Overview
Founded in 2023/4, the company develops compact supercomputing AI servers for AIGC and metaverse inference, emphasizing low power, value and scalability. Its supercomputing architecture and integrated computing/networking form parallel clusters over lossless Ethernet for large-model inference. Claims seventy percent lower construction costs and eighty percent lower operating power than traditional AI servers. Has completed principal validation, prototypes and chip selection, with scaled delivery possible in six months. Technology and architecture, claimed to lead by a year: 1. Servers contain only SOC computing units and a lossless-Ethernet switching backplane. Instead of PCIE, units communicate through high-speed lossless Ethernet using the COCE protocol. 2. Frontier technologies: 1) Parallel computing maps partitioned model logic to SOC networking and hardware, coupling collective computation and communication. 2) Embedded AI uses quantization and memory optimization to run partitioned models efficiently on SOCs. 3) Integrated computing/networking reduces expensive-unit waiting time, with claimed utilization above 75%. 4) Lossless Ethernet enables packet-loss-free, low-latency, nonblocking communication among model-computing nodes. Market advantages: 1. At equal computing output, claims power use of 12% and price of 30% of mainstream AI servers. A fully configured 2U server consumes 300W. 2. Scalable: fewer than sixteen chips in a server, costing roughly a hundred thousand yuan or more, can serve model inference. Plug-in expansion within servers and high-speed data-center switches extend capacity across servers. 3. With 16 computing units, a server can run inference for a 170b model; full configuration delivers 2.5P INT8 inference. A switched cluster exceeding thirty thousand servers could run brain-scale models, claimed to exceed the world’s strongest supercomputer. 4. SOC GPU, CPU and NPU units can be scheduled flexibly for diverse workloads, business expansion and algorithm compatibility. Supply chain: chips are described as resistant to supply restrictions and replaceable with domestic alternatives for local technology ecosystems. Team: former senior Huawei technology, R&D, management and marketing experts with over 20 years of ICT and AI experience. Combines technology and market strengths, complementary leadership, growth management and contacts in AI and telecom. CTO and R&D VP have deep technical and management backgrounds; the CTO holds AI and distributed-computing patents and has published an AI book. Market: extensive customers and promotional partners support rapid scaled sales, with a pipeline of representative industry clients.
Connected profiles
Sources & provenance
English text is maintained separately from the original record. Consult the original source before citing a fact.
No source links are available for this profile.
Original source text
Names may use phonetic renderings when no English name is supplied. The original wording is retained below.
公司成立于2023年4月，公司致力于为AIGC，元宇宙等场景提供超低功耗 、高性价比、易扩展的推理计算产品和系统-小超算AI服务器。我们基于超算架构和算网一体的技术方案研发了超算服务器，实现了在无损以太网上构建并行计算集群服务于大模型的并行推理计算，相比传统AI服务器，超动力解决方案降低了百分之七十的建造成本，降低了百分之八十的运行功耗，能够为大模型的商业化落地运营提供有竞争力的计算解决方案和产品，为客户创造价值。公司已经完成主要的技术验证和产品原型设计和芯片选型，六个月即可进行规模交付 ，能够快速实现商业价值。 一、高的技术门槛和颠覆性的架构，领先行业一年： 1、颠覆性架构。在服务器内部只有SOC计算单元和无损以太网交换背板，计算单元之间不再采用PCIE进行通信，而是采用了基于COCE通信协议的算网一体的高速的无损以太网进行通信； 2、高门槛的前沿技术 1）并行计算。把大模型的拆分后的并行计算逻辑图与SOC组网和硬件特性进行映射，实现集合通信计算和通信耦合； 2）嵌入式人工智能。通过模型量化、内存优化等技术实现拆分后的小模型如何能够高效的运行在SOC中； 3） 算网融合。实现计算和通信的协同以提升计算系效率，降低昂贵的计算设备的等待时间，我们的计算利用率可以达到75%以上。 4）无损以太网。如何实现大模型并行计算各个节点之间数据通信的不丢包，超低延迟和无阻塞。 二、产品场竞争优势 1、产品具备很强的竞争优势，相比行业主流的AI服务器，在输出相等的计算能力下，我们的功耗只相当于行业主流AI服务器的12%，价格只相当于主流AI服务器的30%，2U服务器满配只有300W的功耗； 2、该产品具备很强的扩展性，一个服务器十六片不到（差不多十几万元）就可以实现大模型的推理服务；且产品采用插拔式，很容易在服务器内部扩展，服务器之间采用高速数据中心交换机就可以实现进一步扩展； 3、性能强大，单个服务配置16片计算单元既可以实现170b大模型的推理，满配即可以提供2.5P的INT8的推理计算能力，三万多个服务器通过交换机实现大规模集群就可以跑人脑规模的超级大模型（这个超级集群算力高于当今世界算力最高的超算）； 4、具备足够灵活性：因为soc本身具备包括GPU/CPU/NPU等多种算力单元，因此，针对多样性的算力需求场景，我们可以灵活调度不同的算力类型以满足多样化的算力需求，支持业务扩展的灵活性和不同算法的适配性； 三、供应链优势 该产品的芯片不会被卡脖子，也可以支持信创（把国外的芯片更换成国内的芯片）。 四、团队优势 公司是由前华为资深技术和研发和管理营销专家成立的公司。团队成员均有超过20年的ICT及AI经验，该团队善于通过技术和场的双轮驱动实现公司的商业成功；团队主创人员配合默契、能力互补。创始人善于团队和增长管理，在AI、运营商等领域具有丰富的人脉资源。CTO和研发VP具备非常深厚的专业背景和管理经验。CTO拥有AI、分布式计算等专利并出版了AI方面的专著。 五、场优势 我们有非常丰富的客户资源和场推广合伙伙伴 能够快速实现产品在场的规模销售。目前已经储备了丰富的典型行业客户。