Q9450, GA-P35-DS3L (BIOS 8b), 2x2GB DDR2-800, 2P7K500, XFX 8600GT, S12 550 Energy+, Accelero S1 passive, Ninja RevB TR bolt-thru passive, 2x120mm D12SM-12 case fans, Lian Li A05B, WinXP MCE.

After update to the latest BIOS, it auto-detected the Q9450 "normal" VID to be 1.2375v, which seemed high. The initial temps also seemed high with idle at 58C below Tj and load 36C below. I'm guessing that Tj is 95C as per RealTemp. CPU-z measured cpu voltage around 1.22 and change.

After successful Memtest86+ I shifted to manual BIOS settings, dropping the DDR2 to 667 for 1:1 with the 333 fsb. Speedstep is disabled. At stock the Q9450 has 8 multiplier for 2.66GHz. I began lowering VID and running Prime95 on all four cores for stress testing. Last night it successfully ran 8.5 hours at 1.100 VID with cpu-z measuring 1.04v Vdroop. Max load temps were 45C below Tj and fairly even across all four cores. At idle they were 68C below Tj. That's a 10C drop from temps resulting from the BIOS-set VID with no apparent loss in stability.

I am wondering if the MB is auto-detecting the correct "normal" VID? Are other boards setting VID at 1.2375? RMClock also detected 1.2375v for the 8 multiplier "correct" setting and will only allow dropping VID to 1.15v. I think RMClock gets its optional settings from the Speedstep settings, so I infer that at the 6 mulitplier, SpeedStep could use a minimum 1.15 VID and at the stock 8 it could use a maximum of 1.2375, that is if the MB is capable.

I left the house this morning with VID set at 1.05 (load Vdroop 0.993) with Prime95 cranked up. I am beginning to think that the real fun with this cpu will be on under-volting rather than overclocking. It would be great if RMClock would allow VID's to drop much lower than 1.15 when multipliers are stepped down. Then one could set some interesting performance-on-demand profiles. As of now, at stock fsb, this cpu is just sipping power.

As a comparison I tested a Q6600 G0 in a GA-G33M-S2 and it was Prime95 stable at 1.05V in the BIOS which equated to 0.976V under load.
It’ll be interesting to see how low these 45nm Quads can go.

Today 12 3/4 hours at 1.05 VID no errors. On my third step down from there I am currently Prime95 testing at 1.000 VID 0.960 Vcore idle (6x) and 0.944 load (8x). Prior 2 steps tested 30min Prime95. Load temps now settling at 53C to Tj.

edit 0.994 to 0.944

It ran Prime95 stable for 2 hours at VID 1.000 and I ended the test to approach the limit from the other direction. With VID set to 0.975 Core3 generated a P95 error after 2 minutes. Reset VID to 0.98125 (0.928 Vcore load) and Core1 generated a P95 error after 20 minutes. Reset VID to 0.9875 (0.944 Vcore load) all night but 2 cores generated errors, the first after one hour. Now running long-term test at 0.99375 VID (0.960 Vcore idle, 0.944 Vcore load).

The cpu-z Vcore readings are not changing much at VID's below 1.000. It might be that this budget MB is not able to control Vcore at these low values. I believe the "ultra durable" line of P35 boards (e.g. the DS3) from Gigabyte has a six-channel CPU voltage regulator which may allow further undervolting.

Did you calibrate RealTemp?

It’s a shame you don’t have a power meter to measure how much the lowered voltage affects power consumption. I have a Q9300 arriving this week and will post results in a new thread.

I have not calibrated Realtemp...but that doesnt effect load temps..and I'm looking at delta from Tjunction anyway. I have a killawatt meter...will check power draw across a range of VID tonight and post results.

Nice one.

The system was Prime95 stable for 11 1/2 hours today at 0.99375 VID (0.960 Vcore idle, 0.944 Vcore load), so I'll call that the basement. I didn't have time tonight to run a full suite of VID's with the killawatt, but I did take the end points: 1.225 VID pulls 87 watts idle and 153 watts load and 0.99375 VID pulls 77 watts idle and 115 watts load. That seems like a significant drop for identical performance.

I still don't have a consensus on out-of-the-box VID for the Q9450, it's likely MB and BIOS dependent. The 1.2375 that the DS3L gives me is higher than some of the numbers casually reported over at some of the OCing forums, so I am guessing that 1.22 is "stock" VID.

I still have to clean up the wiring to complete the build, then I will post temps and power draw with Realtemp calibrated for some additional data points. My goal is to set some profiles in the BIOS and with RMClock. This rig is a gift for one of my gamer sons, so I'll have to do some OCing so he has a chance against the E4300/8800GT rig.

edit: for the power readings I had re-enabled speedstep, so idle wattage is at 6x and load is at 8x. Speedstep does not adjust manual VID.

Can you confirm, when you set voltage manually and enable speedstep, it just changes multiplier?

I believe it will lower the multiplier but the board however will not lower the voltage anymore

I confirmed that both SpeedStep and RMClock can only change the multiplier once voltage is set manually in BIOS on the DS3L. And since my lowest stable voltage is lower than the SpeedStep voltage at all multipliers (6x 7x and 8x), that's a good thing.

This might be another useful feature of DFI's "CPU VID Special Control". I don't know if it works for undervolting too.

Robert

My E8200 stock voltage in the mobo is 1.225v which isn't much less than your quads!

I've currently got mine running at 1.1125v @ 3.2Ghz, can't get it any lower than that, but hey... its running 20% faster than stock and should use approx the same wattage due to being undervolted

I'm using a Gigabyte X38 DS4 mobo, and as with your motherboard, once you've manually set the voltage it only changes the multiplier, not sure what it drops it to on stock 6x multiplier, I'm guessing somewhere in the region of 1.1v? Which is around what it is now anyways

I'm thinking the 6x VID from SpeedStep is 1.15, so you are also better off with manual undervolting across all multipliers. Tonight I plan to install RMClock v2.35 which has Wolfdale/Yorkfield support. I couldn't set a P-state pair below 6x / 1.15v with version 2.30.

Hadn't heard of RMClock before, checked out their website but apparently it doesn't work on Vista x64 properly due to the driver file not being signed oh well

6x is probably 1,0875V