You people maybe forgot about crazyness with all that cool F1EE pots, EVGA Classifieds and EVGA W555 boards? I will remind you correct way to bench, how it was started years ago
So back to works, last weekend I had 50L of LN2 and some G80 cards to burn away
For next tests I made wrong decision to compete with 1156 platform. YES, Foxconn socket!
CPU X3470ES (i7 870 analogue)
POT Xtremelabs.org MAGNUM LE
asus M3F
2x1G Kinston DDR3
XFX 8800GTX on green pcb + diy ln2 pot from piece of sh..t Made from copper core of Volcano7+ soldered with bronze to form "pot"
Enermax Revolution 85+ 1250W sponsored by Enermax
I've tested rig on my engineering place, so I had all that crazy devices to measure and control everything.
Tried asus remote overclock tool, with USB cable linked to host PC. It works way better that usual stupid overclock software trash from vendors.
http://www.xdevs.com/images/ln2_cel/rcasus.jpg
For temp measurement I used DMM Fluke 87V with native thermocouple installed to CPU pot, plus extra K-type TC for GPU pot. But later forget about GPU temp probe, coz it's showed temps from -170 to -50. Too little mass of pot.
Most of the time I monitored CPU pot temp, to keep temps from -65 to -80, just to be safe. Clock was at 4.9-5G with 1.57Vcore, 1.35Vtt and 1.65Vmem. I was aware already about burning foxconn 1156 sockets, so disabled two cores and HT, to lower power for CPU.
For control of DC-DC on GPU based on Primarion I used good old oscilloscope Tektronix 2246.
Pic for dram FET monitoring.
http://www.xdevs.com/images/ln2_cel/gtxmem.jpg
Voltage on videomem was set to 2.35V (after simple FB mod), and 1.5 to 1.7V for GPU. Most of tests was done with 1.6Vgpu
Card was prepared as usual for subzero benchies, and it's power units was modified in 7 places. All BGA packaged chips was insulated with bostic.
NVIO chip wasn't cooled, no even heatsink, it's NOT needed with subzero GPU cooling, power consumption of it less than 2W, and is cooled good by heavy copper poligons in PCB from cold GPU chip.
Because I highly supporting idea that all modifications and tricks must be available to every overclocker openly and free I will describe every thing which was done to card. Why so? Just think a bit, what made overclocking mainstream and popular? Hiding secret jumpers, and xtal changes on overclocking 386 CPU's? Not telling anybody results of overclocked systems? No! Only sharing will help overclocking and made this interesting. This is like sport for me. And nobody in pro sport hide abilities and things which affect result, this is cheating.
So lets start to mods.
Vgpu tuning is done by usual and wide-known FB-sense tweak with trimmer. PWM-controller on G80 cards is Primarion PX3540, documents on which are under vendors NDA. But we have short briefsheet, which is enough to do mod. Can be found by 5sec googling.
It's internal parts easy to understand.
So I used nothing new to ramp up my VGPU here, and this mod was used a lot of times by many people. Used here one trimmer with 100ohm range.
OCP
Next I had issues with OCP ,which powerdown card if volts set more than 1.6V and Vgpu clock was above 750MHz. Look again on reference shematics of PX3540 DC-Dc, in same briefsheet. Even not very familiar man will find block with pins ISENx, called "Current sense". So we need to lower that signals to fool PWM chip, to make it measuring less current that fed in real card. These ISEN's are differential, and independent for each phase. Ratio of current from sensor (here current monitoring from ESR of inductors) is set by RCR curcuit, showed in blue boxed on next pic.
How to fool current sense monitor? We need to change resistanse between ISEN's. for each phase. I made that simple - just putting in parallel extra 1.5kohm 0603 resistor on top of capacitor of that RCR-curcuit. On stock resistance on that cap pads was near 700ohm, after mod - dropped to 400 ohm. That's lower monitored current quite awesome, near 50% less I estimate.
OVP
And just to be safe, I also done OVP. Just find resistor hooked to Vmax, and lower it with 100k trimmer to ground. On stock cards it's 0603 res marked "78Х" i guess.
