This site is fictional demo content. It is not real news or affiliated with any real organization. Do not treat it as fact or professional advice.

NextPaper Logo
nextpaper.ccCYBER CHRONICLE
RSS
EN中文

Full article

FULL TEXT

View this issue
BriefTECH

Computing-in-Memory Chips CIM-7nm Enter Mass Production: AI Inference Energy Efficiency Up 50x

Samsung and TSMC simultaneously begin mass production of 7nm Computing-in-Memory chips that integrate computation directly into memory arrays, achieving 50x energy efficiency improvement over traditional von Neumann architectures for AI inference.

Computing-in-Memory Chips CIM-7nm Enter Mass Production: AI Inference Energy Efficiency Up 50x

Samsung Semiconductor and TSMC today simultaneously announced the mass production of 7nm Computing-in-Memory (CIM) chips. CIM-7nm embeds compute units directly into SRAM arrays, eliminating the bottleneck of data shuttling between memory and processors, achieving a 50x energy efficiency improvement over traditional von Neumann architectures for AI inference tasks.

Samsung Electronics Semiconductor President Kyung Kye-hyun said: "The CIM architecture fundamentally solves the 'memory wall' problem. Data no longer needs to move—computation happens where the data lives. This is the most significant paradigm shift in semiconductor architecture in 30 years."

A single CIM-7nm chip delivers 48 TOPS of AI inference performance at just 1.2 watts of power, making it ideal for edge AI devices. First customers include DJI (for next-generation drone vision processing) and Tesla (for Optimus robot local inference).

TSMC simultaneously launched its own CIM-7nm process node based on a different technical approach—ReRAM (Resistive Random-Access Memory) rather than Samsung's SRAM approach. TSMC VP of Business Development Kevin Zhang said: "Competition between two technical routes will accelerate the maturation of computing-in-memory technology."

Market research firm Yole Intelligence projects the CIM chip market will grow from $1.2 billion in 2028 to $18 billion by 2032, driven primarily by edge AI devices' demand for low-power, high-performance computing.

However, CIM chips currently support only inference workloads—not training—and their programming model differs fundamentally from traditional chips, requiring time for the software ecosystem to develop. Nvidia Chief Scientist Bill Dally commented: "CIM is an interesting technology, but the general-purpose GPU's advantages in flexibility and software ecosystem won't be displaced in the short term."