NostaSeronx
Diamond Member
- Sep 18, 2011
- 3,689
- 1,224
- 136
One flop , two flops....
Actually FLOP is both singular and plural
One flop , two flops....
The units are their 8 units = 2*256bit and 2*128bit per core
4 = 4*32bit operands
8 = 8*64bit operands
Operands x Units
i7 2600K4C SB 2600K @ 3.4GHz
(2*4) * 4c * 3.4GHz = 108.8GLOPS
THE END
Actually FLOP is both singular and plural
i7 2600K
64 Flop x 3.4GHz = 217.6 GFlops it is halved if FADD purely or FMUL purely exists
FX-8150
32 Flop x 3.6GHz = 115.2 GFlops
Are you for real?
No, I am a unicorn in a mystical land of fairies
Yes, I thought so. Anyways, you're still wrong but I'm done with you. Enjoy your ignorance. Don't let anyone ever teach you anything, you know everything you'll ever need to know.
plonk
You're still wrong, a SB core can do 2*256bit AVX FP per cycle.
Are you for real?
He s half right at least...
SB ports are not universal , meaning that one port make
FADDs while the other exe port will do FMULs..
If only FADDs or only FMULs are dispatched , only one
port will be active , halving the thoughput.
In contrast , the two Fmac units in a BD FPU can both
do both operations.
If only FADDs or only FMULs are dispatched , only one
port will be active , halving the thoughput.
with two running threads it's less likely to occur, you can have a series of VADDPs in one thread and a series of VMULPS in the other thread at a given moment for example
Thanks to the compiler that is aware than sending two simultaneous
FADD or FMUL in a same thread will be innefficient in respect of the Uarch , even if a computation would normaly require two same ops to be computed first due to data dependencies...
from my experience the scheduling at the code level has nearly no impact on Sandy Bridge, the Intel compiler generally schedule mixed VMULPx and VADDPx, the cases I was refering to where for typical use cases like a scaling loop with only MUL or an offset loop with only ADD
There sure are a lot of non-essential parallel apps opened during a performance test...
In that case , it s possible to do these ops
in two threads if there s no data dependencies ,
otherwise
o/t
Here is a good article from the past. Where AMD went with a dual FX cpu platform to compete with Intel's new (at the time) quad core.
That was interesting at the least.
Maximum PC - Feb 2007