Memory-Centric Computing (Keynote Talk at IDEAS Center) - Prof. Onur Mutlu (14.05.2024)
Why It Matters
Memory‑centric designs dramatically cut energy and latency, enabling scalable AI and genomics while reducing operational costs—critical for enterprises facing exploding data volumes.
Key Takeaways
- •Data movement dominates performance and energy costs in modern systems.
- •Current processor‑centric designs waste up to 95% of potential efficiency.
- •Memory‑centric architectures can cut energy consumption by orders of magnitude.
- •Real workloads reveal 60‑90% of system energy spent on memory.
- •Shifting compute to memory/storage is vital for AI and genomics.
Summary
Professor Onur Mutlu’s keynote at the IDEAS Center introduced memory‑centric computing—also called in‑memory or processing‑in‑memory—as a necessary evolution beyond today’s processor‑centric architectures. He argued that modern workloads such as machine learning, genomics, and large‑scale databases are fundamentally limited by the cost of moving data between separate compute and storage units, creating a performance and energy bottleneck that threatens both scalability and sustainability.
Mutlu highlighted stark efficiency figures: a flagship Alpha processor achieved only 5% utilization on a database workload, while Google’s 2015 study showed top‑of‑the‑line CPUs completing merely 10‑20% of instructions due to memory stalls. Energy disparities are even more dramatic—one 64‑bit floating‑point operation consumes roughly 20 pJ, whereas a single DRAM access can require 800 × that energy, and simple 32‑bit adds can be 6,400 × more costly than the computation itself. In real‑world mobile and edge AI workloads, over 60% of total system energy is spent on data movement, climbing to 90%+ for large neural‑network models.
He illustrated the problem with concrete examples: wafer‑scale chips still lack sufficient memory density, and inexpensive nanopore sequencers now generate more data than today’s transistor budgets can process. Citing Dick Sites’ observation that memory subsystem design will dominate microprocessor design for the next decade, Mutlu emphasized that the industry’s reliance on ever‑larger cache hierarchies and out‑of‑order execution merely masks, not solves, the underlying imbalance between compute and memory.
The implication for businesses is clear: without a shift to memory‑centric architectures—whether through processing‑in‑memory, near‑data accelerators, or smarter storage‑side compute—companies will face escalating energy bills, limited AI performance, and unsustainable growth in data‑intensive applications. Investing in hardware that co‑locates compute with memory promises orders‑of‑magnitude energy savings and unlocks the next wave of AI and bioinformatics innovation.
Comments
Want to join the conversation?
Loading comments...