As with processors, GPUs take advantage of the cache between them and VRAM. On AMD and Nvidia The design of caches in different sizes is just as important.
In order not to adversely affect the performance due to the increasing requirements in various applications, companies develop their cache designs on the one hand, and on the other hand. larger caches positioning. For example, in Nvidia’s A100 GPU at the L2 level 40 MB of storage this figure is 7 times that of the V100.
The guests of the last test were RDNA 2 and Ampere architectures. Measuring the latency GPUs experience in accessing their caches, the team has produced remarkable results.
According to the observations of the test team using the OpenCL-based simple pointer tracking test, Nvidia’s Ampere GPU has an additional 116 ns delay being added.
While the GA102’s massive build is considered to be a major problem for the L2 cache On the AMD wing, this figure is 66 ns. measured as.
You may be interested in
Detailed RTX 3080 Founders Edition Render images shared
Due to the fact that Nvidia includes 2-step caching in Ampere As a result of exceeding 4 MB, VRAMs are dropped where the delay of 268 ns has been reached.
On the other hand, the last stage of AMD, the 3-stage cache structure known as Infinity Cache (L2), offers lower latency in L1, as well as at the beginning when switching to L2 only 20 ns caused delay.
Also Ampere’s required searching data in VRAM period, data in the Infinity Cache of the RDNA 2 architecture GPU with low latency It is worth noting that it continues to access. The waiting time for both chips to access VRAMs is measured equally.
High cache of AMD success in low resolutions It can be said that it has a contribution. On the other hand, looking at the previous generations, the path AMD has come to reveals.
See Also. “RTX 3080 Ti’s on their way”
On the Nvidia side, Maxwell and Pascal were mostly similar, but tested GPU dimensions of the 980 Ti thought to have fallen behind due to. Similar results were obtained in the Turing-Ampere comparison.