is this right, or was my lecturer talking rubbish?

[drunkenmaster]

http://www.theregister.co.uk/2011/02/24/amd_bulldozer_core_isscc/

Well thats bulldozer, clearly FPU is smaller than both the interger cores, its a drop in FPU power by design so older AMD cores had a higher FPU vs Interger ratio in terms of diesize/transistor count, but not much. 80%, utter rubbish being spewed by a moron basically.

This is part of the reason I gave up a comp sys eng degree midway through, in 18 months I'd learned next to nothing and had more incorrect info given to me than useful or new information, I was paying money to be bored to death while learning nothing and being given lots of incorrect information then tested on incorrect information, its laughable.


Itanium is a crazy specialised architecture and has very little fpu power, its all about a very specific type of data, if every single program isn't compiled basically perfectly for it you lose almost all the performance, its even great for certain types of data but near useless in other situations. I cant' find a good recent die shot that's also chopped up into pieces like that Bulldozer shot for anything Intel based but I'm sure they are out there with a few minutes effort on google.

Basically FPU + 80% of the core is utter rubbish.

 

[Code13]

when he was asking what processor speed everyone had, and someone called out 13GHz (adding up his 4 cores) he was reeling back in disbelief

Correct me if I'm wrong, but isn't that like adding together the speeds of all your wheels to find out how fast your car is going?

 

[rypt]

Correct me if I'm wrong, but isn't that like adding together the speeds of all your wheels to find out how fast your car is going?

No it is not, it is more akin to adding up the amount of grip at each contact patch to find the total grip.

If you have a 4 cores at X GHz each, then a true parallel program should run as fast (provided nothing else is limiting it) as 4*X-(overheads) GHz single core

 

[JF-AMD]

Generally speaking, the amount of FP in a typical data center is ~10-20% unless you are a bank or an oil company.

FP does take up a disproportionate amount of die space relative to the amount of benefit for most workloads, but let me throw something back at you: More effort/resource is wasted trying to get high clock speed for games and benchmarks than for floating point.

 

[drifting]

Games calculation collision might still need float arithmetics.