Dual CPU vs Faster Single

[NM]

On Sun, 22 Jun 2003 02:07:25 -0400, JK wrote:
Opteron has a much larger cache, and Opteron motherboards may be quite expensive for a while. If AMD wants to keep the the low end dual processor market, they might still continue to produce Athlon MP chips. The big question seems to be whether they would have Athlon 64 MP chips, or just 32 bit Athlon MP chips.

It's very very unlikely; consider this: Opteron is the best MP processor
out there thanks to its Numa architecture afforded by its 3 HT links.
Athlon64 will only have one, so 1- it's going to suck performance wise
(will have to use a shared bus), and 2- the chipset would have to handle
what is now done in the Opteron chip. In the end it will not be cheaper,
and will have sucky performances.

[Evgenij Barsukov]

NM wrote: On Sun, 22 Jun 2003 02:07:25 -0400, JK wrote: Opteron has a much larger cache, and Opteron motherboards may be quite expensive for a while. If AMD wants to keep the the low end dual processor market, they might still continue to produce Athlon MP chips. The big question seems to be whether they would have Athlon 64 MP chips, or just 32 bit Athlon MP chips. It's very very unlikely; consider this: Opteron is the best MP processor out there thanks to its Numa architecture afforded by its 3 HT links. Athlon64 will only have one, so 1- it's going to suck performance wise (will have to use a shared bus), and 2- the chipset would have to handle what is now done in the Opteron chip. In the end it will not be cheaper, and will have sucky performances.

Then why the heck did they make it?

Regards,
Evgenij
--

__________________________________________________
*science&fiction*free programs*fine art*phylosophy:
http://sudy_zhenja.tripod.com
----------remove hate_spam to answer--------------

[Douglas Bollinger]

On Fri, 27 Jun 2003 16:25:10 -0500, Evgenij Barsukov wrote:
NM wrote: It's very very unlikely; consider this: Opteron is the best MP processor out there thanks to its Numa architecture afforded by its 3 HT links. Athlon64 will only have one, so 1- it's going to suck performance wise (will have to use a shared bus), and 2- the chipset would have to handle what is now done in the Opteron chip. In the end it will not be cheaper, and will have sucky performances. Then why the heck did they make it?

Why, that's easy, the Athlon64 is for single CPU systems.

Very soon, all the CPU's made by AMD will use the Opteron core, although
each product line will be tweaked for each market and usage.

--
"Well, you see, it's such a transitional creature. It's a piss-poor
reptile and not very much of a bird."
- Melvin Konner, from "The Tangled Wing", quoting a zoologist who has
studied the archeopteryz and found it "very much like people"

[NM]

On Fri, 27 Jun 2003 22:46:31 -0400, Keith R. Williams wrote:
It's very very unlikely; consider this: Opteron is the best MP processor out there thanks to its Numa architecture afforded by its 3 HT links. I'd like you to back up this assertion. It's an interesting point, but even AMD doesn't propose that they're NUMA (though the Opteron obviously is). More to the point, why do you belive NUMA is inherently better than coherent memory systems?

In a Numa system, each processor accesses its memory directly, whereas
otherwise there is concurrent access and RAM bandwidth is shared between
the two. The drawback with Numa is when accessing another's CPU memory
from yet another, but apparently hypertransport renders the hit painless
enough.
Athlon64 will only have one, so 1- it's going to suck performance wise (will have to use a shared bus), "Shared bus"? I'd like more information here. YOur point isn't really clear.

Well Opteron has 3 HT links, which allow CPUs to talk to each other more
or less directly, the Athlon 64 only has one, so a SMP system will have
less bandwidth available.
and 2- the chipset would have to handle what is now done in the Opteron chip. In the end it will not be cheaper, and will have sucky performances. More information please! I don't see "sucky" at all.

The Opteron chip integrates many functions which are usually dedicated to
the chipset in other CPUs: each CPU has its own RAM controller, each have
3 HT links throught which they communicate with other CPUs directly. OTOH
in more standard SMP systems, the chipset includes RAM management, and has
to arbiter communication between the CPU, peripherals, RAM etc.

