Pemalite said:
|
It's not all about having higher IPC's either ... Once you need ridiculously large vectors that demand AVX-1024 the IPC argument becomes totally worthless as most applications are simply don't have this structure to take advantange of the high amount of instruction level parallelism.
http://abinstein.blogspot.ca/2010/09/ipc-myths.html
Despite the fact that abinstein has a clear bias for AMD he's absolutely right about this issue. (This abinstein is probably the same one from AMDZone.)
The only way we could probably realistically increase single threaded performance in the future is that we would basically need to increase the clocks or get instruction set extensions dedicated to increasing single threaded performance.
I'm not sure if you would consider this to be a powerhouse but here it goes ... Each Power8 core has 2 VMX units and likely 4 single precision floating point units which basically means that the processor at hand is capable of 12 Fused Multiply-Add instructions so each core can pull off 24 flops/cycle. With 12 cores all clocked at 5 Ghz would mean that it can do 1.44 Teraflops tops! No that is not small and I think this may have being the first commercial CPU to do that. (Aside from the xeon phi ofcourse.)
I would agree with you about IBM controlling alot more of the server market than it has now because of more power but the tools and everything else is so much better on the x86 side of things.