Intel Unveils EMIB-T Development Plans for AI Chips

Intel plans to expand the use of EMIB-T chip packaging technology for large artificial intelligence accelerators. At the ISMP 2026 conference in Busan on November 4, Intel Foundry's packaging director Kyu O Lee presented guidelines suggesting the placement of more than 32 high-speed HBM memory blocks and over 70 silicon bridges in a single module. The company views EMIB-T as an alternative to Taiwan's TSMC CoWoS-L technology.

EMIB is based on small silicon bridges embedded in an organic substrate to connect individual dies. The EMIB-T version adds vertical TSV electrical connections, allowing power to be supplied through the bridges. This approach reduces the need for large silicon interposers and potentially lowers material consumption.

According to Intel, the efficiency of silicon wafer area usage in the production of EMIB bridges exceeds 90%, whereas for designs with large interposers, the figure is about 60%. This is an assessment of material consumption, not the share of finished products that have passed production control.

In developments for 2026–2027, Intel is focusing on modules measuring at least 120 × 180 mm, containing 16 HBM blocks and 38 bridges. A more distant project envisions a size of 240 × 240 mm, over 32 memory blocks, and 70 bridges. These parameters indicate plans rather than already released serial products.

Production Constraints and Competition with TSMC

Interest in alternative technologies is related to the high capacity utilization of TSMC's CoWoS packaging. However, this company is also expanding production: CoWoS-L is expected to remain a common solution for AI chips at least until 2028.

Industry estimates suggest that the yield of usable products using EMIB is around 90%, while CoWoS has reported values of about 98%. These figures pertain to different technological processes and do not provide a basis for directly comparing the reliability of EMIB-T and CoWoS-L. An additional constraint for Intel remains the quality and availability of organic substrates.

Google, Amazon, and MediaTek are considered among potential users of such solutions, but final production agreements for EMIB-T have not been publicly confirmed. Intel introduced the technology in 2025 and aims to expand its application, but actual competitiveness will depend on cost, yield, and available capacity.