QMD/Sonoma experiment.

[muddocktor]

Since I bought a new Sonoma 1.6/533 proc for my P-M rig, I've been doing some testing and experimenting with running it at high fsb speeds, which I couldn't do with my old Dothan 1.7/400. I've now got it at a 240 fsb speed with 2-2-2-5 timings on the ram, which gives a Sandra buffered mem bandwidth of just under 5500 MB/s on both integer and floating point. Since ChasR had stated that QMD wu's didn't do as well on a P-M as they do on a P4 because of the low memory bandwidth, I decided to give it a test to see what they would do on a P-M machine with good mem bandwidth. I tranferred a QMD wu over from my main rig and so far it's done about 10 more frames than when I tranferred it and it's looking pretty darn good. This machine is presently running at 240 X 11 for 2640 MHz (still at stock vcore too :D ) and it's folding at a 425 point/day rate or 161.3 pts/day/GHz. That compares pretty well with my Northwood at 3300 MHz, which folds 1 client running at around 464 points/day or 140.75 pts/day/GHz..

What does this prove? That when Stanford releases the next version client that is hopefully updated to where P-M silicon can get the QMD work without having to manually tranfer them, they will be able to do some really good crunching on them if you can get decent memory bandwidth. With the general shortage of the p147x/p1481 wu's, this makes me hope that Stanford releases the next client soon so I wouldn't have to manually feed QMD's to my P-M rig.

I also plan to be clocking the fsb up some more on this rig too, because I know that the proc isn't even near to being maxxed out yet and I tested my ram to 250 MHz yesterday with no errors in memtest86 at 2-2-2-5 timings. I will see if the pts/day/GHz goes up even more as I have more available memory bandwdth.

For those that are wondering, the ram is 2 X 256 of old skool Corsair XMS3200 C2 BH5, running at 3.3v, using a ddr booster. I also have an X800XT-PE in it and it's all running with a 350 watt Enlight psu, which is rock solid with this setup.

[WarriorII]

Sweet.

Very nice setup.

Keep -a-'Clock'n !

:cool:

[ChasR]

Good news indeed. Long term folding costs should be far lower than a P4 with the reduced current demand of the Sonoma. What does it cost to build one of these from scratch (motherboard, cpu, socket adapter,ram,power supply)?

[Audioaficionado]

http://www.xpcgear.com/asusct479.html $42.49

http://www.xpcgear.com/asusp4p800vm.html $88.99

http://www.xpcgear.com/pentiumm730.html $205.99

So far $337.47

http://www.xpcgear.com/xqpackblknw.html $79.99
or just get a better PS for the same $$
and set it up in an open box as a layer.

2x512 memory ~ $100

Hard drive ~ $50

DVD ~ $30 or CD ~ $15

So I'd say $350 to $600 depending on what's in your spare parts bin ;)

[muddocktor]

As you well know, ChasR, the secret to this is good memory performance, so you won't be able to go with value ram. I'm thinking you will have to have ram that can run 230-250 fsb with low latencies, so it's going to run the cost up somewhat. Here's what I have tied up in this rig:

mobo - P4C800E-Dlx, new on eBay, $130
proc - Sonoma 730, used, pull on eBay, $129
CT-479 - ZZF, $48
Zalman CPNS7000-Cu - ZZF, $32


The ram, hard drive and case/psu were all left over from a previous folding rig. The vid card is an X800XT-PE I got on sale from ZZF for $229.00 a couple of months ago and obvisously it's power isn't needed for a folding only rig. A $20 vid card should do just fine instead. After I do some benching and playing with it, I will most probably pull out the vid card and put it in my main rig, which I'm also thinking of upgrading.

You can trim costs down a bit by going with a P4P800SE instead of my board, but I believe the bios doesn't let you change multi's like mine. You can also probably go with the stock CT-479 hsf with maybe an upgraded fan on it too. For memory, look for something that will run tight timings with at least a 230 fsb speed to optimize your memory bandwidth. These Asus boards only go to 2.85v vdimm, so either a vdimm mod needs to be done or a ddr booster used if you go with winbond based memory to get these high fsb speeds with tight timings.

EDIT: You don't need a high power psu to run these machines, especially if it's a folding only rig. I'm presently powering this machine with that 350 watt Enlight psu and if you make a folding layer you could probably get by with a 225-250 watt, good quality psu.

[AlabamaCajun]

That's awesome Mud! I have to agree that a good 350-400 watt psu will serve well for these rigs. I've used the enight 350 with the P4 rig and the 9800XT but it generated a lot of heat while gaming. While folding the drives are sipping power, the 3D engine in the GPU is off line with mostly the CPU and memmory eating watts you are running at 120-160 watts. You still need 150 overhead to remain stable due to the high surges. the other 50 or so watts are lost to heat! The PSU guys in the other threads would slam us for talking this but you have proven it. I've pulled my 9800XT while to just fold for now and I'm running at 20 watts less then with the card just in the system.

[walaka7]

Nice Mudd! Good to see you got a sweet one :) And you answered some questions I had with the dothan/sonoma with some serious bandwidth and QMD's.

