Digital Design & Comp. Arch: L20: GPU Architectures (Spring 2026)

Digital Design & Comp. Arch: L20: GPU Architectures (Spring 2026)

by Onur Mutlu

Formal & Physical Sciences Computing & Cybersecurity UYComputer scienceUYFComputer architecture and logic design
🎙 Onur Mutlu 👥 64K 📅 May 8, 2026 ⏱ 106 min 👁 3K 📄 lecture 🧭 2026-08-15
Available in: English (current) Français

Keywords

GPUSIMDvector processingarray processingprogramming model

Summary

This lecture, part of the Digital Design and Computer Architecture course at ETH Zürich, provides a comprehensive introduction to GPU architectures. Professor Onur Mutlu begins by revisiting key concepts from previous lectures on SIMD processing, including array and vector processors, memory banking, and bank conflicts. He emphasizes that GPUs are essentially flexible SIMD engines that combine both array and vector processing to exploit data parallelism in both space and time. The lecture highlights the importance of the programming model, contrasting it with the underlying execution model, and introduces the SPMD (Single Program, Multiple Data) model as the basis for GPU programming. Mutlu discusses the historical context of SIMD extensions in CPUs, such as Intel MMX and AVX, and their role in accelerating multimedia and matrix operations. He also touches on other SIMD-based accelerators like Cerebras wafer-scale engines and FPGA-based neural network accelerators. The core of the lecture focuses on how GPUs execute threads in lockstep, using warp scheduling, and the challenges of memory bandwidth and bank conflicts. He outlines the GPU memory hierarchy, including global, shared, and local memories, and explains the importance of coalesced memory access. The lecture concludes with a discussion on the serial bottleneck and Amdahl’s law, emphasizing that GPUs must also handle scalar code efficiently. Throughout, Mutlu provides references to recommended readings and related lectures, making this a valuable resource for students and researchers.

228 words

Critical Evaluation

Value of the Information & Strength of the Argument

The lecture provides high-value information by systematically building on prior knowledge, clearly explaining the evolution from SIMD to GPU architectures. The argumentation is solid, grounded in established computer architecture principles and supported by concrete examples and historical context. Mutlu effectively contrasts programming models with execution models, clarifying why GPUs use a thread-based programming model while executing in a SIMD fashion. He also addresses practical challenges such as memory bandwidth and bank conflicts, offering insights into design trade-offs. The inclusion of recent research directions, like memory-centric computing, adds depth and relevance.

Scientific Rigor, Source Quality, Title Accuracy

The lecture demonstrates high scientific rigor, with clear references to seminal papers and textbooks. The slides and recommended readings are provided, including recent arXiv papers and publications from the lecturer’s group. The title accurately reflects the content, which is focused on GPU architectures. The lecture is well-structured, with logical progression and clear explanations. The sources cited are credible and directly relevant to the topic, enhancing the overall reliability of the content.

176 words

Title / Content Match

The title accurately reflects the content, which focuses on GPU architectures, building on previous lectures on SIMD processing.

Quality & Reliability

9/10

Lecture by a leading academic in computer architecture, with slides and recommended readings from peer-reviewed sources. Content is well-structured, technically accurate, and based on established principles and recent research.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

External References

Contribution & Novelties

The lecture provides a comprehensive and up-to-date overview of GPU architectures, bridging fundamental SIMD concepts with modern GPU design. It emphasizes the importance of the programming model and its alignment with the execution model, a key insight for understanding GPU performance. The lecture also highlights recent trends in memory-centric computing and the challenges of memory bandwidth, which are critical for future GPU designs.

Pour aller plus loin :

  • GPU Computing — Overview of GPGPU computing.
  • CUDA — NVIDIA’s parallel computing platform and programming model.
  • SIMD — Explanation of single instruction, multiple data.
  • Amdahl’s law — Fundamental law in parallel computing.
  • Memory bandwidth — Key performance factor in GPUs.

108 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-balanced and comprehensive lecture. The high scores in information quantity and quality reflect the depth and accuracy of the content, while the strong technical level and reliability underscore its value for an academic audience.

Reliability 9/10