2026-d1-1200-sponsor-talk-meta

Unfortunately I can't share details here because this is part of currently active discussions.

It was roce

Yes, these are RDMA / RoCE NICs.

Has Meta any plan to use UFH rather than ESUN 1.0 header?

MTIA is Meta's custom line of accelerators that are designed specifically for for Ranking & Recommendation (R&R) and GenAI workloads. Resulting in similar or higher performance and cost efficiency. See ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions

As mentioned in the slides, we could potentially use all the NICs for scale-up connectivity depending on the workload requirements.

Thanks for the response, do you have any insights on workload requirements in terms of scale-up/scale-out ratio?

Generally speaking training requires collectives over accelerators in multiple racks and inference jobs tend to be smaller in scale resulting in different requirements.

I'm unable to share this information at this time.

You may find your response in MITA paper accepted for super computing 2026. Arxiv version is available

Thank you Leila for the response

I'd also recommend arxiv.org/html/…. As for the question - typically its a decision made up front on the host side. However, dynamic decision in device is also possible and something we are exploring.