Dual CPU...........which one?

Sunshine · Post by **Sunshine** » Mon Apr 19, 2004 7:58 am

Hi guys!

I am about to buy a Dual CPU system, but don't know exactly which road to take. I'm clearly lacking experience whith such Duals.... in fact I've never built such a system on my own. Personally I consider the Dual Opterons as still too expensive....so the candidates are....

Dual Xeon whith a 875 chipset:
http://usa.asus.com/products/server/srv ... erview.htm

Dual Xeon whith 7505 chiset:
http://usa.asus.com/products/server/srv ... erview.htm

Dual Athlon whith the 760MPX chipset:
http://www.asuscom.de/products/mb/socke ... erview.htm

Has anyone got one of those mentioned systems? Are there any other options? What are your experiences?

Further, I have a question...
According to my knowledge it is preferable to chose Standard PC mode over ACPI, since the IRQs can be assigned individually to each card. But whith a Dual system you have to choose "ACPI Multi Processor" as the installation mode. Is there any other way to change IRQ assignments whith such a dual system?

/Bernhard

arela · Post by **arela** » Mon Apr 19, 2004 1:00 pm

use acpi, standard pc is from stoneage (2 years ago)
Think most modern MB benefitts from acpi mode.

sorry, no dual cpu effort here yet
(have quatro in 4 dif. cases

)

valis · Post by **valis** » Mon Apr 19, 2004 7:26 pm

Using an (older) i860 dual Xeon system here just fine in ACPI mode. Note that my SCSI controllers sit on the 64bit PCI bus and NOT on the 32bit bus (which means that the audio cards pretty much have it to themselves).

Skip the AMD-MPX chipset completely. I waited about 7 months for the MP/MPX chipsets to get stable and while 1-2 people seemed able to get them to work with dsp cards with much tweaking I'd not suggest it for most people. AMD actually stopped making chipsets so these are also tremendously outdated. The current best option for dual cpu AMD is the Nforce3 with its Dual Opteron support, but since there's no data on how well this works with Creamware cards beware you may be the one to research this if you go this route.

As for dual Xeon, notice that most e7505 motherboards (and most dual Xeon in general) offer PCI-X for most/all of their PCI slots which means that you CANNOT use Creamware cards in these slots.

The Asus with the i875 chipset would then seem to be the only choice from the options you have listed.

Also I might add that with Cubase SX/Nuendo I experienced so many instabilities due to load balancing the sequencer and my audio cards that I finally went back to running Logic Audio 5.51. Another interesting factoid here is that I get *better* performance from Logic (it only supports 1 cpu on PC) than I did from SX with BOTH Cpu's enabled. This has to do with the way the audio engines process data and outside of a few rare cases with a single massive vsti I have found this to be pretty much true.

Consider very carefully if you choose to go the Dual cpu route. I have it & love it (no slow SFP redraws here) but I'm not sure I would recommend it for everyone, especially since I don't think you get the performance improvement to justify the cost in most cases. (note that I do a lot of GRAPHICS work which is where I *do* get the benefit).

Counterparts · Post by **Counterparts** » Tue Apr 20, 2004 2:57 am

Hi Sunshine

I've asked somebody at work who a) has built a dual-athlon systrem and b) really knows his stuff.

This is his take: (my questions, his answers)

"Which chipset are you using?"

AMD 760MPX (Tyan Tiger MPX motherboard)

"With a system such as that, AFAIK you have to install Windows as "ACPI Multi Processor". Is there a way to change/force IRQ assignments in such a system?"

You are concerned about all PCI devices sharing the same IRQ. With careful slot selection on my motherboard it is possible to get a time critical PCI card on a unique interrupt. However some slots still share IRQs and there is no way to change that. SMP systems are never configured by windows so that all PCI devices share a single interrupt. The interrupt routing is determined by a special table in the BIOS.

That's straight from the 'horses mouth'

HTH,

Royston

Sunshine · Post by **Sunshine** » Tue Apr 20, 2004 7:55 am

On 2004-04-19 14:00, arela wrote:
use acpi, standard pc is from stoneage (2 years ago)
Think most modern MB benefitts from acpi mode.

sorry, no dual cpu effort here yet
(have quatro in 4 dif. cases )

thanks Arela, but personally I wouldn't put it that way necessarily, still standard PC is an option imo...but I will come back to you later as soon as I have gathered some information about this APIC thing.

/Bernhard

Sunshine · Post by **Sunshine** » Tue Apr 20, 2004 8:12 am

On 2004-04-20 03:57, Counterparts wrote:

AMD 760MPX (Tyan Tiger MPX motherboard)

"With a system such as that, AFAIK you have to install Windows as "ACPI Multi Processor". Is there a way to change/force IRQ assignments in such a system?"

