By: Michael S (already5chosen.delete@this.yahoo.com), October 8, 2012 1:59 pm
Room: Moderated Discussions
SHK (nomail.delete@this.mail.com) on October 8, 2012 5:42 am wrote:
> Robert Myers (rbmyersusa.delete@this.gmail.com) on October 4, 2012 10:24 am
> wrote:
>
>
> One of the things i
> really don't like in HPC since the whole "clusters take over vector processors"
> has taken place is that even the top machines on top500 are stuck with commodity
> DRAM+hardware managed caches. Everybody is complaining about the memory wall but
> something like RLDRAM-3 (AFAIK) has never been used outside networking
> equipment.
> Disregarding the physical implementation seems that a Cray T90 was a
> much better and balanced architecture for HPC.
First, you will have to convince Intel/AMD to include support for RLDRAM into their IMCs. According to my understanding, RLDRAM access protocol is substantially different from SDRAM DDR2/DDR3, or any SDRAM for that matter, so the required effort is not trivial.
Second, RLDRAM would be pretty bad for capacity, except if your opt for some form of fully buffered, which by itself adds more latency than RLDRAM saves.
If I am not mistaken, with RLDRAM you can't currently get more than 512 MB per 64bit "channel". That's like going almost full decade back. With standard unbuffered DDR3 you can easily get 8 GB per channel, with registered DDR3 - up to 96 GB/channel.
Could RLDRAM provide
> Robert Myers (rbmyersusa.delete@this.gmail.com) on October 4, 2012 10:24 am
> wrote:
>
> anon (anon.delete@this.anon.com) on October 3, 2012 6:50 pm wrote:
>
>
> >
>
> > What do you mean by latency?
> >
>
> As I have been
> involved in
> discussions, latency has always referred to the latency of memory
> references,
> whether to cache, to local memory, or to remote
> memory.
>
>
> www.sandia.gov/~rcmurph/doc/latency.pdf
>
> Robert.
>
>
> One of the things i
> really don't like in HPC since the whole "clusters take over vector processors"
> has taken place is that even the top machines on top500 are stuck with commodity
> DRAM+hardware managed caches. Everybody is complaining about the memory wall but
> something like RLDRAM-3 (AFAIK) has never been used outside networking
> equipment.
> Disregarding the physical implementation seems that a Cray T90 was a
> much better and balanced architecture for HPC.
First, you will have to convince Intel/AMD to include support for RLDRAM into their IMCs. According to my understanding, RLDRAM access protocol is substantially different from SDRAM DDR2/DDR3, or any SDRAM for that matter, so the required effort is not trivial.
Second, RLDRAM would be pretty bad for capacity, except if your opt for some form of fully buffered, which by itself adds more latency than RLDRAM saves.
If I am not mistaken, with RLDRAM you can't currently get more than 512 MB per 64bit "channel". That's like going almost full decade back. With standard unbuffered DDR3 you can easily get 8 GB per channel, with registered DDR3 - up to 96 GB/channel.
Could RLDRAM provide
