On November 3, 2022, Beijing-Moore Thread 2022 Autumn Conference was successfully held today in Zhongguancun National Independent Innovation Demonstration Zone, Beijing. At the press conference, Moore Thread launched the new multifunctional GPU chip "Chunxiao", the industry’s first national tide graphics card MTT S80 based on MUSA architecture, MTT S3000 for server applications, and the meta-computing integrated machine MCCX. This is another leap in the iterative innovation of Moore Thread’s multifunctional GPU products after a lapse of 7 months.
Not only hardware, Moore Thread also released a series of GPU software stacks and application tools around MUSA, including MUSA Developer Suite, cloud native sGPU technology and Metauniverse platform MTVERSE, etc., aiming at building an overall solution from the bottom chip to the upper development and application, and realizing the comprehensive upgrade of Moore Thread’s innovative mode of multi-functional GPU software and hardware integration.
At the conference site, Moore Thread also demonstrated more than 40 rich applications based on its multi-functional GPU, covering PC games, AI, digital people, digital twins, physical simulation, 8K multimedia display, cloud games, cloud desktop, digital office and other scenes. Partners such as AOC, Dell, Philips, Gujia, Tsinghua Tongfang, Inspur, Great Wall and Haier provided equipment support for application display, fully demonstrating Moore Thread’s powerful product applications.
GPU is a systematic project, involving hardware architecture, driver development, software ecology, sales and application, etc., with high R&D barriers and long industrial chain. The current GPU ecosystem, after decades of changes, has become huge and complex. If a GPU is to complete the commercial application from R&D to market, it can’t do without the continuous investment in hardware and software, and it can’t do without the strong support of ecology.
In March this year, Moore Thread officially released the first multifunctional GPU chip "Su Causeway", which has been recognized by many markets and ecology. The number of partners of Thread PES Perfect Experience System Alliance is increasing, covering CPU, operating system, OEM, software service provider, cloud service provider and system software developer. Based on the "Su Causeway" chip, Moore Thread and OEM partners have successfully launched a number of personal computers, workstations and data center server products, which are applied in business scenarios such as daily office, digital twinning, artificial intelligence training and reasoning; At the same time, we have joined hands with cloud service vendors to provide GPU cloud computing capabilities for users in different industries, paving the way for the application of Moore Thread GPU in many industries. China Mobile Cloud Capability Center has signed a strategic cooperation memorandum with Moore Thread to carry out extensive cooperation in the fields of cloud computing, cloud applications and accelerated computing; China Telecom Research Institute and Moore Thread will jointly explore the new infrastructure and local ecology of Metauniverse around intelligent computing, scientific research and development, application scenarios, etc., and promote the application of Metauniverse technology in key business scenarios.
Zhang Jianzhong, founder and CEO of Moore, expressed his gratitude to all partners and users on the spot, and further said: "GPU entrepreneurship is a long-term undertaking, full of challenges, and we know the importance of ecology. Based on the advanced MUSA architecture, Moore Multi-functional GPU continuously builds a complete software stack and application ecosystem, aiming at creating friendly support and experience for the open ecosystem. Only when we get together with eco-partners and industry users can we really exert the computing power of Moore Thread and provide the core power for the meta-universe and digital economy. "
Keeping the pace of the speed of light, Moore Thread officially released the second multifunctional GPU chip "Chunxiao", which integrates 22 billion transistors, has built-in MUSA architecture general-purpose computing core and tensor computing core, and can support FP32, FP16 and INT8 computing precision. Compared with the previously released "Su Causeway" chip, the four built-in computing engines of "Chunxiao" have been fully upgraded, which has brought about significant performance improvement: the graphics rendering capability has increased by 3 times on average; The coding ability is increased by 4 times, and the decoding ability is increased by 2 times; ; On average, the acceleration of AI is increased by 4 times, and the performance of physical simulation is increased by 2.5 times. At the same time, new technology is introduced to support narrow-band HD, saving bandwidth by more than 30%.
The newly released Moore thread MTT S80 is based on "Chunxiao" GPU chip, and it is also the first national tide graphics card for gamers. With 4096 programmable MUSA cores, it can provide 14.4TFLOPS single-precision floating-point computing power at 1.8GHz. At the same time, MTT S80 is also the first graphics card product equipped with PCIe Gen5 interface in the industry. With 16GB GDDR6 large-capacity high-speed video memory, supplemented by 8K ultra-high definition and 1080P 360Hz high refresh rate display output capability, it can bring the ultimate game vision and operation experience to gamers.
