SETI on Dual Xeon motherboards (PC-DL), dissappointing - Due to memory bandwidth?

[jatwell]

New system description:

Well, I got in on the Dual LV 1.6 Xeon's (512K Cache) running on a PC-DL at 2.8Ghz (200x14), memory at 200Mhz, 2.5/7/3/3. I had to do the UWire/Insulate mod to get them to 1.6V.

I am having to software overclock the PC-DL by setting it to 166Mhz setting, hold down the insert key every boot, and then bumping it up with ClockGen when Windows boots (runs at startup, no big deal). It won't boot on the 200Mhz jumper setting at all, even if I lower the multiplier right before I try it.

No monitoring program correctly recognizes the new speed except for MBM and ClockGen itself. CPU-Z even reads it at 14x166, or 2324. I'm pretty sure all of this is working correctly.

Basically I have the exact same computer in a Dell 400SC with a 2.8Ghz Northwood, memory at 200Mhz, 2.5/7/3/3.

The Dell single CPU does WU's in 3:00-3:15, two at a time. I was kind of hoping for similar times on the Dual Xeon. It's doing 4 at a time, but more in the 4:10-4:25 range. I guess it's the memory sharing, but I was wondering if others were getting similar times.

Anyone else out there running this setup for SETI?

HMM, may have answered my own question. The Dell gets around 4800 in Sisoft's memory benchmark, where the Dual Xeon is only getting around 2700. I'm pretty sure when I was testing with only one CPU, the Xeon system was getting about the same as the Dell. Is this normal?

**Edit**: Forgot to mention that I have 1GB on the Dell, and only 512MB on the Dual Xeon. I did have 512MB in the Dell and got identical times with it though....

 

[Thoreau]

I've got a similar setup on a server of mine, dual 2.8 Xeons, but with a gig of ram, and also have 4 instances running at a time. It seems to take a little bit longer tha even your setup to crunch each unit but I think that has to do with the fact that it's running Linux. Takes approximately 5 and a half hours per unit. The box isn't totally dedicated to seti tho, so I'm sure the other processes that that box handles have some effect on it. Either way, the box puts out about 16-18 units per day which is better than nothing =)

 

[amdxborg]

Same setup here, dual 2.8 Xeons, but with 2Gig RAM. I also get +-18 wu a day running 4 instances...

 

[GLeeM]

Looks like you know more than I do, but do you have two sticks of ram in the PC-DL? (to enable dual channel)

It is curious that the P4 does so much better.

Is PAT enabled? (acceleration)

Nice OC! What kind of ram do you have?

How did you insulate? Or did you remove the pin?

I have similar setup, but have only moved the jumper, no other mods yet.
156x16=2496, 2/5/2/2 (OCZ Rev. 2)
CM heatpipes with a 120mm fan mounted to blow past the VRM, NB and into the heatpipes. Mid to low 40s C under full load of four Folding @ Home.

I am happy with it! But wonder what mods would get me up to.

 

[jatwell]

I just have some Buffalo CH5 PC3200 RAM, nothing special. It's running on default timings. Dual Channel mode w/ two 256MB sticks and PAT is enabled.

If you want to try the voltage mod, this is the wire I used the insulation off of: http://www.radioshack.com/prod...p;product%5Fid=278-503

I just stripped about 3/8" off of the wire and carefully put it over the C3 pin and then trimmed off the excess. Did the U-Wire mod with B2-B3 & D2-D3 and, easy 1.6V (Actually only 1.56 or so on the PC-DL). With the PC-DL undervolting a little, I don't think that will ever be a problem because some of these CPUs ship from the factory with 1.55v as default.

OH! I've read on other forums that you really don't need to insulate any pins at all with the PC-DL. If you UWire C2-C3 instead of insulating C3 (along with B2-B3 & D2-D3) it will give you 1.6V on the PC-DL as well Check Here

The UWire is a little easier and you don't have to buy anything


Mine wouldn't do very well at all without extra voltage. I know they would hit 2.1, but not a whole lot more. The voltage really helped.

Check the FAQ here: http://www.datamine.tk/

That's where I got most of my info


I'm pretty much convinced the sharing of the memory bandwidth is the killer. Man, I loved RC5 because you never had to worry about memory subsystems. It was all pure CPU Mhz crunching power