252-0028 · Lecture 29Prefetching II, Processing-in-Memory, and Parting Thoughts (ETH Zürich: Digital Design and Computer Architecture, Lecture 25)Onur Mutlu · ETH Zürich1:53:02
Opening Prefetching II, Processing-in-Memory, and Parting Thoughts (ETH Zürich: Digital Design and Computer Architecture, Lecture 25)
Prefetching II, Processing-in-Memory, and Parting Thoughts (ETH Zürich: Digital Design and Computer Architecture, Lecture 25)
Hardware prefetching and runahead execution close Onur Mutlu's Digital Design and Computer Architecture course at ETH Zürich, Spring 2025. Mutlu frames prefetching as a reinforcement learning problem, using the Pythia prefetcher's state features, actions, rewards and Q-learning. Pythia's results show gains over traditional prefetchers and adaptation to bandwidth-constrained systems, followed by contention problems in hybrid and multi-core prefetching. Runahead execution then checkpoints architectural state on a long-latency miss and runs ahead speculatively, generating memory-level parallelism without a larger instruction window. The session ends with industry adoption by Sun, IBM and NVIDIA, and a review of design trade-offs and data-centric architectures.
Hardware prefetching and runahead execution close Onur Mutlu's Digital Design and Computer Architecture course at ETH Zürich, Spring 2025. Mutlu frames prefetching as a reinforcement learning problem, using the Pythia prefetcher's state features, actions, rewards and Q-learning. Pythia's results show gains over traditional prefetchers and adaptation to bandwidth-constrained systems, followed by contention problems in hybrid and multi-core prefetching. Runahead execution then checkpoints architectural state on a long-latency miss and runs ahead speculatively, generating memory-level parallelism without a larger instruction window. The session ends with industry adoption by Sun, IBM and NVIDIA, and a review of design trade-offs and data-centric architectures.
Is it for you
Summary written with AI from the lecture and its description
What you will take away
1Prefetching speculative mechanisms must balance accuracy, timeliness, and system memory bandwidth constraints.
2Reinforcement learning allows prefetchers to adapt online to dynamic workload behaviors and memory bandwidth availability.
3Runahead execution speculatively pre-executes code past long-latency cache misses to uncover independent memory requests.
4Memory-level parallelism (MLP) often dictates overall processor performance more effectively than instruction-level parallelism alone.
5Computer architecture trade-offs require evaluating hardware complexity, cycle time, and energy consumption against performance gains.
Before you start
Good fit if
Students of computer architecture who understand caches, out-of-order execution, and basic memory hierarchies. The recording begins with administrative remarks and course logistics before transitioning to the lecture material. Explains the material in order, conversational in tone. Takes questions along the way.
Prefetching II, Processing-in-Memory, and Parting Thoughts (ETH Zürich: Digital Design and Computer Architecture, Lecture 25)
Lecture 29 of 35 of Digital Design and Computer Architecture.
Choose a node to see how it connects to this lecture. What comes before is on the leftat the top. What comes after is on the rightat the bottom. The ring is what it teaches. People and places are above. Other lectures and books are below.