On September 3rd, as a high-end chip design company based in China and serving the world, Vastai Technologies participated in WAIC 2022 and held the "Empower everything across virtual and physical worlds with chips"—— General Purpose AI Chip's Enterprise Applications Forum. At the conference, Jun Qian, founder and CEO of Vastai Technologies, announced Vastai Unified Compute Architecture, new data center inference accelerator Zaitian VA10, edge AI inference accelerator Zaitian VE1, the extended version of Vastai's software platform-VastStream, and also previewed the domestic 7nm cloud GPU chip SG100.
Vastai Unified Compute Architecture, AI Chip and Software Platform
Powerful chip performance comes from the excellent computing architecture. At the conference, Vastai released its self-developed Unified Compute Architecture (Vastai Unified Compute Architecture). The architecture integrates a variety of high-performance computing engines, with efficient and unified storage management, consistent interfaces and low link latency, complete virtualization capabilities, unified underlying software design, and modular upper-level computing libraries and functional modules. The Vastai unified architecture maximizes hardware acceleration of performance for compute-intensive AI, video, and rendering tasks, while reducing latency end-to-end and integrating to power cloud-based real-time graphics rendering, AI enhancement processing, video codecs.
Vastai also announced the Data Center Inference Card Zaitian VA10, an accelerator with Vastai's self-developed ultra-low latency, ultra-high throughput AI engine, the biggest highlight of which is the Int8 peak computing power of 400 TOPS, once again setting a new industry record. With a power consumption of 150 watts, Zaitian VA10 delivers an overall maximum inference performance that is more than twice that of mainstream GPUs with the same power consumption and a latency as low as 6% of the latter. In low latency scenarios (YoloV3 detection algorithm scenarios below 4 ms), the inference performance of Zaitian VA10 is more than 3 times that of mainstream GPUs at the same power consumption. Zaitian VA10 is suitable for a variety of cloud-based AI application deployments that require high real-time performance, such as live video enhancement, intelligent traffic management, real-time semantic understanding, and other scenarios.
Edge commuting products are another innovation of this announcement, and Vastai released the first edge-oriented inference card, Zaitian VE1 series. It is worth noting that the Zaitian VE1 has both ultra-low latency and ultra-high throughput rate, with a throughput rate twice that of mainstream GPUs, but the latency is only 1/5 of that.
More importantly, Zaitian VE1 has low power consumption and high computing power, with INT8 peak arithmetic power of 100 TOPS under 40~65 watt power consumption. This accelerator card supports 60 channels 1080P video decoding in real time, supports models of mainstream AI frameworks, and provides VastPipe, a full process low-code development framework from model compilation to inference optimization, combining software and hardware to provide customers with end-to-end AI inference solutions.
Based on the advantages of ultra-high throughput rate and ultra-low latency, Zaitian VE1 can play an important role in many key scenarios, such as highway incident detection, "people, vehicles, non-motorized vehicles" video structuring, big data research and judgment, and other ultra-high data volume concurrent scenarios; and then, for example, road real-time dynamic information interaction, industrial defect detection and other low-latency requirements scenarios. Zaitian VE1 is very suitable for intelligent transportation, vehicle-road coordination, industrial quality inspection and other edge computing scenarios.
Hardware needs synchronized support of the software in order for better performance. Therefore, Vastai software platform-VastStream came into being. that accelerates the deployment of various AI applications, including computer vision, video processing, natural language processing, search and recommendation, and custom extension of operators.
The new software components of VastStream provide three major management tools, including system management, to facilitate customer deployment. At the same time, its basic software stack has more functions, including unified interface, flexible scheduling, general AI calculation, multi-channel efficient video transcoding + AI enhancement, Easy programming and fast migration, complete tool chain, etc.
So far, Vastai has formed a system solution covering chips, inference accelerator cards, all-in-one hardware product lines and software integrated system solutions.
Domestic 7nm cloud GPU: SG100
At the conference, Jun Qian also demonstrated the domestic 7nm cloud GPU chip SG100, which has industry-leading graphics rendering performance, world-leading ultra-high throughput, ultra-high quality, and low-latency coding capabilities, and integrates rendering, AI, and video functions providing deep optimization for cloud gaming, cloud phone, cloud desktop, cloud computing, and other Metaverse critical applications. At the same time, SG100 also provides world-class SR-IOV virtualization support, which improves end-users’ visual experience.
It is reported that Jun Qian was a former AMD executive. Before founding Vastai Technologies, he led the AMD China team to design and massively produce the industry's first 7nm graphics processor and the first 7nm GPGPU AI chip. At the same time, Vastai core R&D team has more than 18 years of experience in high-end chip design. The successful development of SG100 is based on the above-mentioned profound and unique technical background and heritage.
Vastai has officially entered the GPU market by previewing SG100. In the future, Vastai will bring more product portfolios based on GPU, AI and video core technologies to better meet the needs of the market and customers.
Join hands with partners to build application eco-system
In the new era of AI and Metaverse, AI chips will play a vital role as a new type of computing power infrastructure. While actively developing industry-leading AI chips, Vastai Technologies has been deploying industrial implementation to promote the development of digital economy.
Since 2022, Vatai has successively carried out in-depth cooperation with industry-leading companies such as Fujian Big Data Group, Guoning Ruineng, GuoNingRuiNeng, and Super Fusion，to provide domestic AI computing power solutions for enterprise intelligentization and digitization in multiple scenarios such as smart city, smart government, smart transportation, smart park, smart energy, etc.
Building eco-system is the top priority in Vastai’s strategic layout. In order to promote the development of technology ecology, Vastai has been establishing cooperation with top tech talents, institutions and universities around the world.
During the meeting, Vastai signed a contract with AI Research Institute of Shanghai Jiaotong University to carry out school-enterprise cooperation. Vastai hopes to establish an intelligent chip ecology by involving more outstanding talents and resources through the alliance of industry-university-research to promote technological progress, talent cultivation and transformation of scientific research achievements.
In the future, Vastai will continue to give full play to the advantages of its own products, such as super computing power and efficient adaptation, and join hands with partners to accelerate the pace of industrialization in various places and various scenarios, and empower more enterprises and government agencies to achieve intelligent upgrades, using vast computing power to boost the development of China's digital economy.