[Info-vax] HP Users Hope Whitman Can Persuade Oracle to Change Itanium Decision

[Michael S]

On Oct 2, 2:57 pm, MG <marcog... at SPAMxs4all.nl> wrote:
> On 2-10-2011 14:44, Michael S wrote:
>
> I forgot to comment on this:
>
> > In fact, 256 blades * 32GB/blade = 8 TB. 128 TB is an address space,
> > not the real memory you could plug into it.
>
> > 256 sockets, 16 TB  is a limitation for directly addressable memory in
> > a single system image.
>
> > 256 sockets = 2560 cores, 2.5 times more than 1024 cores in SGI Altix
> > 4700. And each core in UV is way faster than that of 4700.
> > SpecFp_rate(base):
>
> > [...]
>
> All nice and well, but I quoted those figures directly from SGI.  So,
> perhaps you should contact them and tell them to 'get' their 'facts
> straight'?

I am sufficiently certain about 32GB/blade limit in 4700. Less certain
about 256 blades per SSI domain. It is possible they can do a bit more
with processorless "memory" blades. But not much more and certainly
not 128 TB.

Look, for example, at spec of NASA Columbia supercomputer, which is
the biggest publicly known Altix 4700 installation. On average they
have less than 1 TB of memory per SSI node.

>
> > I don't understand that tech sufficiently well to say more.
>
> Maybe you should study it a bit more before saying anything at all?
>
>   - MG

The info about 1 PB of non-CC globally-addressable memory supported by
Altix-UV came in forum post from SGI employee.
I tend to believe him.
http://www.realworldtech.com/forums/index.cfm?action=detail&id=105363&threadid=105281&roomid=2