You are concerned about all PCI devices sharing the same IRQ. With careful slot selection on my motherboard it is possible to get a time critical PCI card on a unique interrupt. However some slots still share IRQs and there is no way to change that. SMP systems are never configured by windows so that all PCI devices share a single interrupt. The interrupt routing is determined by a special table in the BIOS.

Thank you very much for your efforts Royston!
But unfortunately I've found out that there are different revisions of that chipset which all perform differently... whith one great downfall...There seems to be a problem whith the southbridge PCI bandwidth http://forums.2cpu.com/showthread.php?s ... hlight=IRQ

Concerning the IRQ assignments....
putting the cards into different slots is definitely one way to change the assignments of those IRQs. Anyway there are 6 installation modes...

-ACPI
-Standard PC (APM)

-ACPI Multiprocessor
-ACPI Uniprocessor
-MPS Multiprocessor (APM)
-MPS Uniprocessor

I might be wrong, but the MPS Multiprocessor mode might be the same as Stadard PC whith two processors. So only if you are really willing give up the APIC on your motherboard whith the extra 5 IRQs, this is an option. But I'll get back to that subject later.
Anyway thank you very much!

/Bernhard

Counterparts · Post by **Counterparts** » Tue Apr 20, 2004 8:21 am

Hi again

IIRC It's the Northbridge which manages/determines PCI bus bandwidth. The Southbridge isn't so critical.

Royston

Sunshine · Post by **Sunshine** » Tue Apr 20, 2004 8:44 am

On 2004-04-19 20:26, valis wrote:
Using an (older) i860 dual Xeon system here just fine in ACPI mode. Note that my SCSI controllers sit on the 64bit PCI bus and NOT on the 32bit bus (which means that the audio cards pretty much have it to themselves).

This is a touchy subject imo. I don't know what happens exactly when you put a 32bit/33mhz pci card into a 64-bit/66mhz slot! The existance of two PCI bussus can't be denied (althiough not advertised as such). Anyway there are several possibilities...

a) both busses work independantly. This means putting a 33mhz card into such a 66mhz slot will bring all the 66mhz pci slots into the lower clockrate of 33mhz, but leaving the other 33mhz slots as they are, whith a theoretical throughput of 133mhz/s,

b) both busses are connected to the same multiplier
In this case, the 66mhz slots are set back to 33mhz, BUT also decreasing the other standard 33mhz slots, not by decreasign the rate, but by lowering the general bandwidth of the second bus.

As for dual Xeon, notice that most e7505 motherboards (and most dual Xeon in general) offer PCI-X for most/all of their PCI slots which means that you CANNOT use Creamware cards in these slots.

It's a real pitty they don't work in those slots. I think I've heard similar things about the UAD not working with the AMD 8131 controller, which is the PCI-X controller for those opterons.

The Asus with the i875 chipset would then seem to be the only choice from the options you have listed.

I've come to the same conclusion, but my researches are not yet finished.

Also I might add that with Cubase SX/Nuendo I experienced so many instabilities due to load balancing the sequencer and my audio cards that I finally went back to running Logic Audio 5.51. Another interesting fact here is that I get *better* performance from Logic (it only supports 1 cpu on PC) than I did from SX with BOTH Cpu's enabled. This has to do with the way the audio engines process data and outside of a few rare cases with a single massive vsti I have found this to be pretty much true.

You know.... I've always thought of Steinberg having only second class programemrs at hand. There are several issues one could bring up from a programming few. All those bugs, crossfade clicks,etc. are really hard to take. Also the way they compile their windows versions into mac world. I wonder, how long it'll take them to get their software bugfree to a reasonable level.

Anyway thanks a lot for your detailed descriptions!

/Bernhard

<font size=-1>[ This Message was edited by: Sunshine on 2004-04-20 09:54 ]</font>

Sunshine · Post by **Sunshine** » Tue Apr 20, 2004 8:50 am

On 2004-04-20 09:21, Counterparts wrote:
Hi again

IIRC It's the Northbridge which manages/determines PCI bus bandwidth. The Southbridge isn't so critical.

Am I right in assuming...that for the execuations of those delay based Creamware plugs (reverbs) the northbridge is more crucial than the southbridge?

And for the number of streaming tracks achievable by your soundcard, the southbridge has the main responsibility?

/Bernhard

Counterparts · Post by **Counterparts** » Tue Apr 20, 2004 9:07 am

Sorry, I didn't get it quite right in my posting about this.

