Re: Bulldozer Overclocking Guide (Performance Scaling Charts, max OCs)LN2 Results com
(This BIOS is BIOS G2, it is not available to the public, so there are some settings not available on BIOS F6E)
#3 Now go into advanced BIOS features menu and disable: C1E, K8 Cool& Quiet, C6 support, and any other features there that you donít need. Then boot into windows, if you fail to get past the windows loading screen, then please increase the voltage a bit more.
#4 get into windows and run a stress testing suite for about 30mins with every core being loaded to the max. if you pass, restart and go into the BIOS. Now you want to tune the finer things, and increase the host clock(HTT/FSB). So lower the multiplier of the CPU, and increase the host clock, something like 230mhz to 250mhz is a good point to start. At this point the CPU-NB multiplier needs to be changed, please make sure itís under 2.7ghz. At this time HT multiplier needs to be changed as well. Also your DRAM multiplier needs to be changed.
#5 Now that you have an understanding at what voltage and frequency your processor is kind of stable at, then you can fiddle with the finer knobs. Try using the DRAM EOCP for memory if you donít know how to set your timings, but make sure you always watch the CPU NB frequency, as if itís too high the system will fail to boot. You also will need to increase the CPU-NB voltage; I set to 1.3v for the heck of it, but too high can cause performance issues. If you want a high HTT/FSB you might need to bump the Northbridge voltage as well as CPU NB. DRAM Voltage can be changed as well.
#6 Get back into windows and test, if you fail to boot, at the same frequency you had before for the CPU, that could be due to the increase HTT/FSB, I had to increase vcore to keep up with lower multi and higher HTT/FSB even at the same CPU frequency.
#7 Now you are in windows, run a few benchmarks, write those numbers down. Go back into the BIOS and change up the timings, mix up or match the HT Link and CPU NB frequencies, or try them at different levels. You can try increasing the voltage or decreasing them as well. Go back into windows and run your benchmarks. If performance is higher than you are on the right track, if it is lower, then go back.
By now you can probably pull something off like this:
My Performance Findings of Finer Tuning (CPUNB, HT Link, Multi, Voltage, HTT, DRAM):
Below I will go through some charts of my results. Each benchmark was run more than once, as the results are pretty all over the place with Bulldozer, so they were averaged. I ran SuperPI 1M, wPrime 32M, CineBench 11.5 CPU, and 3DMark Vantage CPU Test 1. My results are probably lower than most have seen, but they are not for showing how BDs performs, you can just look at different reviews, they are just for internal comparison. Windows 7 was put off disc, no updates, just stock windows 7, with all drivers installed. Also not the latest BIOS was used, so it has a kind of outdated CPU Microcode and the CPU is ES. Please look at a performance review if you want to see how BD performs against another processor.
First off memory speedís impact:
So you can see that the memory speed changed alone has a nice impact on performance, there is a huge jump from 1333 to 1600, but not that big from 1600-1866. Some are thinking, that is due to lower CPUNB and HT Link speed, and plz look at CPU NB Scaling for more on that, as I raise the CPU NB alone with memory at 1866.
Other than in SuperPI we don't see any steady scaling of any of the other benchmarks. 3DMark Vantage is close, but not very good. What we did see though, is that CPU NB isn't really bottlenecking the memory speed at 1866. Also please note that we can't pull off a CPU NB speed at or above 2.8 GHz, not on air. In fact with the 8.4 GHz record set, they used a 2.4 GHz CPUNB speed. While CPUNB speed probably will greatly improve performance at a higher level due to the increase in core count, sadly either this uarch doesn't take advantage of it or it will help only under subzero results. I do think that CPU NB helps performance when slightly increased; itís probably not a coincidence that the DRAM E.IO.C.P auto sets it around 2300-2400mhz. CPU NB Speed can help performance, especially subzero, you can maybe hit 3-4ghz.
I kept the CPUNB constant, as well as the speed, and kept memory at 1866. Only thing changed was HT Link speed, and it doesn't seem to be clock limited like the CPUNB is under air/water at 2.8 GHz. For the last two results, I increase the CPU NB, once to come close to a higher HT Link, and once to match HT Link. You see the results are very inconsistent, but that the 2796 HTLink frequency seems to do well, the stock is 2600, and increasing the CPU NB a slight bit seems to help as well. So I am going to suggest, that on air cooling, you take the HT Link to about 2.6-2.8ghz, and either set CPU NB at 2600 or 2400 to match or be close to HTLink.
HT Link and CPU-NB speeds have been changed together, with a much higher memory speed. The result is miniscule at best in all benchmarks, except for Cinebench and 3DMark Vantage. We can see from this that there is a very small increase in performance having those settings close together. So yea it helps.
In the past tests I used HTLink with DRAM of 1866 and HTT of 233, well letís increase those a bit more, because HT Link is the one frequency we can change a lot, I wanted to see if it made a difference changing it with even higher busses. Here is something I find fascinating, can you see it? The performance is best when the HT Link equals the CPUNB in this round.
As we increase the HTT we see that Cinebench as well as 3DMark CPU test seems to scale almost perfectly, even with lower memory speed for some of the tests. We also see a need for higher voltage to remain stable as the same CPU frequency. Even with lower CPUNB and HT link speed for the increase HTT (multiplier limited) we see that performance in those benchmarks still increase. SuperPI seems to have no change, but SuperPI seems to like a higher multiplier.
Conclusion: HTT(FSB) needs to be raised for best performance, even if you can only use the multiplier.
Here is voltage scaling. You can see a steady rise in temperatures of load, but compared to change in idle temperatures that change is very small. That is because what is missing is current at idle, at load there is much more current than at idle, even if the frequency is the same. Some food for thought. Note that performance decreases for the most part as we go down. There is a sweet spot in there if you can find it.
Time for Optimized OC:
So now you are supposed to take those analysis of the different comparison tests, and find your most optimum OC. we put all that you just saw together, and you get this overclock:
You can copy those results for your CPU if you like.
It is a really nice looking system all together isnít it?
Higher Overclocking on Air/Water:
So that is an easy 5ghz, we were stable at 5ghz.
Here is the MAX OC ON AIR/WATER:
Now if that gives you confidence for Bulldozer OCing I donít know what will!
Stability is subjective at best, one guy might say his system is stable with 3 hours of IBT, another with 12, and another with 24. Each will tell each other either your isn't totally stable or that you over did it. Personally i think whatever you do i fine, i don't care if you call it stable, it has nothing to do with me. BUT please note that overclocking in general isn't meant to be stable. One day your system will be good, a week down and its not at the same voltage, because you ran it at 70C with 1.5v on air, and it passed IBT for 24 hours straight.
Thanks for reading, and till next time!
Also if you are looking to buy a GIGABYTE UD7 right now, GIGABYTE has a deal on Newegg for $214 after MIR
LN2 session is happening tonight, should have some type of results like max CPU clock up tomorrow, wish me luck!
GA-X58A-UD5 rev. 2.0
XMS3 2gb x 2 1600
Custom Lian-Li PC 70