[Felger Carbon]

"Keith R. Williams" <krw@attglobal.net> wrote in message
news:MPG.1966c84f877dd16b98a4d9@enews.newsguy.com... (snip) More to the point, why do you belive NUMA is inherently better than coherent memory systems?

Uh, because NUMA provides much greater overall memory bandwidth?

[Keith R. Williams]

In article <pan.2003.06.28.07.02.26.936300
@paris.monnet.biz.nospam>, nico@paris.monnet.biz.nospam says... On Fri, 27 Jun 2003 22:46:31 -0400, Keith R. Williams wrote: It's very very unlikely; consider this: Opteron is the best MP processor out there thanks to its Numa architecture afforded by its 3 HT links. I'd like you to back up this assertion. It's an interesting point, but even AMD doesn't propose that they're NUMA (though the Opteron obviously is). More to the point, why do you belive NUMA is inherently better than coherent memory systems? In a Numa system, each processor accesses its memory directly, whereas otherwise there is concurrent access and RAM bandwidth is shared between the two. The drawback with Numa is when accessing another's CPU memory from yet another, but apparently hypertransport renders the hit painless enough.

Sure. Now show me that separate de-coupled memorys are more
efficient than tightly-coupled shared memories. At some point
they are since tight coupling can only exist across finite space,
but that wasn't what I was asking. The point made was that NUMA
was "better" than a tightly coupled system. ...better at exactly
"what"?
Athlon64 will only have one, so 1- it's going to suck performance wise (will have to use a shared bus), "Shared bus"? I'd like more information here. YOur point isn't really clear. Well Opteron has 3 HT links, which allow CPUs to talk to each other more or less directly,

In a ring or some other topology. This is going to be *SLOW* for
most SMP workloads. It's not going to scale like many assume.
Processor affinity is going to play a large role here, IMHO.
the Athlon 64 only has one, so a SMP system will have less bandwidth available.

Sure, lees isn't better. A64 vs. Opteron was not in any way in
my question.
and 2- the chipset would have to handle what is now done in the Opteron chip. In the end it will not be cheaper, and will have sucky performances. More information please! I don't see "sucky" at all. The Opteron chip integrates many functions which are usually dedicated to the chipset in other CPUs: each CPU has its own RAM controller, each have 3 HT links throught which they communicate with other CPUs directly. OTOH in more standard SMP systems, the chipset includes RAM management, and has to arbiter communication between the CPU, peripherals, RAM etc.

Again, I don't see "suckey" at all. No, I cannot afford to drive
a Porche to work, though my Ranger isn't "suckey" either.

I guess I've lost sight of the argument.

--
Keith

[Keith R. Williams]

In article <DTcLa.115$_c1.2@newsread3.news.pas.earthlink.net>,
fmrfne@jps.net says... "Keith R. Williams" <krw@attglobal.net> wrote in message news:MPG.1966c84f877dd16b98a4d9@enews.newsguy.com... (snip) More to the point, why do you belive NUMA is inherently better than coherent memory systems? Uh, because NUMA provides much greater overall memory bandwidth?

....and complicates cross-CPU communications.

Now Felg, please show me that NUMA is *ALWAYS* a win. Sure, NUMA
is easy, but it is here only because it is easy.

--
Keith

[NM]

On Sat, 28 Jun 2003 21:42:31 -0400, Keith R. Williams wrote:
Again, I don't see "suckey" at all. No, I cannot afford to drive a Porche to work, though my Ranger isn't "suckey" either. I guess I've lost sight of the argument.

It's going to /suck/ for SMP, relatively to the Opteron, because it
doesn't have the architecture that makes SMP very efficient on the
Opteron. If you look at the benchmarks, single processor Opteron systems
are just merely /good/, and worse than Itanium (but for a fraction of the
price ...); but as soon as you go 2 or 4 way, they surpass Itaniums easily
(and still for a fraction of the price). You won't get this advantage with
the A64, and on top of that the lack of dedicated circuitry on the chip
would make a SMP chipset more expensive, so overall it wouldn't even be
competitive WRT cost.

