Intel Foundry Secures Major Microsoft Order for Maia 3 "Griffin" AI Accelerator
Intel has landed a significant contract with Microsoft, marking a pivotal moment for Intel Foundry as it gains the trust of a major external customer. According to industry sources, including Charlie Demerjian of SemiAccurate and X user @Jukanlosreve, Microsoft will manufacture its upcoming Maia 3 chip—codenamed "Griffin"—at Intel Foundry. The production will utilize Intel’s advanced 18A or 18A-P process node, underscoring Intel’s growing role in the competitive semiconductor manufacturing landscape.
Advanced 18A-P Node: Powering Next-Generation AI Chips
The 18A-P process node represents a significant evolution over Intel’s standard 18A technology. It incorporates RibbonFET and PowerVia innovations, which collectively enhance both performance and energy efficiency. Key improvements include newly engineered low-threshold voltage components, optimized elements to minimize leakage, and refined ribbon width specifications. These advancements are designed to maximize performance-per-watt, a critical metric for AI accelerators deployed in data centers where efficiency and scalability are paramount.
Microsoft’s Strategic Shift in Chip Manufacturing
Microsoft’s decision to partner with Intel Foundry for the Maia 3 "Griffin" chip signals a potential shift in its manufacturing strategy. If the Maia 3 project delivers high yields and rapid turnaround, Microsoft may continue to leverage Intel’s most advanced process nodes for future Maia accelerators. This move could see future generations of Microsoft’s AI hardware produced at Intel Foundry, further diversifying Microsoft’s supply chain beyond its previous reliance on TSMC.
Technical Overview: Maia 100 and the Path Forward
Microsoft’s first-generation Maia 100 chips were manufactured using TSMC’s N5 process and CoWoS-S interposer technology. These chips feature an 820 mm² die, a 500 W TDP (with a maximum design power of 700 W), and are equipped with 64 GB of HBM2E memory delivering 1.8 TB/s bandwidth. The architecture also includes 500 MB of L1/L2 cache. In terms of performance, Maia 100 achieves up to 3 PetaOPS at 6-bit precision, 1.5 PetaOPS at 9-bit, and 0.8 PetaFLOPS for BF16 operations. Connectivity is robust, with 600 GB/s backend network bandwidth via twelve 400GbE ports and 32 GB/s host bandwidth through PCIe Gen 5 x8.
Future Prospects: Advanced Intel Nodes and Packaging
Should the Maia 3 "Griffin" project prove successful, Microsoft is expected to explore even more advanced Intel process nodes, such as 18A-PT and 14A. The 18A-PT node is specifically tailored for AI and high-performance computing (HPC) applications that demand sophisticated multi-die architectures. This process introduces a redesigned back-end metallization layer, through-silicon via (TSV) capabilities for vertical interconnects, and support for advanced hybrid bonding interfaces with competitive pitch dimensions. These features enable highly scalable chiplet integration, paving the way for Microsoft to adopt advanced packaging technologies in future AI accelerator designs.
Intel’s collaboration with Microsoft on the Maia 3 "Griffin" chip highlights the growing importance of advanced semiconductor manufacturing processes in the AI and data center sectors. As both companies push the boundaries of performance and efficiency, this partnership could set new standards for the industry.
