So, stock Blender builds are a terrible benchmark unless one wants to go back to, eh, mid-2011 or so. Let's be generous and say 2012.
No not really, you are not understanding what has happened. by your logic 99% real world apps aren't good benchmarks ( think about that)
Blender is a renderer, it uses 32bit 4packed vectors (128bit, x,y,z,a).
SSE is 128bit so that is fine.
Now AVX doesn't have across the board support thax to intel, but avx really only adds 3 operand support and 256bit op support but because we are working on xyza 256bit wont give you much.
AVX2 adds the same as AVX but for integer and some memory/register/data manipulation operations.
Now the build that the stilt has built has advanced Vectorization
WRITTEN BY A INTEL CODE NINJA. They have likely found a way to vectorize code that was otherwise being executed as scaler FP operations. Now the question is does that advanced vectorization require a operation only found in AVX/AVX2 or can in work on SSE. But either way by Packing those operations together and executing them as one operation you remove pressure on Several bottlenecks that exist in Bulldozers design.
the obvious ones:
1. the FPU store bandwidth
2. filling the FPU front end
But there could be others:
1. less pressure on decode
2. less L1i thrashing/aliasing
3. less pressure on store pipeline
etc,etc
Again Blender isn't the problem!
Go build as 2.75 build with AVX2 and watch it performance very close if not the same as 2.75..........
EDIT: CON cores are good at SSE and very good at INT SIMD in general, when they came out that was one of their biggest strengths.
).
Facebook
Twitter