
Amd.com
Every 32-core compute chiplet in AMD's EPYC "Venice" server processor physically contains 33 cores, and AMD will never enable the spare, Mahesh Subramony, the company's corporate vice president of design engineering, said in an interview Chips and Cheese published Nov. 16. The spare core, paired with an extra slice of L3 cache, lets AMD ship a chiplet that has one defective core as a full 32-core part, a yield safeguard on large, expensive dies built on TSMC's 2nm-class process.
The details arrive six days after AMD said EPYC 9006 SP7 platforms are shipping from OEM and ODM partners. They give buyers AMD's own account of three trade-offs that independent testers are now probing: links AMD puts near 1.6 TB/s between chiplets and the I/O die, an I/O die built as two mirrored halves that must behave as one NUMA domain, and a larger L3 cache that adds a few clocks to every lookup.
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 keeping the on-die mesh intact, according to the interview. Subramony framed it as a cost and yield decision: "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 said no product will ship with 33 active. That implies 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 (0.26 square inches) of 2nm-class silicon, and dies that large are more likely to catch a defect, the logic Subramony described.
Disabling spare or faulty units is an established yield tactic: Sony shipped PlayStation 3 Cell processors with seven working synergistic processing elements to raise yields. Venice's spare is designed in from the start, and the interview adds AMD's stated rationale. It gives no defect-rate or cost figures.
Interviewer George Cozma described Venice as AMD's first server CPU with 2.5D packaging, and Subramony did not dispute it. 2.5D packaging places dies side by side on an interposer made of silicon, glass or organic material instead of stacking them. Chips and Cheese had inferred at CES that Venice appears to use more advanced packaging than the organic-substrate routing AMD has used since EPYC Rome.
Subramony said the previous generation moved a little under 700 GB/s between chiplets and the I/O die, versus about 1.6 TB/s on Venice, with each chiplet overprovisioned to keep the link from limiting it. That is more than double. He estimated roughly 10% to 15% of the SoC power budget would have gone to moving those bits over conventional packaging, and said the older GMI links made sense for the previous product.
The new links impose geometry. Dies must match widths at the interface, and each needs enough perimeter for every chiplet connection, a requirement Subramony tied to AMD's decision to split the I/O die. The interview names no specific packaging technology, so whether Venice uses an interposer, a bridge or another 2.5D method remains unconfirmed.
Read more: EPYC Venice Arrives Wednesday: AMD's Zen 6 on TSMC 2nm Resets Server Race
No. The two figures are separate measures that land close together. AMD lists up to 16 channels of 12,800 MT/s MRDIMM memory with a theoretical maximum of 1,638.4 GB/s per socket, while Subramony's figure describes data moving between chiplets and the I/O die, and the interview does not say whether it is an aggregate per socket.
The chiplet link is sized to keep memory fed. Subramony said each chiplet must pull enough bandwidth to help saturate those 16 channels, so the 12-core chiplet gets three links to the I/O die and the 32-core chiplet gets four SDP ports, roughly one per eight cores.
AMD built one I/O die design used twice, rotating one copy 180 degrees, Subramony said. One-off blocks such as the root of trust and boot I/O are duplicated, which adds some area.
The harder problem is visible to software. In NPS1 mode the whole socket must act as a single domain, so AMD had to deliver consistent quality of service, bandwidth and latency across the mirrored halves. In NUMA designs, memory access time normally depends on where the memory sits relative to the processor. Subramony expects tiled designs to become more advantageous as core counts rise.
AMD also built two distinct I/O die designs, one per platform, rather than one superset die, to avoid adding cost to both. SP7 favors compute and memory bandwidth, with more memory channels and fewer I/O lanes. SP8 has half as many memory channels but all 128 PCIe lanes per socket, and targets up to 96 Zen 6 cores or up to 128 Zen 6c cores. The interview gives no SP7 lane count; AMD's Nov. 10 blog lists 128 PCIe Gen 6 lanes in a single socket.
The 32-core chiplet carries 128 MB of L3. Subramony said a larger L3 takes longer to decide whether a request hits or misses, adding a few extra clocks before the request heads to memory, but the path from fabric to memory is otherwise nearly identical and extra hits offset the cost for workloads that benefit. Larger caches hit more often but respond more slowly, a general trade-off in CPU design.
With 3D V-Cache, the 12-core chiplet adds 96 MB for about 144 MB per chiplet. Eight of them account for the 1,152 MB of L3 listed for the 3D V-Cache Venice-X variant. Asked about a 32-core chiplet with stacked cache reaching about 384 MB, a figure the interviewer supplied, Subramony said stacking more cache on the dense design is expensive.
Independent tests are starting to show how these trade-offs land. TechTimes reported that Tom's Hardware found the 256-core EPYC 9996 slower than the 192-core EPYC 9965 in a ClickHouse database test, with cache coherency suggested as a plausible but unmeasured cause. The interview does not address that test, so it establishes no cause. Phoronix's benchmarks, by contrast, put a dual-socket 9996 at 1.58 times a dual-socket 9965 on geometric mean across its suite. Core count alone does not predict speed.
Subramony speaks for AMD, so his explanations are the company's account, and independent measurements of these trade-offs are only beginning. The interview itself gives no packaging vendor, process node, clock or power figures for specific parts, launch dates or pricing.
Some of that is public elsewhere. AMD said May 21 that Venice entered production ramp on TSMC's 2nm process in Taiwan, and announced the 6th Gen EPYC family at Advancing AI on July 23. AMD's Nov. 10 blog lists SP8 as coming later without a date. Taiwan's TechNews reported July 24 that SP8, with 8 to 128 cores, is expected in the first half of 2027, a timing AMD has not publicly confirmed with a date.
AMD's own performance comparisons remain vendor figures. Its 2.24 times claim against Nvidia's Vera compares a 256-core chip with an 88-core one.
Venice trades silicon area and design complexity for yield, chiplet bandwidth and power: a spare core on every 32-core chiplet, duplicated one-off I/O blocks and two I/O die designs. Buyers weighing SP7 systems should test workloads that share data across many threads, where cache coherency costs, the suspected factor in the ClickHouse result, are most likely to show. The next checkpoints are SP8, reported for the first half of 2027, and Venice-X, expected in the second half.
Only the 32-core chiplet, according to Subramony. It physically has 33 cores and 33 L3 slices, but AMD always keeps 32 active and plans no 33-core product. The interview does not say whether the 12-core chiplet carries a spare.
Subramony said it raised chiplet-to-I/O bandwidth to about 1.6 TB/s from under 700 GB/s, and that moving that much data over conventional packaging would have cost roughly 10% to 15% of the SoC power budget. He did not name the packaging technology.
The interview does not say. Subramony described a few extra clocks per lookup that extra hits usually offset, while TechTimes' earlier coverage called cache coherency a plausible but unmeasured explanation for the database result.
AMD has not given a date. Its Nov. 10 blog says SP8 follows SP7, and TechNews reported July 24 that SP8 is expected in the first half of 2027.
