What is AMD Infinity Cache Memory Technology?

We take a closer look at the RDNA 2 architecture and Infinity Cache technology used in the RX 6000 series.

AMD’s RDNA 2-based RX 6000 series graphics cards gained great interest with their introduction, and the competition took on a new dimension. We saw both hardware and software innovations in the presentation of the red team. In addition, we have three new models: RX 6800, RX 6800 XT and RX 6900 XT.

One of the most striking of these was undoubtedly Infinity Cache memory technology. Although GPUs, which are at the heart of graphics cards, are very important, memory (VRAM) is in the second place.

All three graphics cards in the RX 6000 family have 16 GB GDDR6 memory and 256-bit memory interface. So it comes to mind that bandwidth may be very limited. At this point, Infinity Cache technology comes into play.

Although AMD does not go into too much detail in its presentation, we can describe it as a large and very fast chip-unified (on-chip) cache structure, like the L3 cache in the CPU. However, conceptually, it is perfect for large operations that are used as a buffer for reading and writing operations from the main memory and are much faster so that the data can be directly processed without having to go to VRAM. In addition, instead of waiting for the GPU data to come from VRAM, it resorts to caching, improving the amount of throughput per clock cycle on the GPU.

Internal caches for GPU usage are not a new idea, especially for AMD. The company was offering 32 MB of eSRAM cache for its Xbox One and Xbox One S SoCs. Also, the previous Xbox 360 had an eDRAM. However, it is the first time that a large cache has appeared in a GPU used on the PC platform.

Navi 21 GPUs, also known as Big Navi, come with 128 MB of Infinity Cache. The red team has not mentioned different GPUs. However, we will likely see smaller caches.

Infinity Cache Structure

Assuming AMD is using standard 6T SRAM, the Infinity Cache structure of the Navi 21 should have at least 6 billion transistors. This is an important figure even in the advanced 7nm production of TSMC and as a matter of fact, a significant part of the GPU should be allocated to this memory. On the other hand, AMD may have made optimizations that we do not know.

The new Infinity Cache uses the redesigned buses of the GPU to significantly maximize performance while minimizing the power consumed within the GPU.

Energy efficiency

Having this kind of cache on the GPU also has important advantages on the energy efficiency side. As for the performance per watt, this cache improves energy efficiency by reducing the amount of traffic to the high energy consuming VRAM.

It also allows AMD to get rid of a small memory subsystem with fewer DRAM chips and fewer memory controllers. As a result, the power consumed is slightly less.

Result

It was said that memory bandwidth will increase by 117% with Infinity Cache technology. As you know, when we switch to resolutions such as 2K, 4K and 8K, memory becomes more and more important.

NVIDIA party solved this issue with GDDR6X memory technology. AMD took a different route and opted for a fast cache design.

AMD says Infinity Cache is based on the Zen CPU’s L3 cache design, which is the fruit of cross-work between AMD’s CPU and GPU teams. As you know, although HMB based memory is fast, it is quite expensive. Red team thinks Infinity Cache technology is a better engineering investment than going for this kind of expensive solution.

Finally, it was mentioned that this technology will make significant contributions to ray tracing performance. Because most of the processed data sets are kept closer to the computing units in the GPU.