Memory-Centric Computing: Recent Advances in Processing-in-DRAM: IEDM Invited Talk - 09.12.2024
Why It Matters
By eliminating costly data movement, processing‑in‑DRAM delivers dramatic energy savings and performance gains, reshaping the economics of data‑intensive workloads in cloud, edge, and AI environments.
Key Takeaways
- •Data movement dominates energy consumption in modern computing systems.
- •Processing-in-DRAM (PIM) can cut copy latency by >10×.
- •Bitwise majority operations enable programmable logic inside DRAM rows.
- •Commodity DRAM can support PIM with minimal controller modifications.
- •PIM promises 4‑12× speedups for database and ML workloads.
Summary
The invited IEDM talk highlighted memory‑centric computing as a response to exploding data volumes and the growing energy cost of moving that data. The speaker argued that today’s processor‑centric designs waste up to 90% of system energy on memory accesses, citing Google’s data‑center studies and mobile‑device measurements that show 60‑90% of power consumed by data movement rather than computation.
Key insights included the stark performance and energy penalties of traditional data movement, the feasibility of performing operations directly inside DRAM, and concrete experimental results. By exploiting DRAM’s internal sense amplifiers, a 1,000‑page copy can be completed in 90 ns with 0.04 µJ, a 10‑plus‑fold latency reduction and two orders of magnitude energy savings. Triple‑row activation enables bitwise majority logic, which can be composed into arbitrary Boolean functions, delivering 4‑12× speedups in database queries and comparable gains in machine‑learning kernels.
The speaker emphasized that these techniques require only modest changes to the memory controller—no redesign of the DRAM chip itself—making PIM deployable on commodity hardware. Programmable micro‑programs stored in memory translate high‑level operations into sequences of DRAM commands, allowing developers to offload copy, initialization, convolution, and other workloads without rewriting application code.
The broader implication is a paradigm shift toward data‑centric architectures that place compute where data resides, promising substantial reductions in energy use, lower latency, and improved scalability for AI, genomics, and edge analytics. As data growth outpaces Moore’s law, such memory‑centric approaches could become a cornerstone of sustainable high‑performance computing.
Comments
Want to join the conversation?
Loading comments...