Skip to content

mem: add an MI300X-oriented HBM3 x64 interface - #8

Open
Basemism wants to merge 1 commit into
stagingfrom
staging-basem/hbm3
Open

mem: add an MI300X-oriented HBM3 x64 interface#8
Basemism wants to merge 1 commit into
stagingfrom
staging-basem/hbm3

Conversation

@Basemism

@Basemism Basemism commented Sep 2, 2026

Copy link
Copy Markdown

Add a 64-bit HBM3 channel model sized and clocked for the public MI300X memory organization. The interface is intended to be paired by HBMCtrl, giving 128 modeled x64 channels across eight stacks.

Use the public AMD MI300X capacity, interface-width, and transfer-rate data. For timings not published by AMD, derive conservative values from the Ramulator 2.1.0 HBM3_16Gb_8hi and HBM3_6400Mbps presets, scaling command timings to the slower MI300X command clock.

Unlike the generic HBM3 preset in gem5#3271 this interface is parameterized for MI300X’s existing HBMCtrl topology: eight controllers pair sixteen x64 interfaces per stack. Its capacity and burst timing reproduce 24 GB and approximately 665 GB/s per stack, or 192 GB and 5.3 TB/s across the package.

Add a 64-bit HBM3 channel model sized and clocked for the public MI300X
memory organization. The interface is intended to be paired by HBMCtrl,
giving 128 modeled x64 channels across eight stacks.

Use the public AMD MI300X capacity, interface-width, and transfer-rate data.
For timings not published by AMD, derive conservative values from the
Ramulator 2.1.0 HBM3_16Gb_8hi and HBM3_6400Mbps presets, scaling command
timings to the slower MI300X command clock. Document the model assumptions
where gem5 cannot directly represent HBM3 hierarchy or fractional command
cycles.

@mattsinc mattsinc left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In general this is also fine, except perhaps the two defaults that @TomXia should chime in about.

Comment thread src/mem/DRAMInterface.py
tXS = "353.076923ns"

# Existing gem5 HBM2 class uses close_adaptive for pseudo-channel HBM.
page_policy = "close_adaptive"

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@TomXia I know you did some testing about open/close/adaptive for your work. What did you end up using for the default?

Comment thread src/mem/DRAMInterface.py
page_policy = "close_adaptive"

# Existing gem5 HBM2 pseudo-channel class uses 64-entry queues.
read_buffer_size = 64

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@TomXia I know you also played with the read/write buffer sizes. What did you use as the defaults?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants