Originally Posted by diggiddi
What exactly are the FPU's used for then? cos i thought they were needed for games
They most definitely are used quite heavily in games. What I was referring to specifically is AMD's re-arrangement of the FPU with Bulldozer and Piledriver to support AVX, and how game performance is, typically, not affected.
Each "module" has two 128-bit FMAC
apability) floating-point units that can be addressed individually, and can be combined to make a 256-bit FPU to service AVX instructions. Games call single or double precision floating point (32/64 bit, 4-issue/2-issue, respectively) operations routinely, and a some games call a full-width 128-bit "quad-precision" operations occasionally. AVX/AVX2 is still in its infancy and most game developers are finding little use for it, even with the ability to issue 8 single precision calculations in parallel. Well, AVX2 has FMA, but that is besides the point.
So, in their price brackets, AMD CPUs are still very competitive on the FPU front.
FX-8350 - 8 128-bit FPUs combine to 4 256-bit AVX-capable FPUs
i7-4770K - 4 256-bit AVX-capable FPUs
FX-6300 - 6 128-bit FPUs combine to 3 256-bit AVX-capable FPUs
i3-4130 - 2 256-bit AVX-capable FPUs
FX-4300 - 4 128-bit FPUs combine to 2 256-bit AVX-capable FPUs
Per integer core the FX line's FPU throughput IS
weaker than an Intel core, but it is no weaker than older CPUs that did not support AVX:
Phenom II 1100T - 6 128-bit FPUs
i7-920 - 4 128-bit FPUs
To prepare Bulldozer for AVX instructions, AMD either needed to increase the per-core die size with a 256-bit AVX FPU on each core like Intel, or rework the FPU/core arrangement so that two traditionally sized 128-bit FPUs can be combined on the fly to make a 256-bit AVX-capable FPU. AMD made an educated move to favor integer throughput at the expense of floating point throughput to limit the impact AVX would have on increasing die sizes, and therefore cost. This was probably the most wholly-successful design change AMD made from K10.Edited by M125 - 4/2/14 at 11:32am