AVX2 extends Intel AVX by promoting most of the 128-bit SIMD integer instructionshttp://software.intel.com/file/36945
with 256-bit numeric processing capabilities. AVX2 instructions follow the same
programming model as AVX instructions.
In addition, AVX2 provide enhanced functionalities for broadcast/permute operations
on data elements, vector shift instructions with variable-shift count per data
element, and instructions to fetch non-contiguous data elements from memory.
Competitive landscape
AMD,on the other hand ,will be supporting their own XOP(128bit) ISA with Bulldozer Ver1 coming out this summer and with 3 new ISA extensions with Bulldozer Enhanced due out in 2012.There is a chance they will upgrade the XOP ISA to 256bit too.As for AMD's AVX2 support ,by the time Haswell arrives ,AMD should be preparing their NextGen Bulldozer core so this is a solid candidate for AVX2 support.
Since AVX1 is presently done in AMD's Bulldozer core via 2 FMACs working together,it shouldn't be that hard to do the same thing with AVX2,only now the MMX/SSE integer portion(each FMAC has 1 MMX/SSE 128bit unit) will be ganged together to execute one 256bit integer instruction.NextGen Bulldozer should have 256bit capable FMACs (each) ,so doing integer/fp 256bit SIMD will be more efficient and faster. AMD should have an upper hand overall since they support the more powerful Fused Multiply Add ISA(FMA4) that will come even in the first Bulldozer version due out this summer.Haswell will support FMA3 so intel will be playing catchup game in this department. Both designs will have to run recompiled software in order to get the most out of their capabilities(be it FMA or AVX).
In the end,here is a short summary of ISA extensions that will be supported by Bulldozer Ver1 core.Source is here.Haswell should support most of these plus AVX2/FMA3.Intel won't support XOP and FMA4.
No comments:
Post a Comment