You may have first thought the leak we saw in January was a bit “crazy” because it was too early for something so specific, but it was totally real. The filtering was quite correct and what we are going to see is a 50% increase in the L2 and L3 caches, curiously retaining the L1.
Intel Raptor Lake and cache: +50% in Core 13
The first thing we must take into account is that we are talking about a greater number of cores in total, in particular up to 24 between E-Cores and P-Cores. To be precise, the highest-end Core 13 (supposedly i9-13900K) will have 8 P-Core and 16 E-Core
Now it appears.
About p-core L2 cache latency, non-accuracy test result shown in pic2. https://t.co/Gf8o2V04Pa pic.twitter.com/mLarcP9lhl
– Raichu (neOneRaichu) May 18, 2022
Therefore, we will see how the configuration of L1 remains intact with a L1D with 8 x 48 KB + 16 x 32 KB and an L1I with 8 x 32 KB + 16 x 64 KB respectively, making a total of 2.176 KB for the entire L1 cache. This number is increasing, but it does so because there are more cores as such, maintaining parity with Alder Lake in all of them.
In the L2 it changes, we go from 1.25 MB for the P-Core cores to 2 MB, and from 2 MB in the E-Core to 4 MB in a configuration for the first 8 clusters for the first (2 MB per core
An astonishing total: 32 MB of L2 and 36 MB of L3
The sum is simple at this point: 68MB of total cache without adding the L1 where there is no change as such other than the named amount. In other words, we went from 44MB in total with Alder Lake for these two caches to 68 MB of Raptor Lakean augmentation of 54.54%. Also, Raptor Lake will have better latency than Alder Lake as you adjust the cache block size, so performance will take off a bit here.
As the diagram shows in a simpler way, the architecture will actually have the cache assigned per core as follows: 2 MB for each Raptor Cove Core and 1 MB for each Gracemont Corei.e. Intel has put a 2:1 parity between the P-Cores and E-Cores for the L2.
On the other hand, with the L3, a different directive and parity is followed for each 3MB cluster, but individually we have 3 MB for each P-Core Raptor Cove and 1.33 MB for each Gracemont E-Core. This shows that E-Cores need a lot of L2 cache and are very lightly dependent on L3.
Therefore, it is assumed that the Core 14 Meteor Lake Being a major architectural leap, these proportions will change again, but only for the E-Cores, which seem to follow a similar path in their L2 and L3, going from 4 MB for the first to 3 MBwhich should increase the size of the second to compensate for the imbalance produced.