Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 16 additions & 28 deletions reference-architecture/MI3XX/_static/css/custom.css
Original file line number Diff line number Diff line change
@@ -1,31 +1,7 @@
/* Network troubleshooting page */

/* Custom width for rccl errors table */
table.rccl-errors {
table-layout: fixed !important;
width: 100% !important;
}
table.rccl-errors th:nth-child(1),
table.rccl-errors td:nth-child(1) {
width: 40% !important;
}
table.rccl-errors th:nth-child(2),
table.rccl-errors td:nth-child(2) {
width: 60% !important;
}

/* Custom width for rdma errors table */
table.rdma-errors {
table-layout: fixed !important;
width: 100% !important;
}
table.rdma-errors th:nth-child(1),
table.rdma-errors td:nth-child(1) {
width: 40% !important;
}
table.rdma-errors th:nth-child(2),
table.rdma-errors td:nth-child(2) {
width: 60% !important;
/* Hide the primary sidebar toggle (carat) only when the primary sidebar is
empty. */
.bd-container__inner:has(.bd-sidebar-primary.hide-on-wide) .sidebar-toggle.primary-toggle {
display: none;
}

/* Remove white background from topology diagrams in dark mode */
Expand All @@ -43,3 +19,15 @@ html:not([data-theme=dark]) .bd-content img {
background-color: #777 !important;
}

/* Shrink the Legal information section heading */
#legal-information > h2 {
font-size: 1rem;
}

/* Render the Legal information body text smaller (see .. container:: legal-text
in index.rst) */
.legal-text {
font-size: 0.65rem;
line-height: 1.4;
}

10 changes: 8 additions & 2 deletions reference-architecture/MI3XX/conf.py
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,6 @@
if os.environ.get("READTHEDOCS", "") == "True":
html_context["READTHEDOCS"] = True

project = "AMD Instinct Hub"
html_title = "AMD Instinct MI3XX Reference Design"
author = "Advanced Micro Devices, Inc."
copyright = "Copyright (c) 2026 Advanced Micro Devices, Inc. All rights reserved."
Expand All @@ -19,19 +18,26 @@
external_toc_path = "./sphinx/_toc.yml"

external_projects_current_project = "MI3XX-reference"
external_projects_remote_repository = ""

html_theme = "rocm_docs_theme"
html_theme_options = {
"flavor": "instinct",
"show_toc_level": 1,
"show_toc_level": 2,
"navbar_align": "content",
"link_main_doc": True,
"use_download_button": True,
}
extensions = ["rocm_docs"]

# Configure static files path
html_static_path = ['_static']

# Disable sidebars to handle minimal TOC
html_sidebars = {
"**": []
}

html_extra_path = ["llms.txt"]

EXCLUDED_DIRS = {
Expand Down
643 changes: 640 additions & 3 deletions reference-architecture/MI3XX/index.rst

Large diffs are not rendered by default.

40 changes: 0 additions & 40 deletions reference-architecture/MI3XX/legal-information.rst

This file was deleted.

19 changes: 19 additions & 0 deletions reference-architecture/MI3XX/llms.txt
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
# AMD Instinct MI3XX Reference Design

> A common reference for designing GPU cluster networks using AMD Instinct MI300X, MI325X, MI350X, and MI355X series accelerators, supporting up to 8192 GPUs. Covers cluster design principles, network topologies, scalable architectures, subscription ratios, power requirements, and bill of materials for large-scale deployments, including fat tree and rail network designs with diagrams and best practices for high-performance AI/ML workloads.

## Topics

- [Common cluster design principles](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#common-cluster-design-principles): Fat tree and rail network topologies, how each connects hosts to upstream switches, and a comparison of their latency and cross-rail/cross-rank traffic trade-offs.
- [Basic network topologies](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#basic-network-topologies): Building-block layouts for 2-tier and 3-tier rail, tree, hybrid rail, and fully scheduled rail networks, including TH5/J3 and rail-optimized variants.
- [Scaling networks](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#scaling-networks): When spine and super-spine switches are required as cluster size grows across 2-tier and 3-tier tree, rail, and hybrid designs.
- [Network subscription](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#network-subscription): Subscription ratios and rates that relate downstream demand to upstream capacity, including over- and undersubscription.
- [Hardware and software components](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#hardware-and-software-components): Generic bill of materials for 128–1024 and 1024–8192 GPU cluster and network designs, covering platforms, OS, ROCm, storage, NICs, switches, optics, and fabric.
- [Power requirements](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#power-requirements): Power design assumptions for a 4MW, 2K MI355X GPU cluster with 51.2T or scheduled fabric switch options.
- [Network design examples](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#network-design-examples): Topology design diagrams for tree and rail designs from 128 to 8192 GPUs using Jericho/Ramon scheduled fabrics or 51.2T switches.

## Legal

- [Legal information](https://instinct.docs.amd.com/projects/MI3XX-reference/en/latest/index.html#legal-information): Disclaimer, compliance with export laws, and trademark notices.

---
Loading
Loading