So high-res photo of prepared GPU power controller here:
http://www.xdevs.com/images/ln2_cel/ocp.jpg
Vmem - just standard, as was described by all G80 guides. Ramp up from 2.1V to 2.35V. My GTX used hynix drams, which clocked 1080 on stock volts. With 2.35V I've got 1215MHz on memory without problems.
For monitoring of presense of voltaged without need of DMM's or messing with wired I soldered 2-colored LED to Vmem and Vgpu, with common catode on ground. And now if card lost contact with PCI-E I see instantly that there is no voltage on GPU/Mem. Also usable to see if OCP/OVP had power down card, and not guessing if GPU had coldbug, or it's just OCP/OVP.
It's simple , but useful. Hey, EVGA engineers, make same thing to your mobos/GPU's. LEDs today are cheapo . Selling the idea for free.
Also I've added more than 10 Low-ESR polymer capacitors, which don't lose their capacitance under LN2 temps.
http://www.xdevs.com/images/ln2_cel/panas2.jpg - Panasonic yellow cap 820uF submerget to LN2
http://www.xdevs.com/images/ln2_cel/capsepc2.jpg - OSCON SEPC in LN2
Overral card looking next:
http://www.xdevs.com/images/ln2_cel/gtop.jpg
http://www.xdevs.com/images/ln2_cel/gbot.jpg
Than everything as usual LN2 oc.
Mounted stupid pot on GPU insulated and fired stuff up.
BTW, sensor ADT7473 (if i remember it's p/n correct, lazy to look on card now) measures subzero temps, but not sure how it correlate to real temps. not checked that yet.
Card heats much worse to Lynnfield CPU on 5GHz, and worked good with temps down to -55. If lower often had black-screens, of "broken glass monitor" effect, when parts of screen clone other parts of image. Forgot to photo that
Tested Vgpu up to 1.7V, no OCP/OVP issues anymore, with clocks up to 850. But to be safe tested on 810/837 Vgpu, 2GHz shader and 1215MHz mem. Bad GPU pot card heated fast, from -50 on start of aquamark, and to -15 to end of test, with 1.55-1.6Vgpu.
In a result after 6hrs of nonstop night bench I've got stable 810MHz on ROP, 1998 Shader, 1215MHz dram, and some tests, like aqua03 was done at 837-856 ROP, 2025 shader. But I have only one LN2 container to fill pots, so when I refill container I often had overheats on GPU to -10...0 and test halts.
BIOS on 8800GTX was flashed from ASUS 8800Ultra.
Results:
And photos of rig
http://www.xdevs.com/images/ln2_cel/gtx03.jpg
http://www.xdevs.com/images/ln2_cel/gtx053.jpg
http://www.xdevs.com/images/ln2_cel/gtxice.jpg
http://www.xdevs.com/images/ln2_cel/gtxrig.jpg
http://www.xdevs.com/images/ln2_cel/gtx054.jpg
http://www.xdevs.com/images/ln2_cel/gtxr.jpg
http://www.xdevs.com/images/ln2_cel/gtxsn.jpg
A lot of people on nowadays bench 8800GTX btw, because of hwbot rev3,lol. I'm on high ranks a bit, heh.
Also I've used not all LN2 on this 8800GTX, about a half was used to break some very good scores . Will show you on 23 jan.
Thanks to overclocker S_A_V for proving my ideas with useful links.
Thanks to teammate T0lsty from Xtremelabs.org for LN2 supply and help.
Thanks Enermax for awesome PSU.
As a result, 3D-bench under LN2 is much more fun that ugly 2D runs, and much more hard to perform on good level. Everything hard is nice, and bring more interest to defeat difficulties and issues. More complex - more adrenaline
And after rig was disassembled, I had much more adrenaline.
Even on two cores disabled and only two-core 3D benchies Xeon and asus mobo with foxconn socket were heavily damaged. I wasn't running more than 1.57Vcpu by DMM read. EPIC fail on asus mobo for 300$.
http://www.xdevs.com/images/i1156/lga1.jpg
http://www.xdevs.com/images/i1156/lga2.jpg
After some clean up works on socket and CPU with binocular microscope, I've put CPU to mobo and it worked. But I will not return to 1156 ocing anymore. Don't want to overclock to last pin. Enough for me, now only bi1366 Just waiting to Peter to release W555
Bookmarks