Article: Parallelism at HotPar 2010
By: hobold (hobold.delete@this.vectorizer.org), August 6, 2010 7:41 am
Room: Moderated Discussions
Rohit (@.) on 8/6/10 wrote:
---------------------------
[...]
>You seem to be suggesting that the physical width of ALU's was less than the logical
>width exposed to programmer on Cray's.
>
>Is that correct?
I don't know all Cray models in that much detail, but at least initially, they had an ALU width of a single vector element. Later machines might have had more than one ALU of a type to facilitate chaining. But even then, distinct ALUs would process distinct instructions, rather than different vector elements of the same SIMD instruction. Their main memory used to be made of SRAM, so loads and stores achieved a throughput that matched processing capabilities.
Modern SIMD architectures may choose to implement some of their instructions as multipass operations through an essentially scalar ALU, but in general, fully parallel execution in replicated hardware blocks is the goal they are shooting for these days. It's not like we knew much better things to do with all these transistors ...
---------------------------
[...]
>You seem to be suggesting that the physical width of ALU's was less than the logical
>width exposed to programmer on Cray's.
>
>Is that correct?
I don't know all Cray models in that much detail, but at least initially, they had an ALU width of a single vector element. Later machines might have had more than one ALU of a type to facilitate chaining. But even then, distinct ALUs would process distinct instructions, rather than different vector elements of the same SIMD instruction. Their main memory used to be made of SRAM, so loads and stores achieved a throughput that matched processing capabilities.
Modern SIMD architectures may choose to implement some of their instructions as multipass operations through an essentially scalar ALU, but in general, fully parallel execution in replicated hardware blocks is the goal they are shooting for these days. It's not like we knew much better things to do with all these transistors ...