P6 Blade or: The effect of having less cache on a Power6

In an earlier blog entry i had the strong suspicion, that the reduction of the L2 cache to 4 MB and the missing L3 cache won´t be without an impact to performance. There is an document in the internet, called “IBM System p, BladeCenter Performance Report”. I searched a little bit in the document and found some hints that my suspicion may be correct. At first: As a baseline i want to make the following assumption: As the fastest processor of the pSeries is 4,7 Ghz and the fastest processor in the blade is clocked with 4 Ghz, you can´t compare the numbers directly. So lets assume linear clock scaleability : 4 of 4,7 Ghz is 85%. Okay. The cited document contains a SPECint_rate_2006 number of 84,7 for the P6 in the blade, the same number for the p570 with 4 cores at 4,7 Ghz is 122. 84,7 of 122 is 69,5%. The cache reduction of costs you 12 SPECint2006rate points. The STREAMS benchmark is even worse: The JS22 has an TRIAD value of 15,701 MByte per Second, the p570 has a TRIAD value of 29404. with 4 cores in both configurations. I won´t do the math. It´s quite obvious thas the value is halved. Thus even when the Blade is touted as an p6 blade, you won´t get p6 performance out of the blade. I assume, they had to rip off some parts (cache,frequency …) of the P6 to get the proc into the thermal budget of a blade chassis. I think i will do some further studies in this topic to harden my suspicions. At the moment it´s only a thought game based on public informations …