Intel has secured a significant order from Microsoft, putting trust from external customers into Intel Foundry. According to Charlie Demerjian from SemiAccurate—and shared by X user @Jukanlosreve —Microsoft will manufacture its Maia 3 chip codenamed "Griffin" at Intel Foundry using the 18A or 18A-P process. This 18A-P node improves upon Intel's 18A by incorporating the RibbonFET and PowerVia technologies, offering enhanced performance and energy efficiency. Compare to the regular 18A node, improvements include newly designed low-threshold voltage components, optimized elements to reduce leakage, and refined ribbon width specifications, all aimed at boosting performance-per-watt metrics. This is particularly important for AI accelerators used in data centers, which require high efficiency due to the large number of units involved.

Microsoft could relocate the production of future Maia accelerators, following Maia 3, to Intel Foundry. If the Maia 3 project is successful, they plan to continue with Intel's more advanced nodes. As a reminder, Microsoft's first-generation Maia 100 chips were produced using TSMC's N5 process with CoWoS-S interposer technology. These chips feature an 820 mm² die size and operate at a 500 W TDP, with a maximum design power of 700 W. They include 64 GB of HBM2E memory providing 1.8 TB/s bandwidth and 500 MB of L1/L2 cache. The chips achieve peak tensor performance of 3 PetaOPS at 6-bit precision, 1.5 PetaOPS at 9-bit, and 0.8 PetaFLOPS for BF16. Connectivity options include 600 GB/s backend network bandwidth via twelve 400GbE ports and 32 GB/s host bandwidth through PCIe Gen 5 x8.

Update October 18, 17:30 UTC : Charlie Demerjian confirmed that the name of the chip is actually a third generation Maia 3 accelerator codenamed " Griffin ."