Sounds like your on the right track .. cant wait for some more results!
GTX 670 reference PCB - voltmod - Page 5
Also I put a resistor together with current limit feedback resistor. Took it of an old 6600gt.. It's a 10k resistor. The original was 22k.. This makes the current 3 times higher. It might still protect the circuit if a coil saturates..
Now I have +20% for power%.. I did 1.3V, 1372MHz test.. the power% rises to 103% and there is no throttling. I think the throttling might have also been OCP-based... It makes no sense that the card throttles at 105% when my power target is 112%..
I'll do more testing later this evening. Here's a picture after 3Dmark11.
Also here is my current mark11 record: http://www.3dmark.com/3dm11/4863219 --> I clocked my CPU to the limit for this push..
Here is the 3dmark for this 1372MHz run: http://www.3dmark.com/3dm11/5102878 --> Graphics score is rising as it should.. To be continued =)
I do this for the fun of it, I'm not sure if I find the patience in myself to run multiple benchmarks..
I'm gonna run some tests now though, and update this post soon!
Raised core to 1399mhz, then heaven benchmark on. Screen went black, GPUv --> 0V.. OVP? Shouldn't be OCP because I set it so high....
Edited by Olper - 12/2/12 at 7:47am
Sick score for a Lynnfield! Your cooking with fire now Olper You may be right about the speed causing premature OCP issue.. speed will bring current increases quickly.. even with stock voltage. Cant wait to see you get this dialed in a bit more. You have a good shot at some nice bench scores in '01 and '03 .. and probably have enough CPU for '05 too.
Good work! - Dont forget some pics of your completed mod - to post in the Vmod essentials thread
Palit model has only 4-phase buck. MSI has 5-phase and asus has 6-phase atleast on their direct-cu II top model. More phases is an advantage to inductor currents, mosfet currents, ripple voltages and ripple currents. Because this has only 4 phases and I'm driving this card above specification, I'm afraid the inductives saturate. When an inductive saturates, it's impedance collapses very suddenly, very low. It will cause a very large current spike. It can easily burn mosfets. I could fix this by raising the switching frequency. This would lower the max inductive current and at the same time lower ripple voltages and ripple currents... The downside is that it will increase switching losses of the mosfets. I should have a method for measuring mosfet temperatures before I do this mod. I think the fets will be quite cool because I have a full cover block.
Something to do for the next week =) I'll borrow some thermocouples and insert one between a fet and the GPU block. If the temps are low, I'll raise the switching frequency.
I ran 3dmark11: http://www.3dmark.com/3dm11/5105585?key=K8XViuJC8KINtLwiWT5MMw
CPU @ 4300MHz
GPU 1385MHZ(core) / 3703MHz(mem)
Edited by Olper - 12/2/12 at 9:03am
I hope I can fix these issues.. It would be nice to have the card fly over 1400mhz
By the way, I mod the OCP by lowering resistance between RC and ILIM. Maybe I should mod by pulling down IMON... It should not make a difference.. What do you guys think? Does it matter which one I change? Datasheet says that OCP is disabled by connecting ILIM directly to RC.
So long as they are connected on the PCB.. it should not matter ; and your mod should be sufficient. I think you on to something with the inductors and your increase to Sw. Freq. Can you measure voltage at the input and output sides of the inductors to check for droop please?
*Some GPU/IC need "total" OCP mod.. and "per phase" OCP mod .. like some Volterra IC. Maybe Richtek has some sort of total & per phase mod as well I havent bothered to pour into the DS much at all though Im afraid.
I think IMON signals output current information as an analog signal. .datasheet: "IMON: 0 mV to 900 mV analog signal proportional to the output load current. VSN referenced". I think NCP5392 does not use this information at all, but it's there for other circuits to read. Now that I fiddled with ILIM, it means that the NCP5392's OCP limit is higher. Still it sends correct output current information forward. Maybe the GPU reads this signal and throttles at a limit. I'm gonna put a pulldown on this...