PT Performance - Davec I & Davec II Tests

[Mark_Knecht]

Dave,
Yes, I'm a chip designer, or at least that's what I've been pretending for the last 23 years. Used to worked for AMD on some I/O technologies when K6 was emerging. I've done Ethernet, ATAPI, SCSI designs, and for the last 3-4 years some interesting 1394 work. Also worked for an IP startup company that sold PCI and AGP bus interface modules. That said, PTLE is responsible for me getting my guitars back out and doing something worthwhile with me time!!

Don't confuse 64-bit bus operation with 64-bit OS operation. They are totally different issues. PCI is a spec that supports 32 and 64 bit operation independent of the base width that the processor is operating at in the operating system. The North Bridge portion of the chipset reformats data between the front side bus (64-bit), the memory bus (64/128 bit for DRAM, serial interface for RDRAM) and PCI @ 32 and potentially 64 bits. This is just a physical translation. Instead of taking 2 clock cycles @ 32 bits to move 64 bits, you use just one clock cycle at 64 bits. This requires 64 bit adapter cards though, which the PT card is not. My Adaptec SCSI controllers are. (39160 Ultra SCSI-3) Fortunately, 32-bit and 64-bit cards can coexist on the same bus. 64-bit PCI buses can run both 64-bit and 32-bit operations.

With two separate PCI buses, the chipset can make data move on both buses at the same time. DMA controllers on each bus can access system memory at the same time and the system performs better. Additionally, the are often improved noise margins in these sorts of machines, which makes them more reliable.

I'm running NT 4.0 in the machine in question. It's a department server with a single 1GHZ PIII, but the MB is dual processor capable, and I built NT with d.P support, so I can drop in a second processor in the future if I need to.

Hope you found this interesting. If you want to know more, you or anyone else can drop me an email and we'll keep these threads a little cleaner. I'm happy to talk about these things with anyone who's interested.

Mark

[davec]

Mark,

I couldn't find the white papers on the 1400 only the 1400SC which has the SCSI onboard.

Great! I was little confused about you saying 64 bit PCI channels. So DELL is simply doubling the bus speed on two of it's slot. However, the processor can only process 32 bit registers so what is the true gain in a 64 bit bus? Yes you can move data quicker to and from the PCI bus; however, the processor can only process 32 bits of data at a time. The Northbridge is only serving as a buffer/xlator right?

So in a single cycle you're still only processing 32 bits of data, right? The advantage I see is the ability to do an equivalent of a read-ahead as the NB (northbridge) would contain the next 32 bit data slice... interesting.

Man we have an AMD engineer with 23 years of experience on board, do you think we can get Digi to move on the AMD certification? You would definately be our champion.


------------------
--D

[This message has been edited by davec (edited April 24, 2001).]

[This message has been edited by davec (edited April 25, 2001).]

[Mark_Knecht]

Dave,
1st - an ex-AMD engineer! Although I love the company and the processors, I only can speak from past experience about AMD and I do not pretend to represent any of the company's views, nor those of AMD's partners. I would only say from experience that I'm confident that the products are good.

2nd - I think that Digi is handling the compatibility problem EXACTLY right. I do not beleive for one second that there are any problems with the AMD processors. If there were Intel would be killing them with bad press. It ain't so. The issues are with the chipsets, or with MoBo designs using the processors and chipsets. Each MoBo has it's own issues, and these need to be looked at individually by those that have time, money and job description to do so. (Now, if Digi wanted to set me up with a few systems to qualify, I'd be really happy talk about it! I only live about 30 minutes from their office...) ;-)

3rd - while you're _mostly_ correct about 32-bit registers, I think you may be missing the point about the different clock speeds in the system. Please remember that inside the processor, it's processing 32-bit registers at whatever speed your processor runs at. So generally speaking, if I have a 1GHz processor, then I process 1 billion registers a second. (Specualtive exectution effects this heavily...) The processor has the job of deciding if the instruction requires access outside, to memory or a bus, like PCI or AGP. If it does it puts it out on the front side bus. (FSB) The FSB on all of these processors (to the best of my knowledge) is 64-bits wide, and is running at the FSB frequency you purchased (Intel = 100/133, AMD = 100/133/200/266) The North Bridge then routes the data to memory at 64-bits, AGP at 32 bits, or PCI at 32 or 64 bits. when you think of it this way, it's actually easier to do 64-bit PCI than 32-bit.

