三星推出全球首款低功耗内存PIM解决方案
三星电子宣布推出全球首款将计算能力直接嵌入低功耗内存的AI推理技术，即PIM解决方案。该技术旨在加速向低功耗内存的转变，并有望提升AI推理效率。此产品涉及半导体存储与计算融合领域，对元器件供应链中低功耗内存及AI芯片市场具有潜在影响。
原文深度解读
本文报道三星电子将于2024年8月发布全球首款将PIM（存内计算）技术集成到LPDDR5X低功耗内存中的AI推理解决方案。该技术基于三星5月发表的学术论文，旨在解决CPU/GPU与内存间的数据瓶颈，提升AI推理速度与能效。文章涉及三星、SK海力士等公司，关键数字包括GEMV矩阵计算速度提升6.2倍。该产品对电子元器件供应链中低功耗内存、AI芯片及终端设备（如智能手机、笔记本电脑）市场具有潜在影响。
- 三星电子计划在8月发布集成PIM技术的LPDDR5X AI推理解决方案，该方案基于5月发表的学术论文。
- PIM技术将计算单元嵌入内存内部，减少数据搬运，解决CPU/GPU与内存间的数据瓶颈。
- 在LPDDR5X-PIM模拟中，GEMV矩阵计算速度比传统内存快6.2倍，意味着AI任务（如照片分析、文本处理）可提速6倍以上。
- 三星采用双轨策略：HBM面向服务器市场，LPDDR5X-PIM面向个人设备（如笔记本电脑、智能手机）。
- LPDDR5X-PIM与三星自研的4纳米AI加速器GAIA配合时，可进一步提升效率，GAIA负责主要AI计算，PIM提供内存级支持。
- 竞争对手SK海力士已开发自有PIM产品AiM（基于GDDR6-AiM）和AI加速器AiMX。
供应链影响
- 三星LPDDR5X-PIM的发布可能推动低功耗内存市场对PIM技术的需求增长，影响相关内存芯片的供应链布局。
- 该技术可能加速端侧AI设备（如智能手机、笔记本电脑）的更新换代，从而影响终端设备制造商对低功耗内存和AI芯片的采购决策。
- 三星与SK海力士在PIM领域的竞争可能影响HBM和LPDDR系列内存的市场份额分配，进而影响上游晶圆代工和封装测试产能分配。
- PIM技术的商业化可能取决于良率、成本及生态支持，可能影响供应链中相关设备和材料的供应需求。
- 该技术对AI推理效率的提升可能促进端侧AI应用普及，可能影响数据中心与终端设备之间的算力分配，进而影响服务器内存和端侧内存的供需平衡。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
LP5X-PIM模拟器的框图。〈来源：“LP5X-PIM Sim: A High-Fidelity HW/SW Integrated Simulator for LPDDR5X-PIM”〉
三星电子已确保据信是全球首项将计算能力直接嵌入低功耗内存的AI推理技术，此举预计将加速向低功耗、高效率AI新时代的转变。
据行业消息人士7月22日透露，三星电子计划在8月推出一款AI推理解决方案，将存内计算（PIM）技术集成到其低功耗内存芯片LPDDR5X中。
该公告基于三星5月以学术论文形式发布的研究，详细介绍了其LP5X-PIM模拟器。8月的发布预计将更进一步，展示关于速度提升和能效的具体性能数据，以及硬件演示和商业化路线图。
PIM技术旨在解决数据在CPU或GPU与内存之间来回移动时出现的瓶颈。其基本思想是将称为PIM逻辑的计算单元直接嵌入内存内部，使计算在数据所在位置进行，并最大限度地减少来回传输数据的需求。
PIM技术本身并不新鲜。它已持续用于高带宽内存（HBM）。但三星似乎将成为首家将PIM与专为低功耗设计的LPDDR系列芯片结合，并专门应用于AI推理的公司。这表明该技术已远远超出研究阶段，现在瞄准在智能手机等端侧AI产品中的近期部署。
在LPDDR5X-PIM的模拟中，三星发现该芯片执行称为GEMV的矩阵计算的速度比传统内存快6.2倍。实际上，这意味着分析照片或处理文本等AI任务可以运行快六倍以上。
三星正在推行双轨战略，使用HBM瞄准服务器市场，使用LPDDR5X-PIM瞄准个人设备。与HBM不同，LPDDR5X-PIM面向笔记本电脑和智能手机等对功耗敏感的设备，随着公司寻求解决隐私问题和降低网络延迟，端侧AI在这一细分市场变得越来越重要。
当与三星基于4纳米工艺的内部AI加速器GAIA搭配时，该技术可带来更大的效率提升。在这种设置中，以NPU为中心的芯片GAIA将处理大部分AI计算，而LPDDR5X-PIM在内存层面提供高效支持。
竞争对手也在采取类似方法。SK海力士开发了自己的PIM产品，名为加速器内存（AiM），并将其构建到图形DRAM中，命名为GDDR6-AiM，还推出了AiMX，一种将多个AiM芯片封装在一起的AI加速器。
“能够支持AI PC和旗舰智能手机中实时AI功能的有意义技术现在开始真正涌现，”一位行业消息人士表示。“这对端侧AI市场的增长是个好兆头。”
A block diagram of the LP5X-PIM simulator. 〈Source: “LP5X-PIM Sim: A High-Fidelity HW/SW Integrated Simulator for LPDDR5X-PIM”〉
Samsung Electronics has secured what's believed to be the world's first AI inference technology that embeds computing capability directly into low-power memory, a move expected to accelerate the shift toward a new era of low-power, high-efficiency AI.
According to industry sources on July 22, Samsung Electronics plans to unveil an AI inference solution in August that integrates processing-in-memory (PIM) technology into LPDDR5X, its low-power memory chip.
