The paper discusses the challenges and opportunities of reducing data movement in modern computing systems by performing processing in memory (PIM). Traditional systems suffer from a significant energy and latency cost due to frequent data movement between computation units and memory. PIM aims to address this by placing computation mechanisms near or within memory, reducing or eliminating data movement. The paper explores two new approaches to enabling PIM: 1) using the analog properties of DRAM to perform massively-parallel operations in memory, and 2) leveraging 3D-stacked memory technology to provide high bandwidth to in-memory logic. The paper also discusses the challenges of adopting PIM, including programming models, runtime systems, coherence between PIM logic and CPU/accelerator cores, virtual memory support, and data structure design. It presents two examples of PIM applications: Ambit, which enables in-DRAM bulk bitwise operations, and Tesseract, a PIM accelerator for large-scale graph processing. Ambit significantly improves performance and energy efficiency for applications involving bitwise operations, while Tesseract improves system performance and reduces energy consumption for graph processing tasks. The paper concludes that PIM has the potential to significantly reduce data movement and improve performance and energy efficiency in modern computing systems. However, practical adoption of PIM requires addressing key challenges, including programming models, system software, and hardware design. The paper also highlights the importance of exploring low-cost PIM substrates and more sophisticated computational substrates for all types of memory technologies.The paper discusses the challenges and opportunities of reducing data movement in modern computing systems by performing processing in memory (PIM). Traditional systems suffer from a significant energy and latency cost due to frequent data movement between computation units and memory. PIM aims to address this by placing computation mechanisms near or within memory, reducing or eliminating data movement. The paper explores two new approaches to enabling PIM: 1) using the analog properties of DRAM to perform massively-parallel operations in memory, and 2) leveraging 3D-stacked memory technology to provide high bandwidth to in-memory logic. The paper also discusses the challenges of adopting PIM, including programming models, runtime systems, coherence between PIM logic and CPU/accelerator cores, virtual memory support, and data structure design. It presents two examples of PIM applications: Ambit, which enables in-DRAM bulk bitwise operations, and Tesseract, a PIM accelerator for large-scale graph processing. Ambit significantly improves performance and energy efficiency for applications involving bitwise operations, while Tesseract improves system performance and reduces energy consumption for graph processing tasks. The paper concludes that PIM has the potential to significantly reduce data movement and improve performance and energy efficiency in modern computing systems. However, practical adoption of PIM requires addressing key challenges, including programming models, system software, and hardware design. The paper also highlights the importance of exploring low-cost PIM substrates and more sophisticated computational substrates for all types of memory technologies.