The Northbridge manages a "junction of many roads" - comms between the CPU, the RAM (FSB), the AGP bus and the Southbridge.

The Southbridge has a link to the Northbridge and handles the PCI slots, LAN, IDE channels, USB etc.

So - I think that you're correct in your thinking!

Royston

Ricardo · Post by **Ricardo** » Tue Apr 20, 2004 9:59 am

I really don't know much about this but, if you're considering dual processor MoBo, my Creamware supplier was very hesistant. I recently upgraded to Intel 875 on the recommendation that dual processor MoBos and Pulsar cards were tricky to get working properly. In fact they strongly recommended not to get a dual processor board.
Mine works fine in ACPI

valis · Post by **valis** » Tue Apr 20, 2004 3:13 pm

On 2004-04-20 09:44, Sunshine wrote:
On 2004-04-19 20:26, valis wrote:
Using an (older) i860 dual Xeon system here just fine in ACPI mode. Note that my SCSI controllers sit on the 64bit PCI bus and NOT on the 32bit bus (which means that the audio cards pretty much have it to themselves).
This is a touchy subject imo. I don't know what happens exactly when you put a 32bit/33mhz pci card into a 64-bit/66mhz slot! The existance of two PCI bussus can't be denied (althiough not advertised as such). Anyway there are several possibilities...

a) both busses work independantly. This means putting a 33mhz card into such a 66mhz slot will bring all the 66mhz pci slots into the lower clockrate of 33mhz, but leaving the other 33mhz slots as they are, whith a theoretical throughput of 133mhz/s,

b) both busses are connected to the same multiplier
In this case, the 66mhz slots are set back to 33mhz, BUT also decreasing the other standard 33mhz slots, not by decreasign the rate, but by lowering the general bandwidth of the second bus.

Well with my motherboard the 64bit/66mhz PCI bus slots are completely empty because each slot is tied via motherboard addressing lines (shared irq) to the dual onboard scsi controllers. The 64bit bus offers 266Mb/s and has the 32bit bus hanging off of it (which consumes 133Mb/s). However the way the architecture works with having a 32bit/33mhz card installed in the 64bit slots is to just downclock the 64bit slots to match the 32bit bus, thereby losing the advantages of the 64bit/66Mhz.

Also there is some errata on the i860 chipset that says that when the 64bit bus is under heavy load the 32bit bus is reduced to 90Mb/s (instead of 133) due to latency & bus contention. However this is only seen with the dual controllers in raid 7 configuration so I'm not terribly worried (I don't use them in RAID at all). Basically I've got a Pulsar1 & Pulsar2 running alongside my RME hammerfall (&multiface) just fine on my 32bit PCI bus.

On 2004-04-20 09:44, Sunshine wrote:

Also I might add that with Cubase SX/Nuendo I experienced so many instabilities due to load balancing the sequencer and my audio cards that I finally went back to running Logic Audio 5.51. Another interesting fact here is that I get *better* performance from Logic (it only supports 1 cpu on PC) than I did from SX with BOTH Cpu's enabled. This has to do with the way the audio engines process data and outside of a few rare cases with a single massive vsti I have found this to be pretty much true.
You know.... I've always thought of Steinberg having only second class programemrs at hand. There are several issues one could bring up from a programming few. All those bugs, crossfade clicks,etc. are really hard to take. Also the way they compile their windows versions into mac world. I wonder, how long it'll take them to get their software bugfree to a reasonable level.

Anyway thanks a lot for your detailed descriptions!

/Bernhard

It seems to me that Steinberg has EXCELLENT people in their R&D since they have had the ability to engineer ASIO and VST as well as introduce VSTi. In fact I was quite a big fan of Nuendo in the early 1.x days (I never was a huge fan of Cubase VST). However, about the time that VSTi support was introduced (1.5x) I noticed that Nuendo because MASSIVELY unstable on my dual cpu machine. Its unfortunate too because up until that point I was getting about 130-140% the performance in SMP mode (as compared to running in non-SMP mode where only the midi engine is load balanced onto the 2nd cpu). There was a lot of testing on cubase.net and the Nuendo forums during the 1.5x days and it seems what is happening is that during non-SMP mode Nuendo (and SX by virtue of inheriting the engine) loads all VST/DX/VSTi code onto cpu0 and all midi onto cpu1. In SMP-enabled mode VST & DX plugins seem to spread across both cpu's fine (as does midi) but VSTi's seem to stay on only 1 cpu and cause it to be unbalanced and start spiking. Tack onto this the fact that the Nuendo/Sx engine processes everything ALWAYS (regardless of whether there is input to process) and the effect that has on exaggerating the P4/denormal issues in any less-than steller plugins and things would often fall apart on me. In fact I would usually BSOD several times a day, although I suspect that was a result of throwing the Pulsar into the cpu load-balancing mix as well as the RME alone doesn't cause BSOD's (tho the rest of the 'glitchiness' will remain). Keep in mind that the RME was ASIO host and it is only with another taxing application open that I would see BSOD's (this includes attempting to do REWIRE-ing of apps with SMP enabled in Nuendo/Sx).