The successful launch of MTT S80 makes Moore Thread the first GPU company in China to support Windows environment and DirectX graphic interface. Its powerful 3D graphics rendering capability will be able to bring users a smooth operation experience in 4K resolution in Windows DirectX games. At present, the Windows driver of MTT S80 has built-in MUSA DirectX Driver module, and has been adapted to dozens of mainstream games such as diablo 3, League of Legends and Crossing the Firewire.
At the press conference, Moore Thread demonstrated the smooth game running effect brought by MTT S80. At present, Moore Thread is cooperating with game engine developers such as Unreal and Unity, as well as top domestic game developers such as tencent games, Netease Games, Xishanju, Perfect World, and 360 Games (in no particular order), so as to provide better and faster support for game engines and game products, and enable players to get a continuously updated 3A-level game experience. In the future, Moore Thread will continue to update Windows drivers and MUSA DirectX versions to achieve more game compatibility and performance optimization.
Moore Thread announced on the spot that MTT S80 graphics card has completed the first batch of production and stocking, and will start limited sales in JD.COM e-commerce platform on November 11th, 2022.
The newly released MTT S3000 is based on Moore Thread MUSA architecture, and it is also the first multifunctional server GPU product based on Chunxiao. The diverse computing power of Moore MTT S3000 can provide universal intelligent computing power support for AI reasoning and training, cloud games, cloud rendering, video cloud, digital twinning, digital content creation and other scenes with the help of the complete MUSA software stack covering graphics rendering, video processing and deep learning. It aims to build a solid computing power foundation for the construction of data centers, intelligent computing centers and meta-computing centers, and help the innovation and landing of diverse applications in the meta-universe.
Mts3000 contains 4096 MUSA stream processing cores and 128 dedicated tensor computing cores, with a transistor scale of 22 billion, an operating frequency of 1.9GHz and a memory bit width of 256bit；. With 32GB GDDR6 video memory, the bandwidth is 448gb/s; It supports FP32, FP16, INT8 and other kinds of calculation accuracy, among which the calculation power of FP32 can reach 15.2TFLOPS
Thread MTT S3000 and its supporting software and hardware products can cover the whole process from algorithm model to application deployment, and can provide friendly and rich package solutions for AI users. With the blessing of MUSA computing platform, MTT S3000 has multi-dimensional advantages such as ease of use, scalability and compatibility in deep learning training. In the aspect of deep learning reasoning, MTT S3000 supports mainstream AI models in many fields, such as vision, speech, natural speech comprehension and multimodal. At the same time, with the help of CUDA ON MUSA compatibility scheme developed by Moore Thread, users can seamlessly migrate the code developed on CUDA to MTT S3000. Moore also deeply optimized the performance of MUSA software stack and launched TensorX, a self-developed AI reasoning engine.
Ecology is crucial to the promotion of AI applications. At present, MTT S3000 is compatible with many mainstream deep learning frameworks such as PyTorch, TensorFlow, PaddlePaddle, Jittor, etc., and has accelerated dozens of AI models such as Transformer, CNN and RNN. Moore Thread also announced strategic cooperation with Baidu PaddlePaddle, Luchen Technology, Jittor, OpenMMLab and Zhiyuan Research Institute (in no particular order) to jointly prosper the AI ecosystem.
Software ecology is the key to promote the popularization of GPU computing. With MUSA architecture as the core, Moore Thread released a complete MUSA software stack to serve the vast number of developers and end users. MUSA software stack includes functional modules such as graphics rendering, multimedia, artificial intelligence, physical simulation, general computing, etc. It will cover from the bottom driver to GPU acceleration library, and then to application development kits customized for different industries. It is dedicated to meeting the application development needs in various industry scenarios.
MUSA development kit:Including MUSA general computing driver, MUSA compiler, AI operator library, general computing library, performance analysis tools, etc. Through the MUSA development kit, the majority of MUSA developers can develop, optimize and deploy various applications on desktop PCs, workstations, enterprise data centers, cloud platforms and distributed computer clusters based on Moore Thread multifunctional GPU.
CUDA ON MUSA：Moore Thread has developed a set of CUDA ON MUSA compatibility scheme for users who use CUDA language. Through two steps of compilation and running, CUDA source code can be run on Moore Thread MUSA architecture GPU.
MUSA user software:For the vast number of users, it provides multi-level series of software, including: drivers of different system platforms, PES and other platform software series; MUSA-based adaptive and deeply optimized AI framework, AI operator library, general computing library, optimization tools, containerized deployment runtime library, flexible split scheduling, etc. And cloud computing and application software series such as Digital Man, MT OCR and MT Smart Stream.
In order to make it easier for developers to obtain Moore Thread series software stacks, application solutions and technical support, Moore Thread Developer Website (https://developer.mthreads.com/) is officially launched. This website will be the MUSA software product release and download site, and also the MUSA technology exchange community. Moore Thread will build an open application and developer ecosystem around MUSA platform, and promote the prosperity and development of GPU ecology.
