The Next Computing Revolution: Bringing Processing Inside Memory (www.computer.org)

Onur Mutlu argues that the next major shift in computing is memory-centric or processing-in-memory (PIM): moving computation into or next to high-density memories (DRAM, flash) to dramatically cut the massive data-movement costs that dominate modern systems—he cites studies showing >90% of energy for major AI models is spent on moving and accessing data, not on arithmetic. PIM could yield orders-of-magnitude improvements in energy efficiency, performance, robustness, security and sustainability for data‑intensive AI/ML workloads, and several startups and vendors (e.g., UPMEM) are already shipping memory chips with compute capability. Mutlu stresses that the biggest non-technical barrier is the entrenched processor-centric mindset across education, tools and business models; the software stack must evolve to let developers leverage PIM without onerous rewrites. Technically, Mutlu highlights concrete advances and validation infrastructure: experiments showing off‑the‑shelf DRAM can perform functionally complete bulk bitwise operations via timing-parameter manipulation (PiDRAM, HPCA/IEDM 2024), and a multi-pronged research methodology spanning real chip testbeds (SoftMC, DRAM Bender), circuit-level models (CLRDRAM), high-level simulators (Ramulator, Virtuoso, MQSim, Pythia, DAMOV) and prototyping platforms (PiDRAM, Sibyl). These open-source tools enabled discoveries like RowHammer and TRNG in DRAM and underpin efforts to build robust PIM-aware compilers, runtimes and benchmarks—work Mutlu says is crucial for mainstream adoption.