SK 海力士预测 到 2030 年 AI 基础设施性能规模将增长约 20 倍SK Hynix predicts that the scale of AI infrastructure performance will increase by about 20 times by 2030


AI free token/AI免费积分
SK 海力士预测 到 2030 年 AI 基础设施性能规模将增长约 20 倍
感谢IT之家网友 华南吴彦祖、补药吖 的线索投递!

IT之家 10 月 14 日消息,据韩联社报道,SK 海力士开发总管安炫(Ahn Hyun)今日在首尔江南区 COEX 举行的“第 28 届韩国半导体大展”(SEDEX 2026) 上发表主旨演讲。他预测称,2030 年 AI 基础设施表现规模将增长约 20 倍。

安炫指出,当前 AI 推理市面正同时在三个维度快速扩张:任务复杂度的提高、单次处理信息量(上下文长度)的激增,以及推理请求量的井喷。这些因素相互交织,推动了 AI 算力基础设施需要的迅猛增长。

随着 AI 推理需要的暴增,承担海量数据存储与迁移任务的存储器,其关键地位愈发凸显。特别是用于保存并复用前序处理信息的“键值缓存(KV Cache)”容量持续膨胀,迫使平台必须将数据分摊存储于高带宽内存 HBM、DRAM 以及 SSD 等不同介质中。分析表明,在此过程中产生的数据迁移,正演变为制约表现提高并推高平台能耗的全新瓶颈。

对此,安炫作出了这一预测,即人工智能基础设施的表现规模到 2030 年将增长约 20 倍。

安炫强调,行业不能仅停留在提高单一存储芯片的表现上,而亟需采取优化整个存储层次结构的全局思路。

为此,SK 海力士在持续优化现有 HBM 表现的同时,正全力推进 3D 堆叠 DRAM 以及高带宽闪存 HBF 等下一代存储工艺的研发:

IT之家从报道获悉,SK 海力士的目标是将 HBF 的带宽做到同代 HBM 的 50% 左右,同时达成约 10 倍的容量扩容。

本文来源:https://www.ithome.com/1/012/980.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT House October 14 news, according to the Yonhap news, SK Hynix development chief Ahn Hyun (Ahn Hyun) today held in Seoul Gangnam-gu COEX "28th Korea semiconductor exhibition" (Sedex 2026) keynote speech. He predicts that the scale of AI infrastructure performance will increase by about 20 times in 2030.

An Xuan pointed out that the current AI inference market is expanding rapidly in three dimensions at the same time: the increase in task complexity, the surge in the amount of information (context length) processed once, and the blowout of inference requests. These factors are intertwined to drive the rapid growth in AI computing infrastructure needs.

With the explosion of AI inference needs, memory, which is responsible for massive data storage and migration tasks, has become increasingly prominent. In particular, the capacity of the "key-value cache" used to save and reuse preamble processing information continues to expand, forcing the platform to store data in different media such as high-bandwidth memory HBM, DRAM, and SSD. The analysis shows that the data migration generated in this process is evolving into a new bottleneck that restricts the performance improvement and pushes up the energy consumption of the platform.

In response, Anxuan made this prediction that the performance scale of artificial intelligence infrastructure will increase by about 20 times by 2030.

Anxuan stressed that the industry must not only improve the performance of a single memory chip, but also take a global approach to optimize the entire storage hierarchy.

SK Hynix predicts that the scale of AI infrastructure performance will increase by about 20 times by 2030

To this end, SK Hynix is continuing to optimize existing HBM performance while fully promoting the development of next-generation storage processes such as 3D stacked DRAM and high-bandwidth flash HBF:

IT Home learned from the report that SK Hynix's goal is to achieve about 50% of the bandwidth of the same generation HBM, while achieving about 10 times the capacity expansion.

This article is from: https://www.ithome.com/1/012/980.htm For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry