AMD is now lowering expectations for Bulldozer’s performance. Originally, Fruehe claimed a figure of 50% in an official AMD blog:
From a performance standpoint, if you compare our 16-core Interlagos to our current 12-core AMD Opteron™ 6100 Series processors (code named “Magny Cours”) we estimate that customers will see up to 50% more performance from 33% more cores. This means we expect the per core performance to go in the right direction — up. That is all I will say until launch.
AMD’s Thomas Seifert is now giving some more realistic figures:
The Interlagos platform is our first server offering, optimized for today’s cloud datacenters and the architecture excels at compute-intensive and HPC workloads, where it will deliver up to 35% performance improvements compared to our current offerings.
So there we have it, I told you John Fruehe was lying about Bulldozer performance, now AMD has officially confirmed this. Sadly it also doesn’t spell that much good for Bulldozer… Up to 35% more performance from 33% more cores is not all that spectacular, for an entirely new architecture.
Update: Some people came up with the ‘explanation’ that faster AMD processors had come out between Fruehe’s and Seifert’s claims. Ofcourse they assume that I could not think of that myself, and try to do the math to prove that Fruehe was not lying.
Problem is, I DID think of that, I just discarded the idea because the CPUs had not improved that much. I have done the REAL math in one of the comments below, but I will repeat it here, because people may overlook it:
Try these benchmarks for example: http://www.cisco.com/web/DK/assets/docs/presentations/vBootcamp_Performance_Benchmark.pdf
They include some 6176SE vs 6180SE numbers… guess what? It doesn’t add up!
The 200 extra MHz and the slightly faster memory don’t add enough performance to support your theory.
Basically you have this to solve, if both statements mean the same:
X * 1.5 = Y
(X * Z) * 1.35 = Y
X * Z * 1.35 = X * 1.5
Z * 1.35 = 1.5
Z = 1.5/1.35 = 1.11
In other words, for Seifert and Fruehe to say the same, assuming one is taking about the 6176SE and the other about the 6180SE, the 6180SE has to be 11% faster than the 6176SE.
Even with a purely synthetic test such as Dhrystone, it cannot get the required ~11% improvement. It is stuck at ~8%
In the still very synthetic SpecINT, it is only ~5% faster. Nowhere near fast enough.
If we take the 5% of SpecINT, and assume this is one of those HPC scenerios they are talking about, and substitute that, we get:
1.05 * 1.35 = 1.42.
So that would still make Fruehe’s estimate off by almost 20%.
Quite simply… if you cannot make accurate predictions about performance more than a year ahead of time, then just DON’T make predictions!