AMD Venice Uses 2.5D Packaging to Feed Higher Bandwidth and Split I/O Designs
Venice reaches 1.6 TB/s between chiplets and I/O dies, while separate SP7 and SP8 designs balance compute, memory and PCIe needs.

AMD’s Venice server processors put a packaging constraint at the center of the SoC design: moving more data between compute chiplets and the I/O dies makes conventional package connections increasingly expensive in power. AMD’s answer is 2.5D packaging, paired with two I/O dies and different I/O designs for the SP7 and SP8 platforms. The choices reflect a broader engineering tradeoff: bandwidth, core count, memory configuration, product cost and the physical space available for die-to-die links all have to fit together.
Mahesh Subramony, AMD corporate vice president of design engineering and head of server architecture, discussed the design decisions in a conversation with George Cozma at AMD In-Depth. His account shows that the packaging change shaped the layout of the SoC, while the target workloads shaped which chiplets and interfaces each platform received.
Why Venice moves to 2.5D packaging
The bandwidth between compute chiplets, or CCDs, and the I/O die has grown from just under 700 GB/s in the previous generation to 1.6 TB/s in Venice. AMD’s earlier substrate-based approach and GMI links suited the previous data rate. At Venice’s higher rate, Subramony said, conventional packaging could have consumed about 10–15% of the overall SoC power budget just to transport data. That estimate describes the design tradeoff AMD faced; it isn’t a general figure for every package or system.
Short-reach die-to-die connections in a 2.5D package can carry high bandwidth with lower power and latency. But placing the dies close together changes the physical design problem. The CCD and I/O die need compatible interface widths, and the links must remain short, on the order of a handful of millimeters. The dies also need enough edge length, or “beachfront,” for the connections.
That edge space has to serve two needs: connections between the CCDs and I/O dies, and connections that bring external I/O onto the package. A single I/O die has limited perimeter to accommodate both. Splitting the I/O across two dies gave AMD more room to arrange the required interfaces around the package. Subramony described that perimeter constraint as one of the fundamental reasons behind the split.
The change brings its own overhead, but the short, high-bandwidth links make the cost in power and latency manageable, according to Subramony. The result is a layout shaped as much by the physical geometry of the connections as by the functions inside each die.
One I/O design, used twice
For the Venice design discussed, AMD used the same I/O die design twice, with one copy rotated 180 degrees. That approach creates a mirrored arrangement without requiring two separately designed I/O dies for the package.
Some functions still need to exist only once in the system. Subramony cited the root of trust and some boot-related I/O as examples. Putting two identical I/O dies in the package duplicates those blocks, adding area that the SoC doesn’t need for functional reasons. AMD accepted that overhead as part of the split-die arrangement.
The second challenge was making the fabric behave correctly across the two dies. They can be viewed as separate domains, but in an NPS1 configuration the Venice SoC must operate as a single domain. AMD had to make the mirrored topology provide the required quality of service, bandwidth and latency when both I/O dies work together.
That topology question is likely to matter more as processors use larger tiled designs, Subramony said. Dividing a large SoC into tiles can help accommodate scaling, but the fabric has to make those tiles function together without compromising the system’s data paths.
SP7 and SP8 get different I/O dies
Both SP7 and SP8 use two I/O dies, but they don’t share the same I/O die design. AMD considered a single, superset die that could serve both platforms. Subramony said that option would have added cost to both products because their requirements differ enough that each would carry features it didn’t need.
SP7 is aimed at higher compute capacity and more memory channels, with fewer I/O lanes. SP8 puts more emphasis on I/O connectivity: it provides all 128 PCIe lanes per socket and has half as many memory channels as SP7. Its compute targets are also lower than SP7’s. Subramony gave a maximum of 96 cores for the Zen 6 side and 128 cores for the Zen 6c side.
Those configurations point to different priorities within the product family. SP7’s I/O die supports a design with more compute and memory bandwidth, while SP8’s supports a system that needs a large set of PCIe lanes. Separate dies let AMD tailor the silicon to those targets rather than paying the area and cost of a common design built to cover every case.