Now its time to break out with the vmods and see if that little ge will hit 3.0 ghz :D

[ChasR]

I'll build one of these Sonoma things before long. I have one similar to sig rig #1 to complete first. Since I have the motherboard (DFI 875P-T) and 1 GB of crucial ballistics ram, I think I should go ahead and finish it before spending more money on a different project. It'll be a power hog for sure but 160ppd/GHz on QMDs at 3.8 GHz (640 @ 240x16) is a lot of ppd for the money to finish it.

[muddocktor]

I've also noticed that the p147x/p1481 series also seem to be getting a production boost from the increased memory bandwidth too. Wally says his Dothan 1.7/400 at 2805 is folding roughly 900 points/day rate on those wu's, for a rate of 320-321 pts/day/GHz, but my proc is doing these at roughly the same rate as his, but at 2640 MHz instead. That is giving me a rate of 341 pts/day/GHz. Since there are no real architectural differences between the 400 fsb and 533 fsb P-M, I can only attribute the gain to higher memory bandwidth. It's not isolated to just this particular wu either as I've done a few of these with it and the points/day is staying roughly the same. Here's a screenshot of it's present folding rate on this wu:

[ChasR]

I'm thinking it could be a difference in the effective latency of the ram? That would be my guess. They're interrelated in that lower latency ram results in higher memory bandwidth on the same width data path. Hypothetically, if the p1481 simulation requires the transfer 4000MB/sec to and from memory for the cpu to work at 100%, having the ability to transfer 5000MB/sec won't allow the cpu to work any faster. If your cpu has to wait only two clock clycles on a read request but Wally's waits 3, yours will get more work done dispite his higher FSB.

This is probably more or less quibling over terminology. I need to do some testing on the interrelationship of FSB and latency on bandwidth.

I've not read up on the Sonoma. Is it a die shrink of the Dothan or were changes made? Any shortening of the cache pipeline could have dramatic effects, especially with a 2 MB cache.

Edit:

Isn't the sonoma dual channel ram capable while the dothan is single channel? Or does the Asus socket adapter provide dual channel support for the Dothan.

[Audioaficionado]

I've not read up on the Sonoma. Is it a die shrink of the Dothan or were changes made? Any shortening of the cache pipeline could have dramatic effects, especially with a 2 MB cache.

Edit:

Isn't the sonoma dual channel ram capable while the dothan is single channel? Or does the Asus socket adapter provide dual channel support for the Dothan.It's 90nm so it might be a die shrink. As for dual channel, I think it's the chipset that provides that for DDR2 the newer dedicated Dothan mobos. In the case of the Sonoma I don't know. The CT-479 will support DC even with the 400MHz Dothan as it's on an Asus DC 478 mobos.

[muddocktor]

The change to Sonoma was very minor and the only obvious change was the addition of execute disable bit support, besides the obvious increase in fsb speed to 133. They must have also made some changes internally that they don't talk about because the original dothan procs (400 fsb versions) don't overclock well on the fsb speed at all, with most of them topping out in the 160-166 fsb range. Sonoma on the other hand, will handle high fsb speeds well. I see in Sjaak's thread on Dothan to the desktop that at least 1 person is running a 275 fsb speed. BTW, Wally has an original Dothan at 2805 MHz, which works out to a 165 fsb for that 1.7/400. Even if he is running the memory at 4:5 fsb/mem ratio, his ram is still running slower than mine at 1:1.

Talking about memory latency, you might be on to something there. I know that it's a synthetic benchmark, but I was surprised at the memory latency numbers that Everest Home Edition gives on that rig at it's present speed. It's showing memory latencies in the rarified atmosphere of A64 procs, with only nf3 ultra with ram running at 2-2-2-5 timings showing lower latencies (nf4 isn't in their sample). It shows a 54.6 ns memory latency for this P-M rig.

[ChasR]

I see now. Your Sonoma set up has both more bandwidth than Wally's and a much lower effective latency. Enough more bandwidth to fold QMDs efficiently and low enough latency to be better than an already magic setup for p147x/81. It makes one wonder what an A64 would do on p147x with a 2MB L2 cache.

[walaka7]

Both of my desktop dothans are running 2-2-2-5, just lower fsb for less bandwidth. one is a dothan and the other a sonoma. Unfortunatly the sonoma turned out to be a crappy clocker, only yielding 2.4 ghz. I was planning on building two more, but Im stalling a bit to see how are things turn out at years end. Also not getting good memory until around then. And neither of my boards have a multiplier adjust in bios, it has to be done in windows which will freeze if i do that along with a FSB push at same time :/. So i need a different board to do more playing. It would be interesting to see what kind of difference more FSB would make.

EDIT and I have dual channel running on dothan. I think only difference is the FSB/multiplier.

[ChasR]

The effective DDR latency (assuming 1:1 ram) of a rig with a 165 MHz FSB would be CAS 2 x 1/330,000,000 or .0000000006 seconds. Mustanleys effective DDR latency is CAS 2 x 1/480,000,000 or .0000000004 seconds. So even though both are running CAS 2, Mustanley's rig spends 33% less time waiting to act on a command.

[walaka7]

gotcha

Now on the subject of the A64. Is the cache latency around the same as the dothan/sonoma?

[ChasR]

I beleive it may be slightly lower.