Processing in Memory (PIM) - Power and Thermal Challenges and Opportunities
Memory technology is a defining component of modern computing and has a strong impact on performance, power and cost of computing systems. However, the advances in memory performance have not been able to keep up with the performance advances for CPU’s, thus leading to what is known as “the memory wall.” Depending on the application, the memory wall manifests itself both in terms of memory latency, as well as memory bandwidth. An interesting solution to the memory wall problem is to bring memory closer to the processor, or vice-versa, to move some processing capability in the memory itself – this leads to variations of what is known as Near-Memory Processing, Processing in Memory (PiM), etc. This seminar will first introduce a PIM taxonomy along several dimensions of the PiM design space; it will then follow with a history of PIM, then go over several recent PIM examples. The seminar will then go in depth into the Thermal/Power delivery challenges for PIM that are a result of the increased switching activities inherent to the moving of processing into the memory fabric, and exacerbated by the evolution towards 3D integration due to the slow-down of traditional Moore’s law methods. The seminar will conclude with some novel solutions that alleviate the Thermal/Power challenges for PiM.