The announcement builds on research Samsung published in May in the form of an academic paper detailing its LP5X-PIM simulator. The August unveiling is expected to go a step further, presenting concrete performance data on speed improvements and power efficiency, along with a hardware demo and a roadmap toward commercialization.
PIM technology is designed to solve the data bottleneck that occurs when information moves back and forth between a CPU or GPU and memory. The basic idea is to embed a computing unit, known as PIM logic, directly inside the memory itself, allowing calculations to happen where the data lives and minimizing the need to shuttle data back and forth.
PIM technology itself isn't new. It's already been used consistently in high-bandwidth memory (HBM). But Samsung appears set to become the first company to combine PIM with the LPDDR family of chips, which are designed for low power consumption, and apply it specifically to AI inference. That suggests the technology has moved well beyond the research stage and is now aimed at near-term deployment in on-device AI products like smartphones.
In simulations of LPDDR5X-PIM, Samsung found that the chip could perform matrix calculations known as GEMV up to 6.2 times faster than conventional memory. In practical terms, that means AI tasks like analyzing photos or processing text could run more than six times faster.
Samsung is pursuing a two-track strategy, using HBM to target the server market and LPDDR5X-PIM to target personal devices. Unlike HBM, LPDDR5X-PIM is aimed at power-sensitive devices like laptops and smartphones, a segment where on-device AI has become increasingly important as companies look to address privacy concerns and reduce network latency.
The technology could deliver even greater efficiency gains when paired with GAIA, Samsung's in-house AI accelerator built on a 4-nanometer process. In that setup, GAIA, an NPU-centered chip, would handle the bulk of AI computation, while LPDDR5X-PIM provides efficient support at the memory level.
Rivals are pursuing similar approaches. SK Hynix has developed its own PIM product called Accelerator-in-Memory (AiM), building it into graphics DRAM under the name GDDR6-AiM, and has also unveiled AiMX, an AI accelerator that packs together multiple AiM chips.
“Meaningful technologies capable of supporting real-time AI features in AI PCs and flagship smartphones are now starting to emerge in earnest,” an industry source said. “This bodes well for the growth of the on-device AI market.”