Article: Parallelism at HotPar 2010
By: hobold (hobold.delete@this.vectorizer.org), August 6, 2010 4:08 am
Room: Moderated Discussions
Rohit (@.) on 8/5/10 wrote:
---------------------------
[... Cray's vector computers ...]
>From what I read on Wikipedia, it seems awfully like lrb. As for instruction chaining,
>I think in order issue @>2GHz is an adequate substitute for that.
Semantically and concerning their design intent (this is a guess, I cannot read minds :-), LRB and Cray's machines are close relatives. But the microarchitectures are vastly different. Cray used "deep pipelining" to quickly process vector elements one by one. The straight forward parallelism of the vector registers allowed him to pipeline deeper and clock higher than what was possible with scalar pipelines. LRB and other modern SIMD machines have arrays of parallel ALUs and do computation truly in parallel (and they are pipelined on top of that, of course).
But yes, Cray is the father of all current SIMD machines, at least in spirit. Almost no significant improvements have been contributed to his idea of data parallelism since then. But that doesn't stop stubborn crackpots like me from trying.
---------------------------
[... Cray's vector computers ...]
>From what I read on Wikipedia, it seems awfully like lrb. As for instruction chaining,
>I think in order issue @>2GHz is an adequate substitute for that.
Semantically and concerning their design intent (this is a guess, I cannot read minds :-), LRB and Cray's machines are close relatives. But the microarchitectures are vastly different. Cray used "deep pipelining" to quickly process vector elements one by one. The straight forward parallelism of the vector registers allowed him to pipeline deeper and clock higher than what was possible with scalar pipelines. LRB and other modern SIMD machines have arrays of parallel ALUs and do computation truly in parallel (and they are pipelined on top of that, of course).
But yes, Cray is the father of all current SIMD machines, at least in spirit. Almost no significant improvements have been contributed to his idea of data parallelism since then. But that doesn't stop stubborn crackpots like me from trying.
