<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa" -->

---
title: AMDGPU.jl 2.8–2.12: ROCm as artifacts, a performance...
description: AMDGPU.jl 2.8 through 2.12 no longer require a separate ROCm install: starting with 2.12, ROCm runtime and libraries are downloaded as Julia artifacts from...
canonical: https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: AMDGPU.jl 2.8–2.12: ROCm as artifacts, a performance dashboard, and a faster core | daily.dev
og:description: AMDGPU.jl 2.8 through 2.12 no longer require a separate ROCm install: starting with 2.12, ROCm runtime and libraries are downloaded as Julia artifacts from...
og:url: https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa
og:image: https://api.daily.dev/og/posts/3vsX8OjXa.png
og:image:alt: AMDGPU.jl 2.8–2.12: ROCm as artifacts, a performance dashboard, and a faster core
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# AMDGPU.jl 2.8–2.12: ROCm as artifacts, a performance dashboard, and a faster core

**[Julia Bloggers](https://daily.dev/sources/juliabl)** · 6 min read · 1 upvotes · 0 comments

## Summary

AMDGPU.jl 2.8 through 2.12 no longer require a separate ROCm install: starting with 2.12, ROCm runtime and libraries are downloaded as Julia artifacts from AMD's TheRock distribution, so only the GPU driver is needed. A public performance dashboard now tracks benchmarks on every commit, shared with CUDA.jl, Metal.jl and KernelAbstractions.jl. Kernel launch time dropped from about 1.3 µs to 0.9 µs, and GPUArrays-based reductions got up to 12 to 20 times faster on large matrices. The default SVD algorithm switched back to QR iteration after the faster Jacobi method was found to silently lose small singular values. Other additions include hipTensor wrappers, device-side rocBLAS scalars, and fixes for GPU-aware MPI on ROCm 7.1+.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://www.juliabloggers.com/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core>

## Questions this post answers

### Do I still need to manually install ROCm to use AMDGPU.jl?

No, starting with AMDGPU.jl 2.12 the ROCm runtime and libraries (rocBLAS, rocSOLVER, rocSPARSE, rocFFT, rocRAND, MIOpen, hipTensor) download automatically as Julia artifacts from AMD's TheRock distribution. Only the GPU driver needs to be installed separately. A local ROCm 6.0+ install can still be used instead via AMDGPU.set_rocm_version!(local_rocm=true).

_Track changes like this before upgrading a GPU toolchain by following Julia and ROCm releases on daily.dev._

### Why did AMDGPU.jl change the default SVD algorithm away from Jacobi?

Because rocSOLVER's Jacobi SVD algorithm silently loses small singular values (below sqrt(eps) times the largest singular value) without raising an error, making svd, cond, or rank potentially inaccurate. Since AMDGPU.jl 2.12, svd and svdvals default to QR iteration, matching LAPACK's behavior, while Jacobi remains available as an opt-in for well-conditioned matrices.

_Developers choosing GPU linear algebra algorithms can watch for accuracy trade-offs like this via daily.dev._

### How much faster are kernel launches in AMDGPU.jl after recent optimizations?

Kernel launch time dropped from about 1.3 microseconds in version 2.7 to about 0.9 microseconds, due to allocating once instead of 12 times per launch and caching kernel lookups. Short kernels also got roughly twice as fast after switching to reading grid dimensions from hidden kernel arguments, and KernelAbstractions launches skip dynamic dispatch.

_daily.dev helps developers optimizing GPU workloads keep up with performance-focused library updates._

