CS294-252: Architectures and Systems for Hyperscale Cloud Datacenters in the Era of Agentic AI

Fall 2025, UC Berkeley

Location: Tuesdays and Thursdays from 2pm-3:30pm in 405 Soda

Course Overview: Warehouse-Scale Computers (WSCs) host hyperscale cloud services relied on by billions of daily users and power the latest advances in AI/ML, data processing, and web services. While classical WSCs were built as homogeneous collections of servers and networking hardware, modern hardware scaling trends and exponential increases in demand for AI/ML compute have necessitated the introduction of specialized hardware in datacenter environments, including ML accelerators and ML “supercomputer pods”, SmartNICs, GPUs, and custom server SoCs. The challenge of designing these HW/SW systems is vast and ever-growing, but also critical to enabling the continued advancement of AI-powered applications.

This graduate-level course will explore two major themes:

  • How do we architect hardware-software systems at scale to support efficient, practical, AI-powered application pipelines, end-to-end? (i.e. more than just the math)
  • How can AI help us wrangle complexity in designing these HW/SW systems to meet exponential demand, from chip to datacenter and beyond?

Prerequisites: Students must satisfy the following requirements to enroll:

Completion of at least one of: CS252, CS262, CS268, EECS251.

OR

Completion of at least two of: CS152, EECS151, CS162, CS168.

Additionally, if you are an undergraduate, 5th-year master’s, or concurrent enrollment student, please fill out the following form to be considered for enrollment: https://forms.gle/qWfFdmeVUGK2PpJT9.

Calendar / Reading List

August 28
Intro to Warehouse-Scale Computers
Reading 1
L. Barroso, et. al. The Datacenter as a Computer, Third Edition.

September 2
Datacenter-Wide Trends and Workloads
Reading 1
S. Kanev, et. al. Profiling a Warehouse-Scale Computer.
Reading 2
W. Su, et. al. DCPerf: An Open-Source, Battle-Tested Performance Benchmark Suite for Datacenter Workloads.

September 4
Power Management
Reading 1
V. Sakalkar, et. al. Data Center Power Oversubscription with a Medium Voltage Power Plane and Priority-Aware Capping.
Reading 2
P. Patel, et. al. Characterizing Power Management Opportunities for LLMs in the Cloud.

September 9
WSC Networking
Reading 1
L. Poutievski, et. al. Jupiter Evolving: Transforming Google’s Datacenter Network via Optical Circuit Switches and Software-Defined Networking.
Reading 2
D. Firestone, et. al. Azure Accelerated Networking: SmartNICs in the Public Cloud.

September 11
Datacenter-Wide Trends and Workloads Pt. 2
Reading 1
J. Dean, et. al. The tail at scale. +
L. Barroso, et. al. Attack of the Killer Microseconds.
Reading 2
K. Seemakhupt, et. al. A Cloud-Scale Characterization of Remote Procedure Calls.

September 16
Accelerators in WSCs, Pt. 1
Reading 1
I. Magaki, et. al. ASIC Clouds: Specializing the Datacenter.
Reading 2
N. Jouppi, et. al. TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings.

September 18
Accelerators in WSCs, Pt. 2
Reading 1
A. Putnam, et. al. A Reconfigurable Fabric for Accelerating Large-Scale Datacenter Services.
Reading 2
C. Zhao, et. al. Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures.

Weekly Schedule

  • Lecture/Discussion: Tuesdays and Thursdays from 2pm-3:30pm in 405 Soda
  • Weekly Reading Reviews: See Ed for submission links.
    • Due Mondays @ noon pacific for Tuesday lecture papers.
    • Due Wednesdays @ noon pacific for Thursday lecture papers.
  • Weekly Student Presenter Slides: Check your email for submission instructions.
    • Due Fridays @ noon pacific for Tuesday lecture presentations.
    • Due Tuesdays @ noon pacific for Thursday lecture presentations.

Assignments and Grading

The course workload will consist of the following:

  • 25% of grade: Each class, students will be required to read and provide a review of the two papers for that day and attend and participate in the class discussion.
    • Can drop two classes’ worth, no questions asked.
  • 25% of grade: Each student will lead the discussion of a few papers during the semester.
  • 50% of grade: Students will complete a semester-long research project, in groups of 2 or 3, related to the course material.

Instructor