[NM]

On Sat, 28 Jun 2003 21:44:12 -0400, Keith R. Williams wrote:
...and complicates cross-CPU communications. Now Felg, please show me that NUMA is *ALWAYS* a win. Sure, NUMA is easy, but it is here only because it is easy.

It is easier to achieve better performance. Therefore, unless you have
unlimited $$$, NUMA is always a win.

[Rob Stow]

NM wrote: On Sat, 28 Jun 2003 21:42:31 -0400, Keith R. Williams wrote:Again, I don't see "suckey" at all. No, I cannot afford to drive a Porcheto work, though my Ranger isn't "suckey" either.I guess I've lost sight of the argument. It's going to /suck/ for SMP, relatively to the Opteron, because it doesn't have the architecture that makes SMP very efficient on the Opteron. If you look at the benchmarks, single processor Opteron systems are just merely /good/, and worse than Itanium (but for a fraction of the price ...); but as soon as you go 2 or 4 way, they surpass Itaniums easily (and still for a fraction of the price). You won't get this advantage with the A64, and on top of that the lack of dedicated circuitry on the chip would make a SMP chipset more expensive, so overall it wouldn't even be competitive WRT cost.

AMD's position is and has always been that Opteron is for
uni- *or* multi-processor systems and Athlon64 is *only* for
uniprocessor systems. *Nobody* has expressed any interest
or intention in making any kind of chipset that would let
you use an Athlon64 in a multi-proc system. *And* the
Opteron is price-competitive with the current Athlon-MPs, and
more than competitive with the Xeon and Itanic. There is no
reason from either a financial or performance perspective for
anyone to invest a dime in an effort to make Athlon64 work in
multi-proc systems.

So how the hell did this idiotic discussion of using Athlon64
in multi-proc systems start ?

The first time some idiot started talking about Athlon64 in
anything but a uniprocessor system, the first response should
have been "ROTFL", and that should have been the end of it.

[George Macdonald]

On Sun, 29 Jun 2003 14:35:05 +0200, NM <nico@paris.monnet.biz.nospam>
wrote:
On Sat, 28 Jun 2003 21:44:12 -0400, Keith R. Williams wrote: ...and complicates cross-CPU communications. Now Felg, please show me that NUMA is *ALWAYS* a win. Sure, NUMA is easy, but it is here only because it is easy.It is easier to achieve better performance. Therefore, unless you haveunlimited $$$, NUMA is always a win.

For many (some ?) applications, unless the app(s) is/are very carefully
programmed for NUMA (note AMD calls it SUMO - sufficiently uniform memory
organization), you *could* lose badly with it. Basically if two or more
intensive parallel threads need to share large data structures you're sunk!
I don't think AMD is in a very good position to convince developers to
program specially for Opteron SMP. Let's hope there aren't too many of
those "ill-behaved" apps.

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??

[NM]

On Sun, 29 Jun 2003 20:08:36 -0400, Keith R. Williams wrote:
Whoa! The market for A64 is P4, not Itanic! PLease tell me again how it will suck! Hell I could say a full-blown Power-5 will blow away any thing out there, but at what cost? Remember, time is money.

I'm not saying the A64 will suck, I'm saying it would suck as an SMP
processor. It's not designed for this.

[Keith R. Williams]

In article <pan.2003.07.02.14.54.19.266106
@paris.monnet.biz.nospam>, nico@paris.monnet.biz.nospam says... On Sun, 29 Jun 2003 20:08:36 -0400, Keith R. Williams wrote: Whoa! The market for A64 is P4, not Itanic! PLease tell me again how it will suck! Hell I could say a full-blown Power-5 will blow away any thing out there, but at what cost? Remember, time is money. I'm not saying the A64 will suck, I'm saying it would suck as an SMP processor. It's not designed for this.

We now understand each other! ;-)

--
Keith