"跑一个70B大模型，先得准备800万元买显卡？"——这不是段子，是多数企业AI本地化立项书的 第一行数字。 人工智能正以前所未有的深度与广度重塑千行百业，然而当企业投身于AI本地化部署时，两大根本 性瓶颈横亘眼前： 一是 数据"供不上、存不下"的困局；二是在动辄数百GB的AI模型面前， GPU"寸土寸金"的显存墙。 当"显存墙"越砌越高，产业界的梯子却分出两条截然不同的搭法。一条是"算力派"的直梯——把 高端GPU像积木一样继续往上码，用更多的H100、A800去换"寸土寸金"的显存空间；另一条 是"系统派"的折叠梯——在存算之间插入一层"弹性夹层"，把冷数据、温数据、热数据按温度梯 度逐层腾挪，用软件定义和异构硬件去挤出额外的容量与带宽。 国外已有微软DeepSpeed、AMD Infinity Cache探路，国产阵营里，铨兴科技把这套"夹层"思路 拆成两半：先用122 TB QLC eSSD等产品搭起"高性能eSSD矩阵"，把PB级模型权重稳稳落在闪 存里；再用"添翼本地AI超显存融合方案"把FPGA控制器引入PCIe通道，在微秒级完成数据搬运， 仅把当前最活跃的激活张量留在GPU显存——单机即可释放 ...

Core Insights - The introduction of AI integrated machines has significantly reduced the cost for small and medium enterprises (SMEs) to deploy AI and large models from millions to around 200,000 yuan, with expectations for further reductions to the tens of thousands range in the coming months [1] Group 1: Product Development - Lenovo showcased its latest achievement in edge computing at the Tech World conference, introducing the "Lenovo Inference Acceleration Engine," a platform designed for efficient AI PC inference [1] - This engine, developed in collaboration with Tsinghua University and Wuneng Chip, allows a standard PC's local inference capabilities to rival OpenAI's o1-mini cloud model released last year [1] Group 2: Market Impact - Lenovo's CEO, Yang Yuanqing, stated that the demand for AI is experiencing explosive growth, predicting a tripling of edge AI capabilities within the next 12 months [2] - The showcased desktop product, equipped with the acceleration engine, is priced around 40,000 yuan and is expected to facilitate local training of 32B large models, significantly lowering deployment costs for SMEs in sectors like finance, education, and law [2] Group 3: Cost Efficiency - Traditional training solutions for 32B large language models can cost around 2 million yuan, requiring at least eight NVIDIA graphics cards, making it prohibitively expensive for SMEs [2] - The Lenovo solution, utilizing the "AI Studio" software platform, can reduce local training costs by 98%, making it accessible for SMEs to train models with their own databases and personalized information [2]