3D cache was the last great evolution who has experience with this type of memory, which, as many of our readers know, plays a very important role in both processors and graphics cards, as it largely depends on the peak of maximum performance that both components are able to achieve.

Currently, the cache is divided into three large levels, L1 cache, L2 cache and L3 cache. In this article, we have already explained to you what these three types of cache are and what their differences are, so I invite you to read it if you have not had the opportunity to read it at the time, because it will give you a very important basis to better understand what how it works and why 3D cache is so important.

So far, the only company that dared to use 3D cache was AMD and its implementation took place both in general consumer processors and models for the professional sector. This is important because the fact that Sunnyvale released an EPYC CPU with 3D stacked cache means that this technology is very important to them and who really wants to continue to use it and improve it.

What is 3D cache and how is it implemented?

We are looking at a specific type of cache that works in the same way as an L3 cache, which means we are looking at third level memory block. This means that both in terms of proximity to the CPU and in terms of speed, it is inferior to the L2 cache, but has the advantage of being much larger, which means it offers more storage capacity.

Traditional cache is distributed in a horizontal plane, that is, it has a 2D presence. With the 3D cache, things will change completely from then on stacks vertically on another chip and later both will be encapsulated. In the case of AMD, which, as I said, is the first to use 3D cache in its processors, this type of memory is stacked in the central area of ​​the CPU chiplet, where the L3 cache integrated into the processor is located. present, present.

The stacking process requires the use of a TSV (via silicon) interconnect system which allows the upper cache to communicate with the lower cache. Stacking this cache generates an imbalance between this section and the sections where the CPU cores go, since the former happens to be taller.

The packing process cannot be done directly because of this unevenness. For this, it is necessary that all elements of the chiplet remain at the same height AMD resorted to two layers of construction silicon which are the ones that allow homogeneous encapsulation, as we can see in the picture (they are the two transparent blocks). At first glance, this may seem like a simple process, but it is actually quite a complex thing that actually continues to present significant challenges, which we will see later.

How does 3D stacked cache work?

We’ve already seen how this type of cache is implemented, so let’s dive into how it works. We are dealing with L3 or third level cache type which in this case means capacity takes precedence over speed and latencythat is, it is a larger type of memory, but slower and with longer access times.

However, we need to put this information in the right context and it is that although they are considered to be slower than L1 and L2 caches still much faster than RAMthat’s why it has a significant effect on the performance of the CPU when working with certain tasks, because the CPU always tries to call it before looking in the RAM, and if it finds what it’s looking for, it completes its work cycles faster.

The L3 cache is also a type of shared memory that since the Zen 3 architecture is accessible to all chiplet cores. This means that in a processor like the Ryzen 7 7800X3D, which has 32 MB of 2D L3 cache and 64 MB of 3D L3 cache, its eight cores can access a total of 96 MB of shared L3 because it is integrated into a single chiplet. .

It is important to be clear that the concept of shared L3 cache has an important limitation, namely that it is applied to a chiplet. So in a processor like the Ryzen 9 7950X3D only the first 8-core chiplet has access to 32 MB 2D L3 and 64 MB 3D stacked L3. The second 8-core chiplet can only access its 32MB of shared L3.

The L3 cache has a task that is as simple as it is important: to store data, instructions, and already resolved operations that the CPU can access when needed. Each time the processor accesses the L3 cache can be solved by mistake or hit. If it hits, which is ideal, it’s found what it needs in that cache. In case it fails, you will have to search in RAM, and if it is not found there, you will have to perform certain operations again.

Larger amounts of L3 allows you to store more data and instructionsand that means the chances of the processor getting hit while looking for what it needs in that cache increases. Since it communicates with it much faster than RAM, the performance improvement we can achieve in certain tasks is significant.

In other words, more simply, that L3 cache It’s like having a cheat sheet with a specific and perfectly detailed list of questions and answers., while RAM would be like an encyclopedia where you would have to spend more time searching. A larger L3 cache would equal a larger list of questions and answers, allowing you to reduce your reliance on this encyclopedia.