In the Dell North Bridge configuration, there are two separate 64-bit PCI buses coming from the North Bridge. (I hope you'll forgive my simplifying this. I think the machine probably has multiple North Bridge ala the older Pentium Pro designs, or a split North Bridge...) These are available on the MoBo as slots 1-3 and slots 4-6.

4th - AND MOST IMPORTANT - The predominant information moving on the PCI bus is data which is being DMA'ed into and out of memory by PCI controllers. When recording, our PT hardware is running DMA write operations into memory and the processor is setting up DMA read operations by the audio disk controller. When playing back, it's running the same sort of DMA operations in the other direction. Since memory is 64-bits wide (quiet you Rambus supporters!) 64-bit PCI actually make a huge ammount of sense.

Isn't this stuff fun?!?

Mark

[This message has been edited by Mark_Knecht (edited April 24, 2001).]

[This message has been edited by Mark_Knecht (edited April 24, 2001).]

[znelson]

Well you guys are gonna hate me....

First let me give you my specs.

Athlon 1.33Ghz CPU
GigaByte GA-7DX MB
256M DDR 266Mhz
Audio drive is 18Gig Western Digital SCSI2

I successfully had 24 tracks with compression, 4band EQ, slap delay, medium delay, and long delays.. and 16 aux tracks each with the same effects. It kinda hung the system because the SCSI drive was trying to write so many files at once but it recorded them without a hitch, just took a little time to get started and to finish. The only weird thing I noticed was my PCI meter was 100% the whole time even with nothing going on. Is that normal? The other thing I tried was deleting all aux channels and seeing how many stereo Waves RVerb's I could run on the master fader. The answer? Just 1.. it's such a CPU hog. But even with 120 regular effects going and the 1 instance of stereo RVerb, I had about 90% CPU left.

YAY!

[da BaSsTaRd!]

HOLY CRAP!

[JPS]

That's amazing!!! Congradulations on getting your 1.3 ghz GA-7DX up and screaming!!! It totally blows away my 1.2 ghz Athalon. All I could get was a "wimmpy" 24 tracks + 6 aux on the "Davec torture test". I'm sure his cruel mind is already devising a new test for the more powerful athalons. Keep us informed on the platform stability and post often on how its working out. I'm sure there will be alot of interest.

[davec]

That's a smokin box!! I've added you specs to the main post. You are by far leaps ahead of the nearest box which also is a GA but with an Athlon 1.2Mhz. The difference I see is the SCSI-2 audio you have. I'm not saying that's why your rig is a screamer, it could be a combination of things in your BIOS, your 266Mhz DDR RAM (which is double the speed of what most people have) along with the SCSI-2. SCSI-2 is definately a hot setup though.

Did you make any special changes in your BIOS settings?

If someone else out there has a similar rig, we'd all love to see a comparison. For proper triangulation (process of discovery) we actually need at least 3 samples ( much like pin-pointing the epicentre of an earthquake).

Finally, let me be the 3rd person to congratulate you on your amazing numbers.

I'm thinking of posting the test in the 001 Mac forum to compare platforms. This should make deciding on which platform to run on much easier. ....ah the Age of Empires is reborn.


------------------
--D

[This message has been edited by davec (edited April 25, 2001).]

[crs117]

Znelson,

killer system first of all. But i bet you can fix your pci meter from being totally pegged out. I am guessing its your scsi card thats pegging out your pci bus. I dont think this should effect your harddrive performance since scsi 2 is way faster then the amount of data needed to record 24 tracks at a time. Try lowering the data transfer rate of your scsi card to maybe maybe 30 or so megs a second (20 should be more then enough to record 24 tracks of 24 bit audio at 44 kil).

