Today at the TSMC 2026 North America Open Innovation Platform® (OIP) Ecosystem Forum, Alchip Technologies presented a technical paper with Synopsys on how AI data center ASICs can overcome increasingly stringent bandwidth requirements as scale-up and out interfaces become critical to next‑generation system architectures. This approach brings together Synopsys 224G PHY IP with Alchip’s advanced packaging and signal integrity/power integrity (SI/PI) expertise in a protocol-agnostic physical design.

Breaking Through the Bandwidth Wall

As AI accelerators scale to meet growing performance demands, moving data on and off the chip fast enough becomes one of the hardest problems in the design. At 224 Gbps SerDes speeds, the presentation addressed several key physical design challenges:

 

  • Die-edge congestion and escape routing complexity, along with signal and power integrity constraints across both die and package
  • Limited die-edge real estate and asymmetric constraints between East/West (E/W) and North/South (N/S) edges
  • Dense package substrate breakout requirements that must be met without degrading channel performance

 

The collaboration overcame these barriers through a coordinated optimization of escape routing, die-edge utilization, and package breakout — carefully balancing signal and power integrity requirements to maintain overall system performance. Differentiated handling of E/W and N/S die edges further ensured that maximum bandwidth per millimeter of die-edge beachfront was achieved at max reach, consistently across all die edges.

Building on this work, the analysis extended to E/W package breakout techniques needed to sustain 1.6 terabits per second (Tbps) of networking bandwidth, highlighting key routing density and SI/PI considerations at the substrate level. The resulting protocol-agnostic, forward-scalable physical design approach provides a foundation that can ultimately support 448G SerDes implementations.

Pushing Bandwidth Density to Its Physical Limits

Using this collaborative methodology, engineers achieved chip‑to‑chip interconnect bandwidth density approaching theoretical physical limits. This demonstrates the feasibility of ultra‑high‑density 224G scale‑up interconnect implementations within realistic die edge and package constraints.

These results point to broader implications for data center design, as AI accelerators and high‑performance networking ASICs continue to scale, maximizing bandwidth per millimeter directly impacts system throughput and scalability.