Amd ryzen has its weak spot in l3 cache by ccx design

The new AMD Ryzen 7 processors have shown very good overall performance, however there are a few cases where their performance declines quite strangely. Apparently the biggest weak point of AMD's new processors is its memory subsystem, a point where Sunnyvale's have to work very hard before the arrival of Ryzen 3 and Ryzen 5 to improve speed and latencies.

L3 cache is AMD Ryzen's big weak point

Hardware.fr has made an exhaustive examination of the memory system and the cache of the new AMD Ryzen 7 processors. Apparently there is a problem in the implementation of the L3 cache in Ryzen, this memory has very high latencies (100ns) that can be up to 30 ns higher than in the case of the Intel i7 and even the previous AMD FX (70 ns).

We continue to investigate the cache of the new Ryzen and we see that the L1 is still very far from the performance of Intel's processors, on the other hand, the L2 cache of Ryzen manages to offer higher speeds than those of Intel, although with somewhat higher latency. The biggest weakness of the L3 is found in a latency almost three times that of Intel.

In the case of Intel Core i7-6900K processors, which have 32 KB of L1 cache, the performance is maximum until the data to handle does not fit within L1, then they have to jump to the L2 cache that has a size 256 KB, in case the volume of data is greater, it would have to go to the L3 cache that has a capacity of 20 MB. If the data is greater than 16 MB then it is forced into the main memory of the system that has a latency of 70 ms.

In the case of the Ryzen 7 1800X everything works well in the case of the L1 and L2 caches that are 32 KB and 512 KB respectively. However, when we get to the L3 cache, the behavior is totally different, up to 4 MB of L3 utilization, we see an increase in latencies that corresponds to what was expected, however, latencies increase dramatically when 16 MB of the L3 cache. This problem would be derived from the CCX modular design of the new Ryzen processors, each of the modules consists of four cores and 8 MB of L3 cache.

Ryzen L3 cache's uneven performance between using 4MB or using 8MB is due to its modular design causing differences in access time depending on where part of the L3 is accessing the CCX complex. If you are using only the four cores of a CCX complex, you only have access to 8 MB of cache, while if you were using two cores of each CCX complex, you could use the total 18 MB of L3 cache.

AMD Ryzen 7 1700 Review in Spanish (Complete Analysis)

In the latter case, performance would still be limited by the bandwidth of the AMD Data Fabric interconnect bus that links the CCX complexes with a bandwidth of only 22 GB / s, a much lower figure than the 175 GB / s of the cache. Intel's L3 and even that RAM.

New issue for Ryzen, AM4 motherboard shortage

WE RECOMMEND YOU Radeon Software Adrenalin 2019 Edition 19.7.1 Now Available

The new AMD Zen architecture offers excellent performance, AMD has opted for a design that achieves the best balance between performance, cost and scalability thanks to its CCX modules. However, this design would explain the cause of lower-than-expected performance in some highly cache-dependent scenarios, such as gaming.

Source: techpowerup