根据业内最新消息，长鑫存储计划在未来两个月内启动HBM3E量产。今年8月，公司已经开始HBM3E风险试产，并向部分客户提供样品。如今，相关产品正逐步向商业化阶段推进。
HBM是当前人工智能加速器、数据中心GPU以及高性能计算系统最关键的存储技术之一。由于能够提供远高于传统DRAM的带宽，HBM已成为训练和运行大型人工智能模型不可或缺的核心组件。目前全球HBM市场主要由SK海力士、三星电子和美光掌控。
按照JEDEC标准，HBM3E采用1024位总线设计，单引脚传输速度可达到9.2Gbps至12.4Gbps，主流配置通常为9.6Gbps。其总带宽约为1.2TB/s，可提供24GB至36GB容量配置，其中包括8层堆叠和12层堆叠两种主流方案。
不过，量产对于长鑫存储而言并非易事。业内消息显示，当前HBM3E仍面临良率挑战。由于HBM采用多层DRAM晶粒通过TSV硅通孔技术进行垂直堆叠，对制造和封装工艺要求极高，因此良率成为制约产能扩张的重要因素。
消息人士透露，目前长鑫存储HBM3E整体良率约为20%至30%，明显低于国际领先厂商水平。其中，前段晶圆制造阶段良率约30%，进入后段封装后可进一步获得约70%的合格率提升，但最终整体良率仍有较大改进空间。
业内分析认为，长鑫存储当前最大的技术挑战来自TSV技术成熟度。与国际厂商相比，其堆叠工艺仍有一定差距。例如，三星在HBM2时代单层使用约5000个TSV连接，而SK海力士在HBM3产品上已超过8000个。长鑫存储则尝试利用约3000个TSV实现HBM3E设计，这对制造精度和可靠性提出了更高要求。
受限于相关技术因素，长鑫存储短期内预计将专注于8层堆叠HBM3E产品，暂时不会向容量更大的12层堆叠产品快速推进。
与此同时，多家消息机构披露，长鑫存储已经向国内AI芯片企业提供了两种版本的HBM3E样品，其中包括标准版本以及针对良率进行优化的新版本。相关样品目前正在客户侧进行测试和验证。
性能层面，初步测试显示，长鑫存储目前HBM3E产品整体表现更接近SK海力士上一代HBM3，而与国际先进HBM3E产品相比仍存在差距。不过业内人士指出，对于刚刚进入HBM赛道的厂商而言，实现量产本身已是重大突破。
分析人士认为，即使当前良率不高，长鑫存储仍有动力继续推进量产。随着中国AI芯片厂商对高带宽内存需求快速增长，本土HBM供应来源的重要性正在持续提升。面对先进AI存储供应紧张以及进口限制带来的压力，国内市场对于HBM产品存在明显需求。
近期市场已经出现AI芯片价格上涨现象，高带宽内存供应被普遍视为重要原因之一。包括阿里巴巴平头哥、寒武纪等芯片企业此前均被曝测试国产HBM解决方案，希望降低对海外供应链的依赖。
业内预计，如果客户验证顺利，采用长鑫存储HBM3E的国产AI加速器产品最快有望于2027年进入市场。尽管与全球领先厂商相比仍存在一至两代技术差距，但长鑫存储进入HBM领域意味着中国本土AI硬件生态正在补齐关键拼图，并开始向高端人工智能存储市场发起挑战。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to the latest information available in the industry, the long-term storage plan is to start HBM3E production within the next two months. In August this year, the company began HBM3E risk trials and provided samples to some clients. Today, the products are moving towards commercialization.
HBM is one of the most critical storage technologies for the current artificial intelligence accelerator, data centre GPU and high performance computing systems. HBM has become an essential core component for the training and operation of large artificial intelligence models because of its ability to provide bandwidth well above traditional DRAM. The global HBM market is currently dominated by SK Hercules, Tristar Electronics and American Light.
According to the JEDEC standard, HBM3E is designed with a 1024-bit bus, with a single-legged transmission rate of 9.2 Gbps to 12.4 Gbps, with the usual mainstream configuration being 9.6 Gbps. Its total bandwidth is approximately 1.2 TB/s, which provides a 24- to 36-GB capacity configuration, including 8-storey and 12-storey mainstream programmes.
However, volume production is not easy for long-term storage. According to industry sources, HBM3E is still facing a positive rate challenge. Since HBM uses multi-layer DRAM crystals for vertical stacking through TSV silicon perforation technology, which is extremely demanding for manufacturing and sealing processes, a good rate is an important constraint on capacity expansion.
According to sources, the overall positive rate of long-term storage of HBM3E is currently about 20 to 30 per cent, significantly below the level of international leading manufacturers. Of these, the positive rate for the previous crystal circle manufacturing phase was about 30 per cent, and further improvements of about 70 per cent could be achieved by entering the back-sealing, but there was still room for improvement in the overall end.
In-industry analysis suggests that the greatest current technological challenge for long-term storage comes from the technological maturity of TSV. There is still a gap in the stacking process compared to international manufacturers. For example, Samsung uses about 5,000 TSV connections in the single layer of the HBM2 era, while SK Hercules has over 8,000 on HBM3 products. Long condensed storage attempts to achieve HBM3E design using about 3,000 TSVs, which places higher demands on manufacturing precision and reliability.
Subject to relevant technical factors, long-duration storage is expected to focus on eight layers of HBM3E products in the short term, with no rapid advance towards a larger 12 layers of stacking.
At the same time, several news agencies disclosed that the long-stamp storage had provided two versions of HBM3E samples to domestic AI chip enterprises, including a standard version and a new version that optimized the good rates. The samples are currently being tested and validated on the side of the client.
At the performance level, preliminary tests have shown that the overall performance of the current HBM3E product in long-term storage is closer to that of the previous generation of the SK Hercules HBM3 and that there is still a gap compared to the international advanced HBM3E product. However, industry has pointed out that the realization of the production itself has been a major breakthrough for the producers who have just entered the HBM track.
Analysts believe that, even if the current positive rate is not high, long-term storage has the incentive to continue to advance production. The importance of home-grown HBM supply sources is increasing with the rapid growth of demand for high-bandwidth memory by Chinese AI chip producers. Domestic market demand for HBM products is evident in the face of pressure from advanced AI storage supply constraints and import restrictions.
The recent increase in the price of AI chips in the market has been widely recognized as one of the important reasons. The chip companies, including Brother Alibaba, Chiu Wu, were previously exposed to testing the home-grown HBM solutions in the hope of reducing dependence on overseas supply chains.
It is expected within industry that if customer certification is successful, the domestically produced AI Accelerator products that are stored in HBM3E using long-stamps are most likely to enter the market by 2027. Although a one-to-two generation technological gap remains compared to global leaders, long-storey storage into the HBM domain means that China’s indigenous AI hardware ecology is completing key puzzles and is beginning to challenge the high-end artificial intelligence storage market.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.