[Libre-soc-dev] [RFC] horizontal SVP64 vectors

Luke Kenneth Casson Leighton lkcl at lkcl.net
Thu Jul 8 13:30:26 BST 2021


On 7/8/21, whygee at f-cpu.org <whygee at f-cpu.org> wrote:

> mapped
> to registers through the register-mapped memory : the vectors can be ANY
> length
> and reside in cache, instead of requiring crazy numbers of registers...

that makes for a memory bottleneck on the top speed achievable,
because the fanout to addressing is larger than it would be for
registers or a register cache.

if you can solve that then modern high performance could be achieved.

the STAR-100 was a memory only vector engine and it was punished by
the extra memory lookups.

Cray's work was at a time whem memory was the same speed as the processor.

now it is a 10x to 150x difference.

l.



More information about the Libre-soc-dev mailing list