
AMD EPYC Venice: Spare 33rd Core in Each 32-Core Chiplet Protects Yield, Executive Says - Tech Times
AMD's EPYC 'Venice' server processor contains 33 physical cores in each 32-core chiplet, with the spare core never enabled, according to an AMD executive.
Every 32-core compute chiplet in AMD's EPYC 'Venice' server processor physically contains 33 cores, and AMD will never enable the spare core, according to Mahesh Subramony, the company's corporate vice president of design engineering. The spare core, paired with an extra slice of L3 cache, allows AMD to ship a chiplet with one defective core as a full 32-core part, serving as a yield safeguard for large, expensive dies built on TSMC's 2nm-class process.
These details emerged six days after AMD announced that EPYC 9006 SP7 platforms are shipping from OEM and ODM partners. They reveal three key trade-offs that independent testers are now examining: the 1.6 TB/s bandwidth between chiplets and the I/O die, the mirrored I/O die design that must function as a single NUMA domain, and the larger L3 cache that adds a few clocks to each lookup.
Spare Core Design
AMD added both a 33rd core and L3 slice to the 32-core chiplet, with the ability to disable any one core and its cache slice while maintaining the on-die mesh intact. Subramony explained this as a cost and yield decision: 'As CCDs get larger, you're more likely to encounter defects that affect yield.'
As CCDs get larger, you're more likely to encounter defects that affect yield.
AMD keeps exactly 32 cores active on these chiplets, and Subramony confirmed no product will ship with 33 active cores. This means even a defect-free chiplet runs with one core switched off. Chips and Cheese estimated from CES die photos that each 32-core chiplet measures about 165 mm², and dies of that size are more likely to contain defects.
Bandwidth Boost
Venice represents AMD's first server CPU with 2.5D packaging, which places dies side by side on an interposer rather than stacking them. This technology has increased chiplet-to-I/O bandwidth from under 700 GB/s in the previous generation to about 1.6 TB/s on Venice—more than double the previous rate.
Subramony estimated that roughly 10% to 15% of the SoC power budget would have been consumed by moving data over conventional packaging. The new links require dies to match widths at the interface, and each needs sufficient perimeter for every chiplet connection, which influenced AMD's decision to split the I/O die.
Mirrored I/O Design
AMD built one I/O die design used twice, rotating one copy 180 degrees. While unique blocks like the root of trust and boot I/O are duplicated, the greater challenge was ensuring the whole socket acts as a single domain in NPS1 mode, requiring consistent quality of service, bandwidth, and latency across the mirrored halves.
AMD also created two distinct I/O die designs for different platforms. SP7 prioritizes compute and memory bandwidth with more memory channels, while SP8 has half as many memory channels but all 128 PCIe lanes per socket, targeting up to 96 Zen 6 cores or 128 Zen 6c cores.
Cache Trade-offs
The 32-core chiplet carries 128 MB of L3 cache. Subramony acknowledged that a larger L3 takes longer to determine whether a request hits or misses, adding a few extra clocks before the request proceeds to memory. However, he argued that extra hits offset this cost for workloads that benefit, as the path from fabric to memory remains nearly identical.
With 3D V-Cache, the 12-core chiplet adds 96 MB for approximately 144 MB per chiplet. Eight of these account for the 1,152 MB of L3 in the 3D V-Cache Venice-X variant. Subramony noted that stacking more cache on the dense design would be expensive.