Perhaps this will make your rig even faster.

Good luck

Christian

[davec]

Originally posted by Mark_Knecht:
Dave,

2nd - I think that Digi is handling the compatibility problem EXACTLY right. I do not beleive for one second that there are any problems with the AMD processors. If there were Intel would be killing them with bad press. It ain't so. The issues are with the chipsets, or with MoBo designs using the processors and chipsets.

Well sure, but does Digi recommend MOBO's?? I think the specs are just CPU and the 815 chipset, or so says the compatibility posting. I've followed the spec to the "T", yet my rig is inferior to others, to which at this point there isn;t enough info whether it's the CPU or MOBO. If digi community is reporting exceptional results with a CPU+MOBO combo, it's only fair to let the rest of us know, and be able to get support for it, right? That was my point on the certification issue.

3rd - while you're _mostly_ correct about 32-bit registers, I think you may be missing the point about the different clock speeds in the system. Please remember that inside the processor, it's processing 32-bit registers at whatever speed your processor runs at. So generally speaking, if I have a 1GHz processor, then I process 1 billion registers a second. (Specualtive exectution effects this heavily...) The processor has the job of deciding if the instruction requires access outside, to memory or a bus, like PCI or AGP. If it does it puts it out on the front side bus. (FSB) The FSB on all of these processors (to the best of my knowledge) is 64-bits wide, and is running at the FSB frequency you purchased (Intel = 100/133, AMD = 100/133/200/266) The North Bridge then routes the data to memory at 64-bits, AGP at 32 bits, or PCI at 32 or 64 bits. when you think of it this way, it's actually easier to do 64-bit PCI than 32-bit.

I don't think I missed the point. I did say there would be enourmous gains didn't I??

Quote: "On the seperate pci bus issue, there definately would be enromous gains"

Further, with SMP based MOBO's, the 64 bit buss would come in extra handy which had escaped my original thoughts on the subject. Say with a quad box, in single cycle, 256 bits of data could be processed. Which was probably the idea behind the 64 bit buss. Sure it improves single processor use; but ultimately, it's in the SMP boxes that the greatest gains would occur.

4th - AND MOST IMPORTANT - The predominant information moving on the PCI bus is data
which is being DMA'ed into and out of memory by PCI controllers. When recording, our PT hardware is running DMA write operations into memory and the processor is setting up DMA read operations by the audio disk controller. When playing back, it's running the same sort of DMA operations in the other direction. Since memory is 64-bits wide (quiet you Rambus supporters!) 64-bit PCI actually make a huge ammount of sense.

Isn't this stuff fun?!?

Yes, again, I never disputed that the 64bit PCI busses would be useless. But my point is the processor must execute 2 cycles to read or write 64bit info. It works yes, but there is an inefficiency as compared to a CPU which can handle the 64bit registers. Your basically taking a 64bit piece of data, breaking it in half, feeding it to the CPU, which then puts it back together again so that it can make sense. IMHO, I consider that inefficient. That's my only point. I know this is only a transitional stage until we actually have 64bit CPU's/OSes running in our boxes at home. And again, SMP mobo's would greatly benefit from 64bit busses.

Yes, this is fun


------------------
--D

[This message has been edited by davec (edited April 25, 2001).]

[znelson]

Let's see, in the bios I disabled all onboard sound chips, turned the ECC features of my DDR ram on, and that's about it. Tonight I'll ramp the scsi controller down from 40M/sec to 20M/sec and hopefully that will calm the PCI bus down. Oh I forgot to mention I'm running WinME with no optional features installed (not even calculator). The MB itself is simply set to 133Mhz bus, and it only has the Digi001 PCI card, a Tekram SCSI controller, and a Creative Labs Annihilator Pro AGP video card. That's about it, pretty simple, I guess the simplicity keeps things running so fast. Oh I was getting a bunch of 9093 errors until I enabled DMA on my IDE system drive and cd-rw. It's weird how that affects you even when it's your system drive and you're no recording to it..