## Similar posts on daily.dev

- [AMDGPU.jl 2.6 and 2.7: linear algebra, sparse arrays, and RDNA4 matrix cores](https://daily.dev/posts/amdgpu-jl-2-6-and-2-7-linear-algebra-sparse-arrays-and-rdna4-matrix-cores-pywghwbds) · Julia Bloggers · 0 upvotes · 0 comments
- [AMD ROCm 7.12 Tech Preview Brings More Consumer APU & GPU Support](https://daily.dev/posts/amd-rocm-7-12-tech-preview-brings-more-consumer-apu-gpu-support-vv1vqmj2n) · Phoronix · 0 upvotes · 0 comments
- [Better Gaussian Splatting in Julia](https://daily.dev/posts/better-gaussian-splatting-in-julia-n5pedwfwa) · Hacker News · 0 upvotes · 0 comments
- [AMD Jumps From ROCm 7.14 To ROCm 10.0 With ROCm.AI](https://daily.dev/posts/amd-jumps-from-rocm-7-14-to-rocm-10-0-with-rocm-ai-teqeyhyhd) · Phoronix · 0 upvotes · 0 comments
- [State of Julia's GPU ecosystem in 2026](https://daily.dev/posts/state-of-julia-s-gpu-ecosystem-in-2026-p5xt2hw2y) · Julia Bloggers · 0 upvotes · 0 comments

---

Tags: [#performance](https://daily.dev/tags/performance), [#gpu](https://daily.dev/tags/gpu), [#julia](https://daily.dev/tags/julia), [#rocm](https://daily.dev/tags/rocm)

[View this post on daily.dev](https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"AMDGPU.jl 2.8–2.12: ROCm as artifacts, a performance dashboard, and a faster core","url":"https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa"},"datePublished":"2026-11-17T15:45:08.252Z","dateModified":"2026-11-17T16:04:30.458Z","description":"AMDGPU.jl 2.8 through 2.12 no longer require a separate ROCm install: starting with 2.12, ROCm runtime and libraries are downloaded as Julia artifacts from...","image":"https://media.daily.dev/image/upload/s--qPvKM23u--/f_auto/v1722860399/public/Placeholder%2009","thumbnailUrl":"https://media.daily.dev/image/upload/s--qPvKM23u--/f_auto/v1722860399/public/Placeholder%2009","isAccessibleForFree":true,"articleSection":"Julia Bloggers","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Julia Bloggers","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/10a29507451e4542b44721f3be074156","url":"https://daily.dev/sources/juliabl"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"performance,gpu,julia,rocm","timeRequired":"PT6M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Julia Bloggers","item":"https://daily.dev/sources/juliabl"},{"@type":"ListItem","position":3,"name":"AMDGPU.jl 2.8–2.12: ROCm as artifacts, a performance dashboard, and a faster core"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/amdgpu-jl-2-8-2-12-rocm-as-artifacts-a-performance-dashboard-and-a-faster-core-3vsx8ojxa#faq","mainEntity":[{"@type":"Question","name":"Do I still need to manually install ROCm to use AMDGPU.jl?","acceptedAnswer":{"@type":"Answer","text":"No, starting with AMDGPU.jl 2.12 the ROCm runtime and libraries (rocBLAS, rocSOLVER, rocSPARSE, rocFFT, rocRAND, MIOpen, hipTensor) download automatically as Julia artifacts from AMD's TheRock distribution. Only the GPU driver needs to be installed separately. A local ROCm 6.0+ install can still be used instead via AMDGPU.set_rocm_version!(local_rocm=true). Track changes like this before upgrading a GPU toolchain by following Julia and ROCm releases on daily.dev."}},{"@type":"Question","name":"Why did AMDGPU.jl change the default SVD algorithm away from Jacobi?","acceptedAnswer":{"@type":"Answer","text":"Because rocSOLVER's Jacobi SVD algorithm silently loses small singular values (below sqrt(eps) times the largest singular value) without raising an error, making svd, cond, or rank potentially inaccurate. Since AMDGPU.jl 2.12, svd and svdvals default to QR iteration, matching LAPACK's behavior, while Jacobi remains available as an opt-in for well-conditioned matrices. Developers choosing GPU linear algebra algorithms can watch for accuracy trade-offs like this via daily.dev."}},{"@type":"Question","name":"How much faster are kernel launches in AMDGPU.jl after recent optimizations?","acceptedAnswer":{"@type":"Answer","text":"Kernel launch time dropped from about 1.3 microseconds in version 2.7 to about 0.9 microseconds, due to allocating once instead of 12 times per launch and caching kernel lookups. Short kernels also got roughly twice as fast after switching to reading grid dimensions from hidden kernel arguments, and KernelAbstractions launches skip dynamic dispatch. daily.dev helps developers optimizing GPU workloads keep up with performance-focused library updates."}}]}
```

