Meta MTIA 400 (Custom AI Accelerator · Recommendation + GenAI Dual Mandate)
Product Overview
MTIA 400 is the fourth generation of Meta's custom MTIA AI accelerator family, systematically disclosed for the first time at Hot Chips 2026 (2026-08-23~25).
The biggest change in this generation is the mandate expansion: from its inception, MTIA served only the single workload of Recommendation & Ranking; the MTIA 400 also takes on generative AI (GenAI), becoming a "recommendation + GenAI dual-mandate" chip. Meta's production figure given in the talk: hundreds of thousands of MTIA chips already in production.
Architecturally, the MTIA 400 is fully chipletized — 2 compute dies + 1 SoC die + 2 network dies, plus HBM. This is the first time Meta has split its custom accelerator into a multi-die combination, a generational watershed compared with the earlier single-die MTIA 200.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | MTIA 400 (2 compute dies + 1 SoC die + 2 network dies + HBM, fully chipletized) |
| PE Array | 8 × 6 processing element array (with redundant rows) |
| Process | Not disclosed |
| FP4 Compute | 12 PFLOPS |
| FP16/BF16 Compute | Not disclosed (15x the MTIA 200) |
| Memory Capacity | Not disclosed |
| Memory Bandwidth | Not disclosed (DRAM bandwidth 46x the MTIA 200) |
| On-chip SRAM Bandwidth | Not disclosed (5x the MTIA 200) |
| TDP | Not disclosed |
| Scale-up Domain | 72 MTIA 400s per domain |
| Design Partner | Broadcom (multi-generation MTIA collaboration) |
| Announced | 2026 (disclosed at Hot Chips 2026) |
⚠️ Caveat: the "15x / 46x / 5x" figures in the table are multiples relative to the MTIA 200, not absolute specifications. Meta consistently does not publish full MTIA parameters; do not mix these relative multiples with absolute compute figures.
Generational Comparison with the MTIA 200
| Metric | MTIA 200 | MTIA 400 | Change |
|---|---|---|---|
| Mandate | Recommendation / ranking | Recommendation + GenAI | Expanded |
| Package | Single die | 2 compute + 1 SoC + 2 network dies | Fully chipletized |
| FP4 Compute | Not disclosed | 12 PFLOPS | Newly disclosed |
| FP16 Compute | Baseline | — | 15x |
| DRAM Bandwidth | Baseline | — | 46x |
| SRAM Bandwidth | Baseline | — | 5x |
| Scale-up Domain | — | 72 chips | — |
Key insight: DRAM bandwidth up 46x and FP16 compute up 15x — the bandwidth multiple far exceeds the compute multiple. This ties directly to the MTIA 400's new GenAI mandate: the decode stage of generative inference is a memory bandwidth bottleneck, and Meta has clearly weighted its resources toward data movement.
MTIA Roadmap (300 / 400 / 450 / 500)
Meta's published roadmap shows four generations — MTIA 300 / 400 / 450 / 500 — iterating over the next two years:
| Generation | Positioning | Key Changes |
|---|---|---|
| MTIA 300 | Recommendation / GenAI / inference | Roadmap starting point |
| MTIA 400 | Recommendation + GenAI dual mandate | 12 PFLOPS FP4, fully chipletized |
| MTIA 450 | GenAI / inference | HBM bandwidth doubled vs the MTIA 400 |
| MTIA 500 | GenAI / inference | HBM bandwidth up another 50%; per-chip power up to 1700 W |
Accompanying cooling roadmap: Meta targets rack power density of 80 kW or even above 120 kW, using an air-assisted liquid cooling (AALC) + Sidecar CDU architecture. Meta's core aim is not maximum cooling efficiency, but bringing liquid cooling capability to existing air-cooled data centers as quickly as possible — Sidecars can be deployed per rack, scale quickly, and isolate faults, making them better suited to rapidly launching inference services.
Vendor Information
| Item | Details |
|---|---|
| Company | Meta Platforms, Inc. |
| Headquarters | Menlo Park, California, USA |
| Custom Silicon Roadmap | MTIA (recommendation → GenAI → inference) |
| Design Partner | Broadcom |
| Availability | Internal use only (not sold externally) |
| Foundry | TSMC (specific node not disclosed) |
Use Cases
- ✅ Recommendation systems / ranking models (MTIA's traditional home turf)
- ✅ Generative AI inference (new MTIA 400 mandate)
- ✅ Mixed deployment with NVIDIA / AMD GPUs (Meta pursues a multi-chip strategy)
- ❌ External sales / third-party procurement
- ❌ Frontier model pretraining (Meta still relies on NVIDIA GPUs)
Related Cards
- Meta MTIA v3 (Iris) — Previous-generation custom accelerator
- Microsoft Maia 200 — Cloud provider custom inference chip
- Google TPU 8i — Inference-oriented custom ASIC
- AWS Trainium3 — Cloud custom training/inference chip
- Full comparison table
References
- Hot Chips 2026: Meta's Custom AI Silicon — From Recommendation to Dual-Mandate with GenAI
- Industry research report: Custom ASICs enter large-scale deployment at Hot Chips 2026
- Lingke 1+1: Meta MTIA roadmap and AALC liquid cooling architecture
- Reuters: Meta's new AI chip Iris slated for production in September 2026