AMD Ryzen 9 7950X3D Review: AMD Retakes Gaming Crown with 3D V-Cache

Second-Gen 3D V-Cache Technology AMD Ryzen 9 7950X3D

AMD didn’t share any details about its second-gen 3D V-Cache in its Ryzen 9 7950X3D briefing materials, but we found some details at the 2023 International Solid-State Circuits Conference (ISSCC) in AMD’s latest presentation on the Zen 4 architecture, and AMD has also answered a few of our follow-up questions and shared important new details, including that the chiplet remains on the 7nm process and now has a peak bandwidth of up to 2.5 TB/s, whereas the first-gen 3D V-Cache peaked at 2 TB/s (among lots of other new info). We also have new pics and diagrams of the new 6nm I/O Die that AMD uses for its Ryzen 7000 processors.

AMD has moved to the second-gen of its 3D V-Cache, and Intel doesn't have a competing tech. Overall, AMD's second-gen 3D V-Cache technology is an impressive step forward over the first-gen because it allows the company to leverage the now-mature and less-expensive 7nm process node to boost the performance of its cutting-edge 5nm compute die. The new design represents AMD taking the key advantage of chiplet-based design methodologies — using an older and less-expensive process node in tandem with expensive new process tech — into the third dimension.

Image 1 of 8

*The above album contains images of the first-gen 3D V-Cache implementation for illustrative purposes.

First, a quick high-level refresher. AMD’s 3D V-Cache tech stacks an additional SRAM chiplet directly in the center of the compute die (CCD) to isolate it from the heat-generating cores on the sides of the chiplet. AMD's 3D chip stacking tech is based on TSMC's SoIC technology. TSMC's SoIC is a bump-less chip stacking tech, meaning it doesn't use microbumps or solder to connect the two dies. Instead, the two dies are milled to such a perfectly flat surface that the TSV channels can mate without any type of bonding material. We’ve covered the deep-dive details of the first-gen 3D V-Cache technology in our 5800X3D review, and you can read much more about the hybrid bonding and manufacturing process here.

Now for the new stuff: AMD’s previous-gen 3D V-Cache used a 7nm L3 SRAM chiplet stacked atop a 7nm Zen 3 CCD. AMD stuck with the 7nm process for the new L3 SRAM chiplet (called 'L3D') but now stacks it on top of a smaller 5nm Zen 4 CCD (see the table below). This creates a size mismatch, though, which required a few alterations.

First, AMD made the 7nm SRAM die smaller, so it now measures 36mm2 compared to the previous-gen's 41mm2. However, the total number of transistors remains the same at ~4.7 billion, so the new die is significantly denser than the first-gen chiplet.

As we saw with the first-gen SRAM chiplet, the 7nm L3 SRAM chiplet has incredible transistor density — we're looking at almost 3x the density of the first-gen 7nm compute chiplet, and surprisingly, the 7nm SRAM chiplet is significantly denser than the 5nm compute chiplet. That's because, as before, the chiplet uses a density-optimized version of 7nm that's specialized for SRAM. It also lacks the typical control circuitry found in the cache — that circuitry resides on the base die, which also helps reduce latency overhead. In contrast, the 5nm die includes several types of transistors along with data paths and other types of structures not present in the simplified L3 SRAM chiplet.

The stacked L3 SRAM chiplet is connected to the base die with two types of through-silicon vias (TSVs — a vertical electrical connection). The Power TSVs carry power between the chiplets, while the Signal TSVs carry data between the units.

In the first-gen design, both types of TSVs resided in the L3 region of the base chiplet. However, the L3 cache on the base die is now smaller due to the increased density of the 5nm process, and even though the 7nm L3 SRAM chiplet is smaller, it now overlaps the L2 cache (the prior gen only overlapped the L3 on the base die). As such, AMD had to alter the TSV connections in both the base die and the L3 SRAM chiplet.

Image 1 of 7

AMD had to extend the power TSVs from L3 into the L2 region due to the smaller size of the 5nm L3 cache on the base die (a result of the increased density and other factors). For the base die, AMD achieved a 0.68x effective area scaling across the L3 cache, data paths, and control logic compared to the old 7nm base chiplet, so there is physically less room for TSVs in the L3 cache.

The signal TSVs remain inside the L3 cache area on the base die, but AMD shrunk the TSV area in the L3 cache by 50% by applying learnings from the first-gen design along with DTCO improvements to reduce overhead circuitry in the new interface design.

AMD's 3D chip stacking tech is based on TSMC's SoIC technology. TSMC's SoIC is bump-less, meaning it doesn't use microbumps or solder to connect the two dies. You can read much more about the hybrid bonding and manufacturing process here. AMD tells us it used the same fundamental bonding process paired with continued process and DTCO improvements, but the minimum TSV pitch hasn't changed.

The L3 SRAM chiplet also remains on the same power domain as the CPU cores, so they can't be adjusted independently. This contributes to the lower frequency on the cache-equipped chiplet because the voltage can't exceed ~1.15V. You can see far more on that topic on the next page, where we test the chiplets individually. 

Second-Gen 3D V-Cache L3 Latency and Bandwidth Testing

We tested each eight-core CCD separately to measure latency for both the chiplet with the 3D V-Cache (labeled as CCD 0 X3D), and the standard chiplet with the normal amount of L3 cache (CCD 1).

As before, the extra latency from the additional L3 SRAM cache weighs in at 4 clocks, but the bandwidth between the L3 chiplet and the base die has increased to 2.5 TB/s, a 25% improvement over the previous 2 TB/s peak.

Image 1 of 3

The album above outlines our cache and memory latency benchmarks with the Ryzen 9 7950X3D CCDs using the MemoryLatencyTool and MemoryBandwidthTest from the Chips and Cheese team. These tests measure cache latency and bandwidth varying sizes of data chunks. 

The tool measures the L3 latency for the 3D V-Cache equipped chiplet (CCD 0) on the Ryzen 7 5950X3D at 10-12ns, whereas the chiplet without the extra cache (CCD1) measures at 9ns. The higher latency comes from the additional cache capacity infused in the chiplet but is fairly small given the performance uplift you’ll see in the gaming benchmarks.

6nm I/O Die (IOD) for Zen 4 Ryzen 7000 and EPYC Genoa

Image 1 of 4

AMD's ISSCC presentation also included plenty of new details about the 6nm I/O Dies (IOD) used in the Ryzen 7000 and EPYC Genoa processors. In the above album, you can see the zoomed-in images and an annotated die shot from chip detective @Locuza_.

We put the specs in the table for easy comparison, and as you can see, the EPYC Genoa I/O Die is simply massive compared to the Ryzen 7000 variant — that's because AMD can wire up to 12 compute chiplets (CCDs) to the I/O Die for its EPYC Genoa processors.

In contrast, the consumer chips are limited to two chiplets, an immutable limitation because, as you can see in the diagram, the Ryzen 7000 I/O Die only has two Global Memory Interconnect 2 (GMI2) links that connect the compute chiplets to the IOD. That's a bummer — the lower core-count Genoa models with four CCDs can have dual-GMI3 links (wide mode), a new capability that can offer advantages in some memory throughput-intensive tasks. That would've been interesting to add to the consumer chips. This limitation also rules out any three-CCD variants with this IOD.

Image 1 of 31

• MORE: Best CPU for gaming
• MORE: CPU Benchmark Hierarchy
• MORE: Intel vs AMD
• MORE: How to Overclock a CPU