Why not in chip memory controller?

Does any one have insights on this issue? AMD's memory controller
showed obvious advantages over Intel's FSB so far. Why can't Intel just
copy that? Once I saw somebody from Intel saying that have the
controller on board instead of in chip benefits some other stuff of a
computer, such as integrated graphics. Do you really think that's the
reason they refuse to put the controller in their chip?

* :

Does any one have insights on this issue? AMD's memory controller
showed obvious advantages over Intel's FSB so far. Why can't Intel just
copy that? Once I saw somebody from Intel saying that have the
controller on board instead of in chip benefits some other stuff of a
computer, such as integrated graphics. Do you really think that's the
reason they refuse to put the controller in their chip?

Why intel doesn't use integrated memory controllers, well you have to
ask them, they probably have their reasons. But fact is that integrated
memory controllers (which lead to a NUMA config) not only have
advantages but also disadvantages in SMP configurations (NUMA effect).
In short and easy words, accessing the memory which is connected
directly to the cpu is fast while accessing the memory connected to
other cpus is slow. UMA memory systems on the other side have a somewhat
constant memory performance.

Benjamin

"Benjamin Gawert" wrote in message
...
* :

Does any one have insights on this issue? AMD's memory controller
showed obvious advantages over Intel's FSB so far. Why can't Intel just
copy that? Once I saw somebody from Intel saying that have the
controller on board instead of in chip benefits some other stuff of a
computer, such as integrated graphics. Do you really think that's the
reason they refuse to put the controller in their chip?

Why intel doesn't use integrated memory controllers, well you have to ask
them, they probably have their reasons. But fact is that integrated memory
controllers (which lead to a NUMA config) not only have advantages but
also disadvantages in SMP configurations (NUMA effect). In short and easy
words, accessing the memory which is connected directly to the cpu is fast
while accessing the memory connected to other cpus is slow. UMA memory
systems on the other side have a somewhat constant memory performance.

Benjamin


Actually, NUMA has some seriouse advanatages in SMP. Please check out

http://www.anandtech.com/IT/showdoc.aspx?i=2745

* Carlo Razzeto:

Why intel doesn't use integrated memory controllers, well you have to ask
them, they probably have their reasons. But fact is that integrated memory
controllers (which lead to a NUMA config) not only have advantages but
also disadvantages in SMP configurations (NUMA effect). In short and easy
words, accessing the memory which is connected directly to the cpu is fast
while accessing the memory connected to other cpus is slow. UMA memory
systems on the other side have a somewhat constant memory performance.

Benjamin


Actually, NUMA has some seriouse advanatages in SMP. Please check out

http://www.anandtech.com/IT/showdoc.aspx?i=2745

This is no article about UMA vs NUMA but between XEON vs. Opteron. But
then Andandtech like all the other self-proclaimed hardware experts
doesn't have any experience with anything else than PCs.

However, NUMA can have a huge performance penalty in several
applications, it's definitely not always better than UMA...

Benjamin

Carlo Razzeto wrote:
Actually, NUMA has some seriouse advanatages in SMP.

Are there any advantages to NUMA per se, as opposed to just much higher
total bandwidth? My impression is that the big advantage of NUMA is that in
practice it will often allow much higher total bandwidths ( 2x on Opteron
vs. Nocona). But in theory, if you can have the same total bandwidth and
bandwidth per processor with fully uniform memory access, that should have
the highest performance. (In practice, this gets really, really expensive.)

--
Nate Edel http://www.cubiclehermit.com/

"What's the use of yearning for Elysian Fields when you know you can't get
'em, and would only let 'em out on building leases if you had 'em?" (WSG)

"Nate Edel" wrote in message
...
Carlo Razzeto wrote:
Actually, NUMA has some seriouse advanatages in SMP.

Are there any advantages to NUMA per se, as opposed to just much higher
total bandwidth? My impression is that the big advantage of NUMA is that
in
practice it will often allow much higher total bandwidths ( 2x on Opteron
vs. Nocona). But in theory, if you can have the same total bandwidth and
bandwidth per processor with fully uniform memory access, that should have
the highest performance. (In practice, this gets really, really
expensive.)

--
Nate Edel http://www.cubiclehermit.com/

"What's the use of yearning for Elysian Fields when you know you can't get
'em, and would only let 'em out on building leases if you had 'em?" (WSG)

Well, I guess there are a couple a problems with that... First like you said
is cost... Technically Intel could make a system with as much bandwidth
available to each CPU as an 8 way opteron gets, the problem is the same
configuration in say, 16 would be offering each CPU half of what a similar
16 way opteron would. So in the end, the real problem with UMA v. NUMA is
scaling. Currently Intel's position on the IMC issue is "we don't need it,
more cache will do the same thing". I guess for uniprocessor, DC and small
scale multi-socket systems they have a point, they just end up losing in
higher end configurations. As the benchmark I linked to demonstrates
currently it only takes a 4 socket (8 core) configuration for intel to start
feeling some serious hurt with their UMA arch.

I guess it should be noted that what Ben pointed out is technically an
issue. In some cases, particularly if your OS is not NUMA aware you can end
up with slightly higher latencies when a process requires data stored in
memory controlled by another CPU. However, first off most OSs *are* indeed
NUMA aware, so this is minimized. Also, at least in AMD's case this type of
situation doesn't lead to the kind of high latencies and stalls you might
think, the system is fairly well optimized. I would venture a guess that
most people developing a NUMA system would consider this and optimize as
best they can for it.

Carlo

"Benjamin Gawert" wrote in message
...
* Carlo Razzeto:

Actually, NUMA has some seriouse advanatages in SMP. Please check out

http://www.anandtech.com/IT/showdoc.aspx?i=2745

This is no article about UMA vs NUMA but between XEON vs. Opteron. But
then Andandtech like all the other self-proclaimed hardware experts
doesn't have any experience with anything else than PCs.

However, NUMA can have a huge performance penalty in several applications,
it's definitely not always better than UMA...

Benjamin

Technically yes, but these are really the only NUMA v. UMA benchmarks that
are readily available. They do however demonstrate how UMA systems tend to
become more quickly and easily memory starved than NUMA systems.

Carlo

Carlo Razzeto wrote:
"Benjamin Gawert" wrote in message
...
* Carlo Razzeto:
Actually, NUMA has some seriouse advanatages in SMP. Please check out

http://www.anandtech.com/IT/showdoc.aspx?i=2745
This is no article about UMA vs NUMA but between XEON vs. Opteron. But
then Andandtech like all the other self-proclaimed hardware experts
doesn't have any experience with anything else than PCs.

However, NUMA can have a huge performance penalty in several applications,
it's definitely not always better than UMA...

Benjamin

Technically yes, but these are really the only NUMA v. UMA benchmarks that
are readily available. They do however demonstrate how UMA systems tend to
become more quickly and easily memory starved than NUMA systems.

I think it's fair to say that NUMA scales well, I believe that UMA will
be faster for two core and about the same for four.

The real advantage of the Intel controller is that you can customize it
for the environment, memory type, etc. And it allows other vendors to
offer custom chips as well, such as the IBM cache snooping chipset. With
AMD you have your choice of the controller on the chip. Period.

In actual practice I don't think it's a huge difference until scaling
effects become significant. For most systems that's not an issue.

--
bill davidsen
SBC/Prodigy Yorktown Heights NY data center
http://newsgroups.news.prodigy.com