FluxML/Flux.jl

Benchmark optimizer latency on GPU

Open

#1,699 opened on Aug 24, 2021

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Julia (4,725 stars) (619 forks)batch import
cudahelp wantedperformance

Description

Ref. https://github.com/JuliaGPU/GPUCompiler.jl/issues/65#issuecomment-903155469. We should do this for both the current mutating optimizers and for Optimisers.jl. It may be that TTFG is only part of the problem.

Contributor guide

Tech stack
None
Domain
machine learningperformance
Issue type
performance
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
JuliaGPU programmingFluxbenchmarking
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Start by reading the linked comment in GPUCompiler.jl (issue #65). Investigate the current mutating optimizers in Flux and compare with Optimisers.jl. The goal is to measure latency on GPU, possibly using benchmarking tools like BenchmarkTools.jl. Consider both the time to first gradient and overall iteration performance.