Samsung Electro Mechanics expands into data center AI market by producing high performance FC BGA substrates for Qualcomm AI200 accelerators
Samsung Electro Mechanics has begun production of high performance semiconductor package substrates for Qualcomm's first data center AI accelerator. As reported by ZDNet Korea, the company is producing the Flip Chip Ball Grid Array package substrates in the company's main facilities in Busan. This move marked a major broadening of the two companies' collaboration from their traditional emphasis on mobile and PC components and towards the enterprise data center market.
Target Hardware The target hardware is the general purpose accelerator the Qualcomm AI200. Its an award winning accelerator for the artificial intelligence inference market. It is integrated on silicon with a custom Oryon CPU, a Hexagon NPU and low power LPDDR5. The buy volume as a first product is modest, but the short term importance of venturing into server infrastructure as a substrate manufacturer is very significant, given that the company has previously been a traditional application processors provider at the heart of consumer mobile devices.
An industry source noted the importance of the supply agreement as follows
"The established relationship between Samsung Electro Mechanics and Qualcomm allowed them to secure this supply chain agreement smoothly. With Qualcomm intending to release the AI200 followed by the subsequent AI250 chip, the substrate manufacturer stands to secure a highly diversified stream of revenue."
The AI200's FC BGA substrates connect the semiconductor die to the primary circuit board via tiny conductive bumps instead of wire bonding. This yields huge performance benefits for its electrical and thermal performance, and is now used as the baseline architecture for all high end processors. The substrates for the AI200 have a low to mid 10s layer count, with one copper layer every other layer as insulation (Ajinomoto B.U.F.) but high end inference silicon does not require such high complexity and has a 20+ layer count.
This factor of differentiation has enabled the other competitors to follow along this supply chain. At present, LG Innotek has started making a qualification process to qualify its own FC BGA substrates for the Qualcomm hardware, in order to adopt mass production in the near future. An industry analyst elaborated
"The AI200 focuses primarily on inference, which requires less bandwidth and lower substrate performance compared to accelerators built on high bandwidth memory. This characteristic lowers the technical entry barrier, allowing newer players like LG Innotek to participate in the supply chain."
