and deployment."
GTIis addressing the fast-growing IoTcategory, which is expected to grow onaverage 12 percent in the number ofdevices annually according to IHS Market.The company's enablement of thesmallest and most cost-effective IoT deviceswith edge AI is enhanced byself-learning capabilities necessary to maximizeproductivity. Data models withchip-adaptive training device setup is as simpleas plugging in, powering upand getting devices started with a voiceprint,selfie or gesture. After initialsetup, smart devices can take their cues fromtheir surroundings, for example,a lock learning the residents of a home througharrival and departure patterns.
Lightspeeur®2801S is shipping today, andis already in development with more than 10customers, including LG, Fujitsuand Samsung, in countries where GTI has officesin Asia and North America. Alsoshipping to customers developing edge AIsolutions is GTI's USB3.0 dongle withan embedded Lightspeeur® 2801S thataccelerates customer development processeswhen creating next generation AIenabled equipment and devices.
AboutGyrfalcon Technology Inc. GyrfalconTechnology Inc. (GTI) is the world's leadingdeveloper of low-cost, low-power,high-performance Artificial Intelligence (AI)processors. Founded by veteranSilicon Valley entrepreneurs and ArtificialIntelligence scientists, GTI aimsto expand the power of cloud ArtificialIntelligence to local devices withgreater performance and efficiency, and tomake Artificial IntelligenceProductization possible. For more information onGTI, visit www.gyrfalcontech.com.
Gyrfalcon Technology Lightspeeur®2801s AI Accelerator features high
performance and superior energy efficiency (9.3 Tops /Watt) ,enabling ultra-fast AI edge computing and data center machine learning
News from siliconAngle.com
Specializedcomputing chips that can be customized for modern workloads such as artificialintelligence processing are all the rage. Now, a startup called GyrfalconTechnology, Inc. is joining in with a new application-specific integratedcircuit for “internet of things” devices that enables high-performance andlow-power AI at the network edge.
GTIsaid today it’s taking what it calls an “edge-first” approach to AI-enabled IoTdeployments by shifting the processing work from the data center to the devicesthemselves. The idea is that doing so eliminates the latency that results fromrelaying data across cloud networks. Its ASICs can therefore generate insightsfor users faster and also cut down on the costs of transferring data to andfrom the cloud.
ASICs areintegrated circuits that can be customized for a particular use, as opposed toregular central processing units that are intended for general-purpose use. Forexample, an ASIC chip can be designed specifically for powering AI workloadssuch as natural language processing. They’re also commonly used mining bitcoinand other cryptocurrencies, among other workloads.
GTI’sfirst ASIC, the Lightspeeur 2801S, is powered by the company’s MatrixProcessing Engine. It enables “AI processing in memory” for IoT devices it’sdeployed in. The Lightspeeur 2801S is also extremely energy-efficient, enablesa wide range of use cases and sells at a price point that’s 10 times lower thansimilar competitive hardware, the company claimed.
Inaddition, the hardware is compatible with many of the most popular deeplearning frameworks used to create AI software, including TensorFlow and Caffe.
“Popularapplications include image and video classification and identification, facialrecognition, character identification, image enhancement for viewing,manufacturing defect detection, use with audio interfaces, segmentation orimages and video, captioning and other implementations where edge AI can drivenew experiences, convenience and productivity for individuals andorganizations,” GTI co-founder and President Frank Lin told SiliconANGLE.
GTIreckons that its ASICS are a “breakthrough” for IoT device makers requiringsmall and inexpensive processing power that doesn’t compromise on performance.Examples of such devices include smart home locks and cameras.
“Onebig difference between our offering and what is commonly provided is the high performancecombined with low power use,” Lin said. “By having a very compelling size,price, energy usage profile and high performance, there are an extremely broadrange of devices we can help our customers improve with our AI Accelerator.”
TheLightspeeur 2801S ASIC is available now. GTI says it’s already working withmore than 10 customers to get its chips in their products, including the SouthKorean electronics giants LG Electronics Inc. and Samsung Electronics Co. Ltd.and Japanese computing firm Fujitsu Ltd.