The IBM Analog AI team is global team focusing on advancing the forefront of in-memory computing technologies to overcome the Von-Neumann Bottleneck. By performing vector-matrix multiplication (VMM) at the location where data is stored, we save time and energy by reducing data communication between memory and the compute unit. This efficient VMM computation is particularly powerful in deep neural network (DNN) applications, where the VMMs can tolerate a certain level of errors, for example, due to reduced-precision or noisy computations, without effecting accuracy of the DNN model. The team demonstrated near software-equivalent DNN accuracies both in phase change memory (PCM) hardware and in software simulations. We published high profile papers with >7000 citations since 2015 and authored numerous patents every year.
The Analog AI global team includes members from Almaden/California, Yorktown/New York, Albany/New York, Tokyo, and Zurich. The Almaden Research team holds technical leadership roles in a broad range of topics, including PCM hardware testing, circuit design, algorithm simulations, application exploration, architecture definition, and software development. The team has hosted many intern students and visiting scholars in the past from around the world, including US, Europe, Brazil, Japan, and Taiwan. Some students are now regular employees at IBM and many continued to advance their careers in related fields.
As part of the IBM Research Semiconductor team, you will conduct world-class research on AI Hardware using in-memory computing for Deep Neural Network acceleration. The IBM Almaden Analog AI team published an in-memory computing (IMC) chip with phase change memory (PCM) integrated in the metal stack on top of 14nm CMOS circuitry for Deep Learning acceleration. (reference: https://www.nature.com/articles/s41586-023-06337-5) This chip, containing 1 million PCM devices per tile and 34 tiles per chip, is an important stepping stone towards a scalable and configurable architecture. (reference: https://ieeexplore.ieee.org/abstract/document/9957094)
In this internship, the student will develop software stack and/or hardware components for an in-memory computing architecture. Research could include Deep Neural Network accuracy simulations, hardware demonstrations, and power performance modeling. The student will gain hands-on experience implementing DNN workloads, such as CNN and Transformers, for deployment onto the IMC chip and in simulation using the AIHWKIT. (reference: https://github.com/IBM/aihwkit)
From June 16 to September 5, 2025 (adjustable at the discretion of the organisation)