At its Financial Analyst Day event on Thursday, Advanced Micro Devices said that its chips in 2014 will be able to automatically decide which of its heterogeneous cores to use for a particular task. The context switching between x86 and stream processing cores will allow AMD to greatly speed up performance of its chips and optimize their power consumption.
"Ultimately, as we complete our roll-out of heterogeneous system architecture, we will have a very fine-grain control of where is computing [performed], it is an optimization," said Mark Papermaster, chief technology officer of AMD.
At present AMD's accelerated processing units - which integrate x86 general-purpose processing cores and highly-parallel stream processors on a single chip - can accelerate certain tasks using Radeon stream processors only when software is made to specifically use them. Given the limitations of the current architecture, x86 and Radeon cores should use dedicated memory, which tends to be inefficient. Essentially, nowadays the software completely controls which compute resources to use and when.
But after years of evolution, which will involve development of both hardware and software/compilers/tools, accelerated processing units in 2014 will be able to dynamically decide (possibly, when it comes to new programs) which task is better to execute on a particular core thanks to new software as well as special features of the chips.
Dynamic context switching between different types of cores will not only greatly speed up performance of such chips, but will also optimize power consumption as the most efficient hardware will be used to perform an operation.
Earlier AMD expected to release "fully fused" Fusion chips in 2015 or beyond.