Whole lectures, as used: every JAX lecture on an Apple CPU, a Turing GPU and an x86 CPU

未关闭
#4 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
24/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
jupyter-notebook, python

调研方向

Start with lectures/run_lecture.py and the existing lectures/ benchmarks, then check precision/constraints.txt and the lecture sets in lecture-jax and lecture-python.myst. Run each lecture in a fresh process across the listed platforms and precisions, recording per-cell timings and key outputs. Done means results are saved under lectures/results/, the device comparison table identifies dominant cells, and a summary is posted on meta#350.

由索引模型根据 Issue 内容生成。

描述

Run every JAX lecture as a reader or a CI build runs it, on an Apple CPU, a Turing GPU and an x86 CPU, and record where each lecture's time goes. The aim is to know which lectures run faster on a CPU than on the GPU they are built on, and why.

Why

  • The first two lectures measured point that way. As written, aiyagari_jax ran 18× faster on qe-mini1's M6 CPU than on a TITAN RTX, and ifp_egm_jax's JAX solve 13× faster (lectures/, #3). Both are bound by launching small kernels and syncing in Python loops, not by arithmetic or memory.
  • meta#350, question 1: which lectures should default to the CPU in CI.
  • Readers run the lectures on laptops. A lecture-by-lecture table of CPU against GPU, with the cells that carry the time, tells the lecture series which lectures need a GPU at all, and which could be restructured.
  • It folds in #1's W4, the sargent_surico NUTS cell. That lecture can now be run whole, in float32 and float64, on CPU and GPU: the 2 × 2 that lecture-python.myst#1022 could not separate.

Lectures

  • All 29 lectures in lecture-jax, at a fixed commit. That is every lecture except intro, status, troubleshooting and zreferences.
  • The JAX lectures in lecture-python.myst, sargent_surico among them.
  • Each runs as written. The 13 float64 lectures also run in float32, where the lecture tolerates it (aiyagari_jax does, slowly; see #3).

Platforms

Device Where When
M6, 12 cores, 32 GB (CPU) qe-mini1, with its agents paused First
TITAN RTX, one card qe-wharton1-gpu, card 0b:00.0 First
Threadripper 2990WX (CPU) qe-wharton1-gpu First
M5 Pro, 64 GB (CPU) qe-wharton2 Once it is onboarded in infrastructure
T4 and its host CPU RunsOn g4dn.2xlarge, the lectures' CI instance Once RunsOn is configured for this repository

Method

  • lectures/run_lecture.py: every code cell in one fresh process, compile counted, timed per cell. Three runs per configuration.
  • One environment per platform, pinned to precision/constraints.txt's JAX release. Each lecture's imports are installed beforehand, and the runner skips !pip install cells.
  • Outputs are kept for a cross-device check. Each lecture's key results are recorded with --record, so that a device or precision that changes the answer is flagged.
  • Results go to lectures/results/<host>/<date>/<lecture>/. The dashboard (QuantEcon/project-benchmarks-jax) gets a per-lecture view.

Relation to meta#289

meta#289 tracks lecture build times over releases, on one platform. This is one point in time, across platforms. run_lecture.py's per-cell timings may serve meta#289's tool question.

What the results would change

If Then
Most lectures run as fast or faster on a recent Mac CPU The lectures can say so, and CI can run them on CPU instances. That is meta#350's question 1
A lecture's time sits in a few cells Those cells are candidates for restructuring, which qe:benchmark can assess one lecture at a time
A lecture runs much faster on the GPU It keeps the GPU, and the table says by how much

Deliverables

  • the per-lecture results, under lectures/results/;
  • a table of every lecture by device, with the cells that dominate;
  • a summary comment on meta#350.

🤖 Generated with Claude Code

主要语言
Python
星标
0
派生
1
平均合并
3 小时 29 分钟
30 天内合并 PR
64

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

QuantEcon/benchmarks 的其他 Issue

查看 QuantEcon/benchmarks 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。