252-0028 · Lecture 23GPU Architectures II & Memory Overview and Technology (ETH Zürich, Digital Design and Computer Architecture, Lecture 20)Onur Mutlu · ETH Zürich1:51:12
Opening GPU Architectures II & Memory Overview and Technology (ETH Zürich, Digital Design and Computer Architecture, Lecture 20)
GPU Architectures II & Memory Overview and Technology (ETH Zürich, Digital Design and Computer Architecture, Lecture 20)
Professor of Computer Science at ETH Zurich, ETH Zürich · 63 lectures here
Mutlu merges diverged GPU threads through dynamic warp formation, then descends DRAM from channels to rows and columns.
Read moreShow fewer
GPU branch divergence and DRAM organization are the two halves of Mutlu's session in the ETH Zürich Digital Design and Computer Architecture course, Spring 2025. He starts with mask bit vectors and dynamic warp formation (Fung et al., MICRO 2007), showing how lane conflicts in partitioned register files limit thread merging. Two-level warp scheduling, the growth of Nvidia streaming multiprocessors from the GTX 200 to the V100, and dense and sparse tensor cores follow. The second part argues that memory dominates latency, energy and reliability. It then builds large memories from banked hierarchical arrays, compares SRAM and DRAM cells, and traces the DRAM subsystem through modules, ranks, chips and banks. The session assumes pipelining, SIMD and vector execution, and decoders.
Mutlu merges diverged GPU threads through dynamic warp formation, then descends DRAM from channels to rows and columns.
GPU branch divergence and DRAM organization are the two halves of Mutlu's session in the ETH Zürich Digital Design and Computer Architecture course, Spring 2025. He starts with mask bit vectors and dynamic warp formation (Fung et al., MICRO 2007), showing how lane conflicts in partitioned register files limit thread merging. Two-level warp scheduling, the growth of Nvidia streaming multiprocessors from the GTX 200 to the V100, and dense and sparse tensor cores follow. The second part argues that memory dominates latency, energy and reliability. It then builds large memories from banked hierarchical arrays, compares SRAM and DRAM cells, and traces the DRAM subsystem through modules, ranks, chips and banks. The session assumes pipelining, SIMD and vector execution, and decoders.
Is it for you
Summary written with AI from the lecture and its description
What you will take away
1Branch divergence in GPUs reduces SIMD utilization because inactive threads must be masked off during execution.
3Memory access is significantly more expensive in both latency and energy compared to arithmetic operations.
4Monolithic memory arrays are divided into independent banks and hierarchical structures to reduce access latency and enable parallelism.
5DRAM cells use an access transistor and a capacitor to store bits, requiring periodic refreshing due to charge leakage.
Before you start
Good fit if
Students of computer architecture who know about SIMD processing, program counters, and basic execution pipelines. The recording includes a significant intermission/break in the middle starting around 52:00. Explains the material in order, conversational in tone. Takes questions along the way.
Keep goingMemory Organization, Technology, Hierarchy & Caches (ETH Zürich, Digital Design and Computer Architecture, Lecture 21)Next in the course, 24 of 35.
GPU Architectures II & Memory Overview and Technology (ETH Zürich, Digital Design and Computer Architecture, Lecture 20)
Lecture 23 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.