For some reason they seem to have never integrated the VSTi code well because I still notice the same glitches even in the 2.x products. I'm not sure if this is a programmer issue, or a problem of management & marketing insisting on features being tacked on quickly rather than integrating well with the code. I have noticed that they seem to have added quite a bit of additional functionality (keeping up with the Joneses so to speak) and yet the load balancing issues remain. I got tired of reporting the bugs ages ago and have just gone back to using Logic 5.51 PC for the time being. I'll worry about 'keeping up with the Joneses in 5 years or when I land a major contract (whichever is first

).

Anyway yw for the info. I spent 7-8 months researching these issues before I jumped in last time. A single cpu P4 would have been easier but I tend to prefer machines that will last 2-3 years vs. having a constant flurry of parts through my main machine.

siberiansun · Post by **siberiansun** » Tue Apr 20, 2004 7:49 pm

On 2004-04-19 14:00, arela wrote:
use acpi, standard pc is from stoneage (2 years ago)

yeah, i remember 2 years ago.
woke up in my cave, carved a spear, cheered a friendly "oga boga" to my neighbour.

sorry 'bout the extreme OT.

Sunshine · Post by **Sunshine** » Wed Apr 21, 2004 12:39 pm

Well, I'll post my ACPI or Stadard PC findings into the other thread:

http://www.planetz.com/forums/viewtopic ... forum=19&6

Sunshine · Post by **Sunshine** » Wed Apr 21, 2004 1:28 pm

On 2004-04-20 16:13, valis wrote:

It seems to me that Steinberg has EXCELLENT people in their R&D since they have had the ability to engineer ASIO and VST as well as introduce VSTi. In fact I was quite a big fan of Nuendo in the early 1.x days (I never was a huge fan of Cubase VST). However, about the time that VSTi support was introduced (1.5x) I noticed that Nuendo because MASSIVELY unstable on my dual cpu machine. Its unfortunate too because up until that point I was getting about 130-140% the performance in SMP mode (as compared to running in non-SMP mode where only the midi engine is load balanced onto the 2nd cpu). There was a lot of testing on cubase.net and the Nuendo forums during the 1.5x days and it seems what is happening is that during non-SMP mode Nuendo (and SX by virtue of inheriting the engine) loads all VST/DX/VSTi code onto cpu0 and all midi onto cpu1. In SMP-enabled mode VST & DX plugins seem to spread across both cpu's fine (as does midi) but VSTi's seem to stay on only 1 cpu and cause it to be unbalanced and start spiking. Tack onto this the fact that the Nuendo/Sx engine processes everything ALWAYS (regardless of whether there is input to process) and the effect that has on exaggerating the P4/denormal issues in any less-than steller plugins and things would often fall apart on me. In fact I would usually BSOD several times a day, although I suspect that was a result of throwing the Pulsar into the cpu load-balancing mix as well as the RME alone doesn't cause BSOD's (tho the rest of the 'glitchiness' will remain). Keep in mind that the RME was ASIO host and it is only with another taxing application open that I would see BSOD's (this includes attempting to do REWIRE-ing of apps with SMP enabled in Nuendo/Sx).

For some reason they seem to have never integrated the VSTi code well because I still notice the same glitches even in the 2.x products. I'm not sure if this is a programmer issue, or a problem of management & marketing insisting on features being tacked on quickly rather than integrating well with the code. I have noticed that they seem to have added quite a bit of additional functionality (keeping up with the Joneses so to speak) and yet the load balancing issues remain. I got tired of reporting the bugs ages ago and have just gone back to using Logic 5.51 PC for the time being. I'll worry about 'keeping up with the Joneses in 5 years or when I land a major contract (whichever is first ).

Oh thank you very much for the information. This sections contains many new things I've never heard of...I've read this part 4 times.

/Bernhard

Casper · Post by **Casper** » Sat May 01, 2004 5:35 pm

Me 2 !

I'm searching for the most capable motherbord meself... any good advices in general ? grtz

Casper