More details about 32-way Horus chipset for Opteron

  • Thread starter Thread starter Yousuf Khan
  • Start date Start date
Y

Yousuf Khan

Looks like Newisys is using a directory-based cache coherency chip to take
Opterons beyond their 4- or 8-way internal limitations. The people who
designed this chipset seem to be the same people who designed IBM's Summit
chipset for Xeon.

There's also some interesting chatter about why IBM chose not to use
Newisys's designs when most of the people who started Newisys were ex-IBM.
Some suggestion about protecting their Power and Xeon turf, or something or
another. :-)

http://www.cbronline.com/article_news.asp?guid=2B3F952C-AC0D-4CCB-A285-D9E87E37A90B

Yousuf Khan
 
Yousuf Khan said:
Looks like Newisys is using a directory-based cache coherency chip to take
Opterons beyond their 4- or 8-way internal limitations. The people who
designed this chipset seem to be the same people who designed IBM's Summit
chipset for Xeon.

There's also some interesting chatter about why IBM chose not to use
Newisys's designs when most of the people who started Newisys were ex-IBM.
Some suggestion about protecting their Power and Xeon turf, or something or
another. :-)

http://www.cbronline.com/article_news.asp?guid=2B3F952C-AC0D-4CCB-A285-D9E87E37A90B

Yousuf Khan

It is absolutely positively not true that the folks from Newisys were
the ones who designed the Summit chipset. The Summit chipset was
designed by a group in Rochester, MN. I believe the Newisys guys
history is Austin. And Phil Hester hasn't been IBM for a long time.

del cecchi
 
del cecchi said:
It is absolutely positively not true that the folks from Newisys were
the ones who designed the Summit chipset. The Summit chipset was
designed by a group in Rochester, MN.

If Summit is derived from Sequent's ccNUMA, then it wasn't a design
from scratch - the original designers would have been Sequent
engineers.
 
Looks like Newisys is using a directory-based cache coherency chip to take
Opterons beyond their 4- or 8-way internal limitations. The people who
designed this chipset seem to be the same people who designed IBM's Summit
chipset for Xeon.

There's also some interesting chatter about why IBM chose not to use
Newisys's designs when most of the people who started Newisys were ex-IBM.
Some suggestion about protecting their Power and Xeon turf, or something or
another. :-)

Note that IBM's Opteron design is not an in-house design as the
article seems to suggest, it's actually an MSI server. Note the
resemblance:
 
Looks like Newisys is using a directory-based cache coherency chip to take
Opterons beyond their 4- or 8-way internal limitations. The people who
designed this chipset seem to be the same people who designed IBM's Summit
chipset for Xeon.

There's also some interesting chatter about why IBM chose not to use
Newisys's designs when most of the people who started Newisys were ex-IBM.
Some suggestion about protecting their Power and Xeon turf, or something or
another. :-)

http://www.cbronline.com/article_news.asp?guid=2B3F952C-AC0D-4CCB-A285-D9E87E37A90B


Err.. ok... Attempt number 2 at posting this message, hit the wrong
key last time :>

IBM's server is not actually an in-house design, it was designed by
MSI. Here's a link to the two servers:

http://www-1.ibm.com/servers/eserver/opteron/325/index.html

http://www.msi.com.tw/program/products/server/svr/pro_svr_detail.php?UID=510


So really there are no conspiracy theories required for why IBM did
not chose Newisys' design, they simply went for a lower-cost option
from MSI.
 
M. Ranjit Mathews said:
If Summit is derived from Sequent's ccNUMA, then it wasn't a design
from scratch - the original designers would have been Sequent
engineers.

And if my aunt had nuts she'd be my uncle. Summit was a clean design by
Rochester. I don't recall Sequent even having been acquired when the
project started. I vas dere Charlie. The system was a joint effort by
Rochester, Austin and Raleigh. Beaverton was late to the party. The
chips were designed in Rochester.

del cecchi
 
del said:
And if my aunt had nuts she'd be my uncle.

That's one of the best lines I've heard it a while - have to remember
that one

Cheers,
Mike

Summit was a clean design by
 
del cecchi said:
And if my aunt had nuts she'd be my uncle.

uhh... never mind. :-)
Summit was a clean design by
Rochester. I don't recall Sequent even having been acquired when the
project started. I vas dere Charlie. The system was a joint effort by
Rochester, Austin and Raleigh. Beaverton was late to the party. The
chips were designed in Rochester.

