[Applied
Materials Blog] Industries from transportation and healthcare to retail
and entertainment will be transformed by the Internet of Things, Big
Data and Artificial Intelligence (AI), which Applied Materials
collectively calls the AI Era of Computing.
The previous computing eras—Mainframe/Minicomputer, PC/Server and Smartphone/Tablet—all benefitted from advances in Moore’s Law whereby 2D scaling was accompanied by simultaneous improvements in performance, power and area/cost—also called “PPAC.”
While AI Era applications are booming, Moore’s Law is slowing; as a result, the industry needs breakthroughs beyond 2D scaling to drive PPAC in new ways. Specifically, we need new computing architectures, new materials, new structures—especially area-saving 3D structures—and advanced packaging for die stacking and heterogeneous designs.
The AI Era is Driving a Renaissance in Semiconductor Innovation (Applied Materials Blog)
AI
Era architectural changes are influencing both logic and memory.
Machine learning algorithms make heavy use of matrix multiplication
operations that are cumbersome in general-purpose logic, and this is
driving a move to accelerators and their memories. AI compute includes
two distinct memory tasks: first, storing the intermediate results of
calculations; and second, storing the weights associated with trained
models.
Performance and power are important in the cloud and in the edge, and innovations in memory can help. One approach using existing memory technologies is “near memories” whereby large amounts of working memory are condensed, placed in close physical proximity to logic, and connected via high-speed interfaces. As examples, 3D stacking and through-silicon vias are gaining traction. One major drawback of SRAM and DRAM as “working memories” in these applications is that they are volatile and need a constant supply of power to retain data—such as weights.
To reduce power in the cloud and edge, designers are evaluating new memories that combine high performance with non-volatility so that power is only needed during active read and write operations. Three of the leading new memory candidates are magnetic random-access memory (MRAM), phase-change RAM (PCRAM) and resistive RAM (ReRAM).
Performance and power are important in the cloud and in the edge, and innovations in memory can help. One approach using existing memory technologies is “near memories” whereby large amounts of working memory are condensed, placed in close physical proximity to logic, and connected via high-speed interfaces. As examples, 3D stacking and through-silicon vias are gaining traction. One major drawback of SRAM and DRAM as “working memories” in these applications is that they are volatile and need a constant supply of power to retain data—such as weights.
To reduce power in the cloud and edge, designers are evaluating new memories that combine high performance with non-volatility so that power is only needed during active read and write operations. Three of the leading new memory candidates are magnetic random-access memory (MRAM), phase-change RAM (PCRAM) and resistive RAM (ReRAM).
Full article: Applied Materials Blog LINK
Additional read: Manufacturing Requirements of New Memories LINK
No comments:
Post a Comment