Meta unveils MTIA 300, its first training chip with built-in NIC chiplets
Meta_Engineers · x · 2026-08-31
Meta's engineering blog details MTIA 300, the first of its in-house accelerator family optimized for training ranking and recommendation models.
- Unlike LLMs, recommendation models are communication-bound: embedding tables can hold over 99% of parameters, driving frequent AllReduce/AllToAll/AllGather collectives across hundreds of accelerators.
- On GPUs, these collectives compete with compute for resources, leaving hardware underutilized.
- MTIA 300's built-in NIC chiplets and communication-offloading engines make communication a first-class citizen, beating general-purpose GPUs on these workloads.
- Meta co-designed the HCCL communication library alongside the chip.
Related event: Meta Unveils MTIA 300, Its First Training-Focused AI Accelerator(2 posts)→
More from Infra
- RL train-sample systems are frontier AI's hardest infra problem — and a trade secret — brianryhuang · 2026-10-16
- Local tuning of GLM-5.3-Flash on 2×GB10 lifts decode from 49 to 58 tok/s — EAccelerate_42 · 2026-10-16
- 3x DGX Spark cluster runs GLM-5.3-Flash locally behind a proxy, 1.75M tokens/day at 119W — Teknium · 2026-10-16
- Optuna creator Akiba: when 2 hyperparameters cost $20M, what's left for BO/HPO — Stefania_druga · 2026-10-16
- Amazon Proposes Per-Request Validated Evidence Caching for Production RAG — _reachsumit · 2026-10-16
- LinkedIn's TransX report: sequence transduction for recsys, +6.02% CTR in A/B tests — _reachsumit · 2026-10-16