Meta and AMD Co-Design MI350X for Recommendation Model Training

Architecting for Hardware Diversity: Meta's End-to-End Co-Design for High-Performance AI Training on AMD GPUs
Haoyu Zhang, Arshad Rahman, Adnan Aziz et al. (Meta and AMD), 2026

arXiv 2610.22230

Why recommenders are a harder target than LLMs

Click or hover a stage of a DLRM training step.

Select a stage.

The paper's three ideas

Hover a card.

Select an idea.

MI350X vs B200 on paper

Pick a spec. Bars start at zero.

Co-design = adapting the recipe to 288 GB

Toggle activation checkpointing. Illustrative segments Capacity lines (180 / 288 GB) are reported.

Recipe-change results (not like-for-like)

Reported, each a single model without variance.

Kernel and communication toolbox

Is 101.2% a fair parity?

Slide the assumed power-boost contribution (paper reports 3 to 6% end to end). Computed as 101.2 ÷ boost, assuming uniform effect, which the paper does not state.

Spread across ten models

Counts are reported; per-model values are not shown here.