The CPU, GPU, and H100 Boundary | Berkeley CS285 Topic Guide

TL;DRFive assignments move from CPU-friendly imitation learning to H100-based LLM RL and six-hour offline-RL runs; self-learners should use three compute tiers instead of copying the entire enrolled workflow.

🌏 中文版

Video status: Official entry or recording index only. Source details

The public Spring 2026 starter repository contains HW1–5 and code for two default final projects under an MIT license. That makes independent implementation possible.

The series overview owns the complete public-versus-enrolled access boundary. This article owns compute and project tradeoffs.

Compute ledger

Work	Implementation focus	Official compute signal	Self-study choice
HW1	MSE, DAgger, flow-matching imitation	README found local CPU faster in testing	Complete on CPU
HW2	Policy gradients, reward-to-go, baseline	Seconds to about ten minutes per run	Complete on CPU; add seeds
HW3	DQN and SAC	MsPacman and HalfCheetah may take about three GPU hours	Do small environments; choose one expensive task
HW4	REINFORCE, GRPO, LLM rewards	Modal wrapper defaults to H100; four required runs	Reduce format-copy first; set a cloud budget
HW5	SAC+BC, IQL, FQL, OGBench	One run may take about six hours; tuning accumulates	Start with one task and seed

These times are estimates for the official environments, not guarantees across hardware. Cloud pricing changes, so check current rates and define a stopping condition before launching paid resources.

A three-tier route

First, complete HW1 and HW2 and establish logging, evaluation, and multi-seed habits. Second, run HW3 on cheap environments and choose at most one expensive environment. Third, select one reduced experiment from each of HW4 and HW5, with a time or spending cap.

Record environment versions, seeds, commands, wall-clock time, hardware, and failed runs. Saving only the best curve discards the reproducibility lesson that matters most in RL.

Choosing a final project

The public final project outline accompanies default Offline-to-Online RL and LLM RL options. The first extends a fixed dataset with limited interaction. The second involves preference data, reward modeling, and policy optimization. Course documents mention enrolled-student Modal resources and an H200 option; neither is a public entitlement.

A self-study project can be one baseline, one modification, three seeds, one principal plot, and a limitations section. Without a GPU, choose a small simulated offline-to-online setting rather than assuming LLM training costs.

Delivery boundary

The public path can reproduce a self-evaluated version of the assignments. It cannot claim completion of the Berkeley credit-bearing course. The series overview maintains the full list of video, discussion, grading, and course-support limits.

Course video sources

This post covers the homeworks and the final project, with no single corresponding lecture. Checked live on 2026-10-10: the instructor published the Spring 2026 lecture recordings on the RAIL YouTube channel (27 videos). The playlist has lectures and guest talks only, no homework walkthroughs. The course syllabus still says recordings are on bCourses and the course site still links the Fall 2023 playlist.

Course and recording entries:

Checked: 2026-10-10.

Update Log

2026-10-10: Added explicit video status and checked recording sources and access notes.
2026-10-10: Rechecked video status. Found the public Spring 2026 playlist on the RAIL channel and linked it; this post has no single lecture, so the status stays official entry only.