By: DaveC (davecuny.delete@this.gmail.com), November 14, 2011 1:13 pm
Room: Moderated Discussions
A while back, I remember reading a presentation slide that mentioned a clock-normalized per-thread performance for Bulldozer at about 0.9 of a Magny-Cours core. Now that the SPEC scores are out, that seems just about right:
Clock normalized (all numbers are Rate_Base) --
at 140W Bulldozer is 0.92 of a MC core in Int
at 115W Bulldozer is 0.90 of a MC core in Int
at 85W Bulldozer is 0.93 of a MC core in Int
at 140W Bulldozer is 0.87 of a MC core in FP
at 115W Bulldozer is 0.85 of a MC core in FP
at 85W Bulldozer is 0.88 of a MC core in FP.
Per sockect comparison (still Rate_Base) --
at 140W, Bulldozer is 1.29 of a MC socket in Int
at 115W, Bulldozer is 1.19 of a MC socket in Int
at 85W, Bulldozer is 1.10 of a MC socket in Int
at 140W, Bulldozer is 1.21 of a MC socket in FP
at 115W, Bulldozer is 1.13 of a MC socket in FP
at 85W, Bulldozer is 1.05 of a MC socket in FP
So . . . Bulldozer *could* be decent if they can get the clock speed and/or thermal issues resolved. Right now, however, when adding 33% more "cores" gives you at most 29% more performance (and as low as 5-10% more performance) per socket, things aren't looking good.
Comments?
DaveC
Clock normalized (all numbers are Rate_Base) --
at 140W Bulldozer is 0.92 of a MC core in Int
at 115W Bulldozer is 0.90 of a MC core in Int
at 85W Bulldozer is 0.93 of a MC core in Int
at 140W Bulldozer is 0.87 of a MC core in FP
at 115W Bulldozer is 0.85 of a MC core in FP
at 85W Bulldozer is 0.88 of a MC core in FP.
Per sockect comparison (still Rate_Base) --
at 140W, Bulldozer is 1.29 of a MC socket in Int
at 115W, Bulldozer is 1.19 of a MC socket in Int
at 85W, Bulldozer is 1.10 of a MC socket in Int
at 140W, Bulldozer is 1.21 of a MC socket in FP
at 115W, Bulldozer is 1.13 of a MC socket in FP
at 85W, Bulldozer is 1.05 of a MC socket in FP
So . . . Bulldozer *could* be decent if they can get the clock speed and/or thermal issues resolved. Right now, however, when adding 33% more "cores" gives you at most 29% more performance (and as low as 5-10% more performance) per socket, things aren't looking good.
Comments?
DaveC



