AMD explains how the brand new 3D V-Cache improves over the unique

In context: AMD launched the Ryzen 9 7950X3D late final month and welcomed the enthusiastic response to its second-generation 3D V-Cache, though opinions fluctuate on its usefulness in 16-core CPUs. Now they’ve shared some technical particulars explaining its efficiency.

AMD began mixing nodes in 2019 when it used the 7nm node for the core advanced die (CCD) and the 12nm node for the Zen 2 microarchitecture’s IO die. AMD lately confirmed to Tom’s Hardware that Zen 4 brings it as much as three nodes: a 5nm node for the CCD, a 6nm node for the IO chip, and a 7nm node for the V-Cache.

The V-Cache is in the course of the CCD, and the eight cores are on the perimeters.

In a current ISSCC presentation, AMD defined among the challenges of stacking one node onto one other. Both the 7950X3D and the unique 5800X3D put their V-Caches on prime of their common L3 caches to permit them to attach. This association additionally retains the V-Cache away from the warmth generated by the core. However, whereas the V-Cache sits proper above the L3 cache within the 5800X3D, it overlaps the L2 cache on the fringe of the core within the 7950X3D.

Part of the issue is that AMD elevated the quantity of L2 cache in every core from 0.5 MB in Zen 3 to 1 MB in Zen 4. But it solves the additional house limitation by punching holes within the L2 cache for through-silicon vias (TSV) to energy the V-Cache. The sign TSV nonetheless comes from the controller on the middle of the CCD, however AMD has additionally tweaked it, decreasing its footprint by 50%.

The Zen 4 L2 cache is bigger due to its bigger capability, but additionally as a result of it has TSVs operating by means of it.

AMD shrinks V-Cache from 41 mm² to 36 mm² But maintain the identical 4.7 B transistors. TSMC manufactures caches on a brand new model of the 7nm node developed particularly for SRAM. Thus, the V-Cache has 32% extra transistors per sq. millimeter than the CCD, regardless of being fabricated on the smaller 5nm node.

All of the enhancements and workarounds applied by AMD add as much as a 25% improve in bandwidth to 2.5 TB/s and improved effectivity, however aren’t specified. The nine-month interval between the first-generation and second-generation supplementary chiplets is not unhealthy. Hopefully it is going to present its price when the Ryzen 7 7800X3D arrives in a month.