GPU is the key computing power needed by the current data center, and the application of cloud native technology in cloud computing, cloud desktop, cloud games and other fields is becoming more and more common. Therefore, Moore Thread released a series of GPU cloud native solutions based on Moore Thread’s innovative MT Mesh 2.0.
MT Mesh 2.0 can automatically allocate GPU computing and memory resources according to the cloud center application load, and realize the flexible expansion of GPU computing power. It can not only divide a GPU into multiple containers or virtual machines at will, but also support one container or virtual machine to schedule multiple GPUs.
Flexible containerized GPU (sGPU): Based on Kubernetes ecology, MT Mesh 2.0 is used to flexibly and effectively allocate containerized GPU resources.
Virtual GPU(vGPU): Using MT Mesh 2.0, it pioneered the resource elastic segmentation technology in the industry, which can dynamically allocate and modify GPU virtualization resources without restarting, and realize the on-demand call, dynamic expansion and release of computing power. The new "time-space split" feature is introduced, which supports hardware virtualization (SR-IOV), secure physical partitioning, up to 32-way virtualization, Windows cloud desktop GPU virtualization, and GPU pass-through of Unified Communications and Kirin operating systems.
Android container cloud acceleration technology ACX: Using GPU Android container transparent transmission technology and rendering coding integration technology can accelerate Android cloud mobile phone solution, reduce application delay and increase the number of concurrent channels. By natively supporting OpenGL ES, OpenGL rendering framework, and hardware acceleration of Android game material compression algorithms such as ETC/ETC2, better rendering effect and compatibility of Android cloud games can be provided.
Cloud and cloud games are important application scenarios of GPU in the cloud. Moore Thread has cooperated with many cloud desktop eco-partners to provide cost-effective cloud desktop solutions, serving the vast number of users in education, office, finance, geographic information and other industries. In the field of cloud games, Moore Thread and Tencent Pioneer have jointly created a leading cloud game solution for Android, and jointly created an excellent cloud game solution for PC with C-TREK Times, and are committed to bringing high-quality and smooth cloud game experience to the majority of game lovers.
At this conference, Moore Thread also released the MTVERSE Metauniverse platform and many software and hardware products specially built for Metauniverse applications, including the MCCX Metacomputing All-in-One Machine based on MTT S3000.
Based on Moor Thread MUSA GPU cluster, MTVERSE provides users with computing infrastructure services, including big data, AI training and reasoning, graphics rendering and physical simulation. It provides full-stack solutions from hardware cluster, software infrastructure to SDK tool chain, covering many core elements such as people, scenes and content in meta-universe. On the other hand, the massive SDK tools on the upper layer can help developers and applications conveniently call these capabilities, and realize a series of functions such as digital people, document understanding, speech recognition, visual recognition, natural semantic understanding, dialogue interaction, physical simulation, AIGC content generation, etc., further simplifying the development cycle and difficulty of applications and solutions.
Upgrade of AlphaCore physical simulation engine:Two new physical simulation products are added, including Flood Dynamics, a liquid interactive simulation tool, and Storm System；, a meteorological and cloud dynamics simulation tool; The effect and efficiency of gas and fluid simulation tool CatalystFX and cloth hair making tool VeraFiber have been comprehensively improved. Moore announced the deep integration with RaysEngine, the cloud native rendering engine, to realize digital twin real-time 3D rendering and physical simulation, and to make more realistic digital scenes.
DIGITALME digital person solution upgrade:Build a complete production line of digital people, including Nuwa’s digital life maker, Painted Skin’s expression driving engine, Suiying’s digital people’s action driving engine, and on-demand digital people’s dialogue system.
AIGC content generator ma Liang:It supports bilingual graphic generation and editing in Chinese and English, provides users with a zero threshold AIGC creation platform, and helps the integration of science and technology and humanities. Moore also cooperated with Yunnan Arts University to provide end-to-end "computing power and algorithm" solutions for artists, jointly set up "AI Art Creation Laboratory of Design Institute", and set up Maliang course to train design students and help artistic creation.
MCCX all-in-one computing machine:The purpose is to provide powerful and easy-to-use meta-computing power. For the meta-computing power application scenarios such as AI, rendering, coding and decoding, MCCX has designed a reasonable system architecture and resource ratio to provide users with cost-effective meta-computing power hardware solutions; By means of integrated delivery of software and hardware, a complete basic software stack, development environment, AI and rendering framework are preset, which supports one-click application deployment and upgrade, and realizes out-of-the-box meta-computing computing power.