AMD's Venice moves to 2.5D packaging to push 1.6 TB/s between chiplets and I/O

Talking Venice with Mahesh Subramony at AMD In-Depth

AMD's Venice moves to 2.5D packaging to push 1.6 TB/s between chiplets and I/O

AMD's Mahesh Subramony explains why EPYC Venice ditches conventional packaging: moving 1.6 TB/s over standard links would burn 10–15% of the SoC power budget. The fix—short-reach die-to-die links—forced the I/O die to be split in two, built from one mirrored design, with a 33rd spare core on the 32-core CCD for yield harvesting.

If we had tried to move that much data using conventional packaging, the energy spent just transporting those bits would have been somewhere around 10–15% of our overall SoC power budget, regardless of the picojoules-per-bit numbers you see claimed in papers.

More from this day

2026-11-20