CCD links follow bandwidth needs
The number of links from a CCD to the I/O die also depends on what the CCD must feed. The 32-core design has four SDP ports, following a rough guideline of one port for every eight cores. AMD sized those links to give the CCD enough bandwidth for the cores to meet their performance and instructions-per-clock targets.
The 12-core CCD has three SDP ports, which may appear generous when viewed only as bandwidth per core. Its planned pairing with MRDIMMs changes the calculation. Each CCD needs to draw substantial bandwidth to help saturate 16 memory channels running at 12,800 MT/s. The link count is therefore meant to support the memory bandwidth available to the whole SoC, rather than scale simply with the CCD’s core count.
A spare core helps yield on the 32-core CCD
The 32-core CCD physically contains 33 cores, along with 33 slices of L3 cache. AMD keeps the number of active cores at 32; Subramony said there won’t be a 33-active-core product. The extra core is there to provide a way to harvest a defective core and its associated cache slice while keeping the mesh intact.
Large dies on advanced process nodes are more likely to encounter defects that affect yield. If one core is faulty, disabling it can allow the CCD to remain a functional 32-core part. AMD’s ability to remove any one core and its L3 slice while preserving the mesh gives it more flexibility when manufacturing these expensive, larger dies.
Keeping the active core count fixed at 32 also limits design complexity. Subramony said it helps keep cache associativity and mesh topology relatively straightforward. The spare core is thus a manufacturing provision, not a way to raise the advertised core count.
More L3 adds lookup time, but can reduce memory trips
The 12-core and 32-core CCDs differ in L3 capacity and mesh structure, and the larger cache carries a latency cost. Subramony said checking a 128 MB L3 takes a few extra clock cycles before a request can continue farther out. Techniques that speculate on the request may hide some of that time, but the larger cache still adds lookup latency. The path from the fabric to memory is otherwise almost identical, he said.
A DRAM latency measurement on an otherwise idle CPU can expose that additional lookup time, particularly with a closed-page load. But a single latency figure doesn’t capture how a working processor behaves. A larger L3 can serve more requests without going to memory, especially when more threads are resident on the CCD. Those cache hits can make up for the extra time spent checking the larger cache in workloads that benefit from its capacity.
That’s the rationale for increasing the CCD and its cache together. The tradeoff depends on whether a workload can use the additional capacity: the cache lookup can take longer, while a successful hit avoids the longer trip to memory.
Why more stacked cache isn’t the priority for dense CCDs
AMD’s 96-core offering pairs a 12-core CCD with X3D technology, adding another 96 MB of L3 and bringing the total to about 144 MB per CCD. Subramony framed the prospect of adding a large stacked cache to a denser CCD as a question of cost and workload benefit, rather than a straightforward extension of the existing design.
With fewer cores active on a dense CCD, those cores can already share the existing L3, increasing the cache available per active core. A further large X3D die would add expense, so the relevant question is how much performance it would return. In a many-core part operating under a tight per-core power budget, AMD might instead use that budget to increase core power and frequency.
Subramony described the 256-core part as designed for balanced throughput with a very tight power budget per core. For that product, adding more cache isn’t AMD’s first performance lever. More bandwidth or higher frequency would take priority, in his view; a large stacked cache isn’t a current focus.
Venice’s design choices follow a consistent pattern: spend package area and silicon where the target configuration can use them. The 2.5D links address the energy cost of rising chiplet bandwidth, separate I/O dies match SP7 and SP8’s different balance of compute, memory and PCIe, and the spare core gives the 32-core CCD room to absorb a defect. The remaining tradeoffs—especially L3 lookup time against hit rate—will depend on how each workload uses the processor.
Choosing Single, Double, Welded Flap and Repair Keder for Fabrication
Houston Engineering Team Builds a Welcoming Golf Tournament at East River 9
MIG Transfer Modes: Choosing Between Short Circuit, Spray and Pulse