It is important to pay attention not all applications benefit equally from having more L3 cache because they simply don’t work the same. Currently, games are the ones that get the most out of this type of cache because they have a high dependency on the CPU, and this can improve its performance thanks to this increase in traffic when using the L3 cache, which is closer to its cores and communicates with them faster than RAM .

What advantages and disadvantages does this type of cache offer?

3D stacked L3 cache has very important advantages go beyond just improving performance in certain applications, although this is undoubtedly one of the most important. To give you a clearer and more visually explicit vision, I will share with you a list explaining each of its advantages and also talk about its disadvantages.

I will try to be as clear as possible, but if you have any doubts, you can leave them in the comments and I will help you solve it.

Comparative average yield. We can see the difference that the 3D composite L3 cache makes.

Benefits of increasing L3 cache with 3D stacking

• Allows you to increase the total amount of L3: this has a positive impact because, as we’ve already seen, it allows the CPU to have more storage capacity for data, instructions, and operations that can be accessed more quickly, reducing its reliance on RAM.
• Overcomes space limitations on silicon: L3 cache can take up a huge amount of space on a chip, in fact the 32 MB of L3 cache that the Ryzen 5000 and Ryzen 7000 add already takes up more space than the CPU cores and their L1 and L2 caches. Increasing the L3 cache horizontally would create a huge chip that would also be more complex and take up more space on the wafer.
• Reduce production costs and maintain scalability: this is a direct consequence of the above, as it allows the modularity and scalability of the chip to be preserved by considering the 3D composite L3 cache as an extra that can be mounted on top of the conventional L3 cache.
• Improves performance: and this improvement can be very large in certain scenarios. Within the mainstream consumer market, the applications that benefit the most from this are games, to the extent that a processor like the Ryzen 7 5800X3D can perform almost as well as the Ryzen 7 7700X, despite having a lower IPC, using DDR4 and running at lower frequencies. .
• Good efficiency and temperatures: processors with more stacked L3 cache perform better, but at the same time register very good temperature and consumption values ​​with respect to the peak performance they can achieve.

Disadvantages increase L3 cache using 3D stacking

• Product price: and the truth is that the price difference can be important. For example, the Ryzen 7 7700X has an approximate price of 350 euros, while the Ryzen 7 7800X3D currently costs 470 euros. In return, it offers much higher performance in games, and quite a big one, as the latter outperforms the former by 25.6% at 720p, 25.3% at 1080p and 19.6% at 1440p.
• Forces to reduce frequencies and eliminate overclocking: This is one of the most important compromises of integrating another 3D-stacked L3 cache block, as it cannot handle high voltages. Obviously, this means that CPU performance ends up being lower in those applications that aren’t able to take advantage of this extra cache. Continuing with the previous example, the Ryzen 7 7800X3D has a more limited turbo mode than the Ryzen 7 7700X, and for this reason it performs slightly lower in certain applications, such as Cinebench R23.

Final notes: cache and specialization

In general, adding more cache to a processor using 3D stacking It’s a way to specialize and focus it on a very specific scenario. Within the mainstream consumer market, this specialization is focused on gaming, as we’ve seen throughout this article, and it’s very likely that AMD will continue to maintain this focus and offer new Dedicated gaming CPU thanks to the role played by the L3 cache.

It is also interesting to see cache gaining importance in the graphics card sector from both NVIDIA and AMD, and this leads me to believe that it may be that in a few years they will also end up implementing 3D stacked cache on the GPU itself. It would not be unreasonable, in fact AMD has already resorted to external chiplets with L3 cache connected to the GPU, but it would be completely complete, especially in terms of homogeneity at the silicon level and the contact area of ​​the package.

AMD solves this problem in Ryzen with the structural silicon we mentioned, but with this technique it reduces the efficiency of the heat transfer and dissipation process, which could lead to cooling problems in certain cases. We’ll see what the future holds, but I’m pretty confident that chiplets and 3D stacking will eventually become big drivers of innovation not only in the CPU sector, where they’re already making a difference, but also in the CPU.GPU sector.