AMD server architecture leader Mahesh Subramony explains the design choices behind the Venice EPYC processors, especially the shift to 2.5D packaging. The move addresses a jump in chiplet-to-I/O bandwidth from just under 700 GB/s to 1.6 TB/s: conventional packaging would have consumed an estimated 10-15% of the SoC’s power budget simply moving data. Short, dense connections in the new package impose strict requirements on die dimensions and available connection edges, helping drive AMD to split the I/O system across two dies. Using the same I/O design twice, with one rotated, simplifies production but duplicates some one-time functions and requires careful fabric design to preserve bandwidth, latency, and service quality.
The interview also explains why SP7 and SP8 use different I/O dies: SP7 emphasizes compute and memory channels, while SP8 targets fewer cores, half as many memory channels, and 128 PCIe lanes per socket. CCD link counts likewise reflect workload needs: the 12-core version needs extra bandwidth to help saturate high-speed MRDIMM memory, while the 32-core design scales links with core count. Its 33rd physical core is a spare for defect tolerance and yield, with only 32 active. The larger CCD’s 128 MB L3 adds lookup latency, but its extra cache hits can improve effective performance. Subramony says stacked cache on dense CCDs is not a current priority; bandwidth or frequency offers a more attractive use of cost and power.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.