[新聞] 三星突破記憶體瓶頸 公布HBM-PIM成果

[kyle5241 (Kyle Korver)]

https://tinyurl.com/ywfs2xpy

三星突破記憶體瓶頸　公布HBM-PIM、LPDDR-PIM研究成果

https://tinyurl.com/ymux929s

Samsung breaks through memory bottleneck; announces research results on
HBM-PIM and LPDDR-PIM

At the 2023 Hot Chips forum, in addition to Intel's announcement of its data
center chip product, the latest report from Korea's TheElec also pointed out
Samsung Electronics has announced its research results on high bandwidth
memory (HBM)-processing-in-memory (PIM) and low power DDR (LPDDR)-PIM as part
of its efforts to focus on the AI sector.

三星公布了高頻寬記憶體內運算(HBM-PIM)和低功耗動態記憶體內運算的成果

Previously, Samsung and AMD began a collaboration related to PIM technology.
Samsung equipped HBM-PIM memory onto AMD's commercial GPU accelerator card,
the MI-100. According to Samsung's research results, applying HBM-PIM to
generative AI will more than double the accelerator's performance and power
consumption efficiency compared to existing HBMs.

之前三星和AMD開始在記憶體內運算(PIM)合作，三星把HBM-PIM整合到AMD
GPU MI-100中。根據三星的研究，HBM-PIM在同樣的攻耗下，運算效能可以提升一倍

To solve the memory bottleneck that has begun to appear in the AI
semiconductor sector in recent years, next-gen memory technologies like
HBM-PIM have received significant attention. The HBM-PIM conducts computation
processing within the memory itself through PIM technology. This simplifies
the data movement steps, thereby enhancing performance and power efficiency.

為了解決AI運算中的記憶體瓶頸，記憶體內運算受到很大的注目，HBM-PIM是直接在
記憶體內執行運算，這樣簡化了數據傳輸的步驟，因此增加性能和功耗

Furthermore, to verify the Mixture of Experts (MoE) model, Samsung used 96
HBM-PIM-equipped MI-100 units to build a HBM-PIM cluster. In the MoE model,
the HBM-PIM accelerator doubled the performance and tripled the power
efficiency compared to HBM.

在MoE模型中，三星的HBM-PIM展現了2倍性能以及提升了三倍的功耗比

Industry sources explained that the speed of memory development has been
slower compared to the advancements in AI accelerator technology. To
alleviate this memory bottleneck, it's necessary to expand the application of
next-gen semiconductors like HBM-PIM. Additionally, in sectors like LLM, many
data sets are frequently reused. Therefore, utilizing HBM-PIM computation can
also reduce data movement.

記憶體速度的增速遠低於AI加速的進展，為了抒解記憶體瓶頸，HBM-PIM的使用是必要的
此外，對於AI模型LLM來說，很多的數據都會重複的使用，所以使用HBM-PIM可以大幅
的減少數據移動

On the other hand, Samsung also introduced the "LPDDR-PIM," which combines
mobile DRAM with PIM to enable direct processing and computing within edge
devices. Notably, because LPDDR-PIM is designed for edge devices, it offers
lower bandwidth (102.4GB/s) and saves 72% of power compared to DRAMs.

另外對於邊緣運算裝置，三星結合了手機記憶體在PIM中，創造了可以在邊緣運算
裝置中使用的LPDDR-PIM，雖然頻寬比較小，但能耗也降低了72%

Previously, Samsung revealed its AI memory plans during its 2Q23 earnings
call. It not only mentioned that the HBM3 supply was undergoing customer
verification but also stated that it's actively developing new edge AI memory
products and PIM technology. Looking ahead, both HBM-PIM and LPDDR-PIM are
still some time away from commercialization. Compared to existing HBMs, PIMs
are quite expensive.

之前，三星說它的HBM3在接受客戶的驗證，而且它也在開發新的邊緣運算的AI記憶體
產品和PIM技術。目前，HBM-PIM和LPDDR-PIM離商業化還有段距離，主要是PIM還是
非常的貴

The Hot Chips forum is a prominent academic event in the semiconductor
industry. It's typically held in late August. Apart from Samsung, other major
companies like SK Hynix, Intel, AMD, and Nvidia also participated in this
event.

心得：
如果PIM做出來，就完全不需要先進封裝了...

--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 111.249.86.213 (臺灣)
※ 文章網址: https://www.ptt.cc/bbs/Tech_Job/M.1693495145.A.DA9.html
→ tony890415  : 美光GG
→ payneblue   : 貴到用不起 不用想太多
→ samm3320    : 那個裡面要一大堆ADC，很難降吧
推 darkangel119: 三星王朝再起
噓 PowerGeMOS  : LK99也突破極限，但只有韓國做出來
推 shinyi444   : 掌握了超導體和記憶體 韓國好強（？
推 ru04hj4     : 記憶體 三星就真的強啊
推 havochuman  : 好了啦 邏輯晶片輸 講其他都多的
→ haoboo      : 話說LK99後來還有啥下文嗎
→ magicee01   : 0<+>0
→ kyle5241    : LK-99 理論預測如果有doping 可能會有一些有趣的性
→ kyle5241    : 質 畢竟銅高溫超導沒doping 也是Mott insulator
推 ubei        : 美光到底行不行？被韓國二大家追著打
→ transonic   : 美光最大的利多就是美國爸爸，不可能倒啦
推 ljsnonocat2 : 這樣電腦架構要修改嗎?
→ TaiwanUp    : LK99的機構只是一種類似俱樂部號稱的研究中心
→ TaiwanUp    : 三星的話                    好像有時候差不多
推 jkl852      : 大內宣都喜歡說技術領先韓國
噓 pf775       : 韓國新聞能信？
→ bbinbbin    : 超導記憶體嗎？