BLAIZE, which provides semiconductors for autonomous edge AI, starts in Japan
Compared to NVIDIA's Xavier, the system efficiency is 15 times, and the power consumption is one -fifth.
Blaizeが提供するGSP、低遅延でメモリアクセスを減らすことで高い電力効率を実現する(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)BLAIZE, a US AI solution, announced on November 5 to hold a press conference online and launched business activities in Japan as a sales agency.
BLAIZE is a venture company that offers semiconductors and software for AI founded by engineers involved in the development of semiconductors in Intel and NVIDIA, etc., and the company is dedicated to edge AI that the company calls GSP (Graph streaming processor).It offers semiconductors and software that use them to automakers and Tiwan parts manufacturers.
BLAIZE CEO Dinaka Munagara says, "GPU cannot satisfy the performance and cost per power required by customers," and use the GPU provided by NVIDIA to infer the edge AI.In comparison, he appealed that the same performance could be achieved with low cost and low power consumption.
Blaize CEO ディナカー・ムナガラ氏BLAIZE CEO Dinaka Munagara, who was on stage at the beginning of the press conference, introduced the company's outline and characteristics.According to Munagara, the company was founded by engineers who worked as semiconductor manufacturers such as Intel and NVIDIA.Looking at the executive's background on the company's website, Mr. Munagara himself is the same, but it is an architect or programmer with built -in GPU built into the Intel processor that is commonly used in PCs.It will be a company created by those who are familiar with.In this presentation, it was revealed that NSSMTEXTEXS will be operating in Japan as a distributor, and Hideki Sugimoto, the highest technical manager (CTO), is a video.We received a welcome comment on the partnership.
株式会社エヌエスアイテクス 最高技術責任者(CTO)杉本英樹氏(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)According to Munagara, "I am currently investing in keeping layers in the industry, such as Denso, Daimler, Magna, and Samsung. It is planned to start mass production in the fourth quarter of this year.It is approaching the start of semiconductor production with customers (Proof of Concept, conceptual demonstration, semiconductor using semiconductors to verify that products are actually realized).He said that he was actually starting a prototype of the product with the customer.
Blaizeの概要(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)Mr. Munagara said, "Now, the use of AI is shifting from the data center to the edge. In such a situation, we thought that we needed a product that meets the needs specializing in Edge AI, which cannot be realized with the GPU. CPU.And GPUs are too powerful and cost to obtain the necessary performance. We believe that it is necessary to have a product that is suitable for realizing edge AI in the future, and customers need it.I would like to provide solutions, "he said, and for AI, which will be required on the edge side like an autonomous car in the future, it is not a general -purpose CPU or GPU, but in terms of power consumption.However, he emphasized that the company's solutions that specialized in edge AI would be needed.
エッジAIには新しいアプローチが必要(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)Blaize ストラテジック・ビジネス・デベロップメント担当副社長 リチャード・テリル氏Vice President Richard Terril, Vice President of BLAIZE Stratagic Business Development, explained the solution for edge AI provided by the company.According to Teril, the company provides a semiconductor product specialized in AI's inference called GSP (graph streaming processor) and software products to use it.
GSPの仕組み(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)According to the company, GSP is a processor specialized in the de -plastic infusion processing on the edge side.If the CPU or GPU processes multiple images, etc., it will be processed sequentially, so access to the memory to place the data will occur each time.However, in GSP, the processing of the process can be reduced as much as possible by processing the process in parallel and exchanging data in the processor.As a result, it is possible to process it with higher performance and low power consumption compared to when using GPUs.
GSPは1チップ(16コア)で16TOPSを実現し、消費電力は7W(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)According to Terril, "GSP can achieve 16 TOPSs with only 7W power consumption in GSP. This is 60 times more efficient than other solutions such as GPUs."He emphasized that high efficiency is a feature of GSP.
As a solution that uses such GSPs, BLAIZE will provide three products: Pathfinder P1600, which is a board (SOM) product, and XPLORER accelerator (X1600E EDSFF and X1600P PCIE), which adds an accelerator as an accelerator.
Pathfinder P1600(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)The Pathfinder P1600 is a product with a GSP, memory, and various I/O (Ethernet, USB, etc.) on one module for built -in boards, which are provided in a smaller size than a credit card.It has 16 cores with GSP, achieving 16TOPS performance.The use of network equipment and edge devices is considered.
Xplorerアクセラレータ(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)Xplorer X1600E EDSFF is PCI Express 3.A product that has a 0 x4 interface and is conscious of incorporating it into a server.After all, it has 16 cores of GSP, achieving 16TOPS performance.
Xplorer X1600p PCIE has a PCI Express expansion card, and can be used by incorporating it into a desktop PC or server.The GSP cores can be used up to 16 to 64 core, and the performance is 16 to 64TOPS.
ソフトウェア開発環境(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)"Software development tools are important in the development of AI. We have provided AI STUDIO and Picasso SDK by working to enhance such software development kits, and have been evaluated for two years."He emphasized that the software development kit, which is important in AI semiconductors, is also focusing.
According to Terill, Picasso SDK is a conventional software development kit for programmers who are used to software development in such development kits.The unique is AI Studio, and this is a tool that can easily create AI programs using visual tools, making it easier and easier to build AI programs.
Blaize フィールド・アプリケーション・エンジニア 小林達也氏BLAIZE Field Application Engineer Tatsuya Kobayashi explained such a Blaize solution.Mr. Kobayashi said, "Automobile is an important application for BLAIZE's GSP. It has been calling for the recruitment of OEM manufacturers and Tiwan parts manufacturers, making it possible to cover multiple functions with one chip, making it a severe demand for cars.It can also be dealt with, "he said, especially in terms of electricity efficiency, there is an advantage for semiconductor products currently adopted for automatic driving cars, such as GPU and FPGA.
自動車に最適(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)Xavierとの比較(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)Helm.aiの事例(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)As a specific example, we introduce the comparison between Pathfinder P1600 and NVIDIA's Xavier, and if the same performance is the same performance, the system efficiency is 15 times, power consumption is one -fifth, higher power efficiency and low cost.He said that it will be realized.And Helm.Introducing products that provide level 2 + and level 3 autonomous driving provided by AI, and realizing a system that processes images from six cameras at a low delay of 50 ms or less, solutions using GPUs.He explained that it could reduce cost reduction by 50 % or more and 80 % or more.
各種事例(出典:性能・効率を追求したAIエッジ・コンピューティング、Blaize)In addition, in addition, objects detection of objects with multiple cameras, cases of detecting human and robot movements, cases of detecting people and vehicles in smart city, security cases in retail, sensor fusion examples, and cases of home delivery.Several cases were introduced.