"Sequent" or IBM Beaverton did have a project underway at that time but it
was Itanium based only, the viper crossbar chipset was to build a 16x
NUMA/SMP using seperate address/data buses to connect 4x quads. Sequent put
all its eggs into the Monterey/Win64/Itanium basket.

The early NUMA-Q design for x86/Dynix is a modified SCI protocol in a ring
topology using the "datapump" chip.


-brig
 
Tony said:
So really there are no conspiracy theories required for why IBM did
not chose Newisys' design, they simply went for a lower-cost option
from MSI.

No, but IBM has a reputation for going with best of breed products. The
Newisys servers with their management processors and high sophistication
would've fit into IBM's regular approach to server design.

Yousuf Khan
 
Looks like Newisys is using a directory-based cache coherency chip to take
Opterons beyond their 4- or 8-way internal limitations. The people who
designed this chipset seem to be the same people who designed IBM's Summit
chipset for Xeon.

You scare me. Did you hear what the Summit's NUMA factor is, once
you are outside of the first box? I heard ridiculous claims of 150.

-- Pete
 
No, but IBM has a reputation for going with best of breed products. The
Newisys servers with their management processors and high sophistication
would've fit into IBM's regular approach to server design.

That may well be IBM's reputation, but their product line-up doesn't
necessarily back up that reputation, at least when looking at low-end
x86 servers. Not that the MSI servers are bad at all, and they do
indeed have build in management processors. Newisys' design might be
slightly higher end, but the specs on the two servers aren't really
all that different.

FWIW the e325 Opteron server is not the only system that MSI produces
for IBM. The near-identical e335/e336 Xeon servers are also produced
by MSI, and apparently also one of their 2U Xeon servers.

IBM's x86 server business has been kind of losing out to HP and
especially Dell over the past while. That likely has a lot to do with
going for a lower-cost option on these servers.
 
No, but IBM has a reputation for going with best of breed products. The
Newisys servers with their management processors and high sophistication
would've fit into IBM's regular approach to server design.

IBM has also been using MSI boards quite a bit for some of their
workstations and stuff.
--
L.Angel: I'm looking for web design work.
If you need basic to med complexity webpages at affordable rates, email me :)
Standard HTML, SHTML, MySQL + PHP or ASP, Javascript.
If you really want, FrontPage & DreamWeaver too.
But keep in mind you pay extra bandwidth for their bloated code
 
del cecchi said:
It is absolutely positively not true that the folks from Newisys were
the ones who designed the Summit chipset. The Summit chipset was
designed by a group in Rochester, MN. I believe the Newisys guys
history is Austin. And Phil Hester hasn't been IBM for a long time.

del cecchi

Just to set the record straight, I lead the taskforce in IBM that
established the Summit direction for the xSeries (It actually had an
internal code name different from Summit). I then went on to lead the
team that laid out the original high level design, the function
partitioning across the various chips for both Xeon and Itanium II,
the system design and simulation/performance efforts. My team was also
responsible for overseeing the design activity (which was done in
Rochester) to insure that it met both the architectural requiements
and performance objectives for Summit. I was an IBM Fellow out of
Watson Research. And all of this started before Sequent came on the
scene at IBM.

I retired from IBM in 2000 and joined Newisys.

The big difference between the IBM design and the Newisys/AMD design
is that Summit has a local L3, while Horus has a remote L3 (Remote
Data Cache - RDC) and a Remote Directory (RD). These Horus structures
were made possible by the cHT architecture and the AMD/Opteron
implementation that offered additional cache state change status to
Horus. (Such a design was not possible with the Intel FSB, either its
architecture or its implementation).

Our simulation studies show that some applications do very well with a
RDC, others, very well with a RD. So we adopted a belt and suspenders
approach in Horus. The implementation boundary between the RDC and the
RD is adjustable in our Horus design at boot time. We also expect that
the size of the system (how many nodes) will also influence the
tradeoff betwen these two structures.

The model we see is that systems built with Opteron and our Horus
hardware and system managmement software will be tuned for
performance. This tuning can be quite narrow when it is a single
application (e.g. a data base server with a particular data base) or
rather broad when a family of applications (e.g. web server doing many
different types of transactions) is run.

We await our first hardware which should occur before the end of this
year so that we can validate our design and its performance.

Rich
 
Back
Top