From: already5chosen@yahoo.com   
      
   On Sat, 20 Sep 2025 07:56:49 GMT   
   anton@mips.complang.tuwien.ac.at (Anton Ertl) wrote:   
      
   >   
   > Yes, organizing the interconnect in a hierarchical way can help reduce   
   > the increase in interconnect cost, but I expect that there is a reason   
   > why Intel did not do that for its server CPUs with P-Cores, by e.g.,   
   > forming clusters of 4, and then continuing with the ring; instead,   
   > they opted for a grid interconnect.   
   >   
   > - anton   
      
      
   I don't know for sure, but would imagine that the reason is that their   
   server CPUs with P-core have the same design for low-to-mid end "cloud"   
   models and for high-end "enterpise" models. High-end models have OLTP   
   and similar enterprise workloads as rather important market. Flatter   
   LLC is better for OLTP/enterprise than dozen or two of separate L3   
   caches. Besides, their current L2 caches are rather big, so if they   
   make those separate L3s true exclusive, which is optimal for reduction   
   of cc traffic, then there would be rather big waste of total cache   
   capacity.   
      
   An alternative is to left LLC intact and instead make L2s shared by   
   pairs of cores. That is unacceptable because of yet another market   
   addressed by the same Xeons line - computations/HPC, where being   
   limited by L2 bandwidth is not rare even now. With shared L2 it will   
   become very common.   
      
   3 different uncore designs for 3 different markets can solve that   
   nicely, but of course in the Intel's current financial situation that   
   is unthinkable. Probably even current arrangement with 3 Xeon lines   
   (Xeon-E = desktop chips with E-cores fused off, Seirra Forrest = plenty   
   of Crestmont cores and "normal" Xeons currently represented by Granite   
   Rapids) could be unsustainable.   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)