Microsoft Details Maia Dataflow Architecture: 4719 Tflop/s fp8, Over 99% of Peak
thoefler · x · 2026-11-17
At SC26, Microsoft presented "Maia: A Software Defined Dataflow System for Large-scale AI Acceleration," revealing new details of its Maia accelerator architecture:
- Core principle: efficiency must come from both data movement and computation, via Software-Defined Dataflow. Microsoft proposes a new data-centric taxonomy (LSLA/SDLA/LSGA/SDGA) based on data access × data movement — Maia sits in SDLA, framed as a sibling to Flynn's 1972 taxonomy.
- Maia 300 specs: multi-chiplet N2P packaging, >13 Pflop/s, 8× HBM4 stacks delivering 18 TB/s at 950W TDP, two-tier Ethernet scale-up with 32× 400G NICs (3.2 TB/s bi-directional), 96 Maia per rack, 12k accelerators per scale-up domain, >166 Eflop/s cluster. Maia 200 is monolithic N3P with 6× HBM3E, 216 GB, 2.8 TB/s.
- Results: bf8 matrix-matrix multiply across 6,143 relevant matrix sizes hits 4719 Tflop/s — over 99% of theoretical fp8 peak.
More from Infra
- YC F26 Startup Maritime Launches a Managed Cloud for Running AI Agents — ycombinator · 2026-11-17
- Why Big Tech Fears Local AI: Security, Performance and Cost Advantages — DavidLinthicum · 2026-11-17
- Echelon optimizer: Muon-level results with up to 10.5× faster steps, Lean-proven — xennygrimmato_ · 2026-11-17
- Open-Weight Models Now Handle 30% of Enterprise Usage, Up From Near Zero Since July — Terminator857 · 2026-11-17
- What developers should check before deploying models on Cerebras — SpottyMoggy · 2026-11-17
- SkyPilot makes The Information's top 50 startups list at $345M valuation — skypilot_org · 2026-11-17