The Stilt

That is a low leak part, which are generally optimal for underclocking. The required voltage will be slighty higher, however the current draw is much lower. Puts less stress on the VRM too. After the estimated load-line effect, the actual stock voltages would be (roughly): 3.9GHz - 1.29875V 3.6GHz - 1.24125V 3.1GHz - 1.15625V Only beyond that it is undervolting (technically). Those frequencies and voltage should be "factory stable" with load-line close to 0%.

The software will not run on Kaveri due the hardware checks. The registers on Kaveri and Carrizo are different anyway. AFI is likely to be the last software I make so I cannot promise support for the upcoming parts. If I happen to be still around I will add support for Indus, Bald Eagle, Steppe Eagle, Mullins and Golden Eagle too.

AFI R1.02 Changes: - Fixed the issues with Trinity APUs. - Fixed the faulty PState7 frequency reading. - Added support for the new Richland APUs / NPUs ("20"-series, K-series).

Yeah, Trinity and some Richland models are broken. Will fix it in a new version, with couple of new SKUs added.

7.5€ (10.2$) per litre, excluding the handling fees and rent of the container

2D does not tell anything about the IMC stability on APUs. 3DMark06 GT-1 is the best stability test for NB/IMC/Memory, better than memtest, linpack or prime combined. Artifacts = Unstable memory System or display driver crash = Unstable IMC / NB

LSlowmotion: I'll fix that issue with 5600K asap. High leakage parts are a must for high frequencies. Otherwise you will run into a voltage wall. The best parts for high frequency have ultra high to extreme leakage and ultra low (<1980) leakage scalar. Lower the leakage scalar value, higher the silicon quality (despite the leakage). In my personal experience: 1.3375V = 7.6GHz 1.3125V = 7.8GHz 1.2875V = 8.0GHz 1.2750V = 8.2GHz(+) 1.2500V = ?.?GHz

BApm itself has no effect on Windows behavior. Does it happen at the stock GPU frequency also? Freezing during the welcome screen is usually because of a too high GPU frequency.

AFI R1.01 Changes: - Added support for FM2 NPUs (Athlon & Sempron). All FM2 socket APU/NPUs are now supported (excl. FirePro).

It works only on APUs at the moment. Athlon & Semprons are NPUs.

Good good, keep them coming Hiwa: The "PNY1" should go a bit higher...

Unfortunately this is possible on APUs only. CPUs need JTAG to extract the same information (thru the "Engineering Sandbox"). 8GHz only?! That won't do as I was thinking of ~ 8.4GHz

I've been thinking about it. Currently I am crippled due lack of a proper driver. The driver I am currently using cannot access physical memory. Therefore I cannot read APIC to determine the clock frequency. TSC and other performance counters on AMD is far too inaccurate for the purpose. I have another driver which can access the physical memory, but it is unsigned and therefore cannot work with 64-bit Win 7 or 8 without activating the test mode. I am hoping that people would show the "AFI" screen together with CPU-Z.

Once decoded, it is.

Abbreviations: - "Leakage Scalar" - A part specific reference value for leakage behavior (ILC). - "DSTT" - Dielectric Strike-Through Threshold. * All of the values are part specific and based on actual production data. The program supports all desktop (i.e FM2) Trinity & Richland APUs. NPUs are not supported at the moment. The voltage displayed in the DSTT field is not an absolute limit. It should be used for reference only. The voltage is the highest "safe" (long term) voltage level in that temperature. The voltage can be exceeded for a short periods of time, however silicon level damage can occur. Pressing "Copy Hash" button will copy the displayed hash-string to clipboard. In case the UAC is enabled (Vista / 7 / 8) please run the program as admin. AFI R1.00

Run it as admin. Is that on Windows 8 anyway?

Use 7-Zip, it is a LZMA zip.

Had some nasty flashbacks from the time when I did the ACI... Nothing has changed since those days. None of the ODMs seem to follow the design guidelines, but instead use their custom configuration. And the custom configurations are rarely working perfectly and therefore the results may vary... Tried three different boards and came across of three different configurations. None of them followed the default configuration defined by AGESA. Didn't try ASRock as I don't have one yet, but I think I managed to cover all of the possible scenarios The fix itself is very simple, but to make it all work perfectly I had to build a logic around it. The logic required around 20 times more rows than the fix itself. The BApm controls all of the Dpms (dynamic power management) on 15h APUs. There are different Dpms for NCLK (NB), SCLK (GPU), LCLK (Link), DCLK (Display), VCLK and few different Dpms for the UVD clocks too. If the BApm is disabled you will loose all of those Dpms. No Dpm (PState) change can occur after that. This includes both the clocks and the voltage. The software will configure the Dpms prior disabling the BApm in a way that there will be no performance impact. For NCLK Dpm for example it will determine the highest performance resulting PState and lock the NBPll to that frequency prior disabling the BApm. After the BApm has been disabled no frequency change can occur (other than CPUPll & GPUPll, i.e frequency). As the NCLK PState (and frequency) is tied to the GPU activity it is continously varying. In case the BApm would be disabled directly, the NCLK could lock into a low performance (frequency) state, which naturally would cause a performance impact. Notes: - In case the display driver crashes and is recovered the BApm needs to be disabled again - Running simultaneously with CPU-Z or HWMonitor might cause a system crash. - All of the changes will be overwritten by a reset - Unlocking the NCLK frequency might require a cold reset - The NCLK frequency cannot be changed in flight (e.g with PSCheck) after BApm has been disabled - In case your are already running at critical UNB voltage levels (1.55V or above), do note that disabling BApm sometimes results a 25mV increase in the GPU voltage. It is tested on Catalyst 13.9 WHQL and on 13.10 Beta 2 drivers. AMD has changed and might change again the behavior of BApm in different driver versions, so the fix might not work on future or older drivers. https://www.dropbox.com/s/22vdtd3ruwsd54s/DPT_R1.00.zip

Several people have reported that the 3DMark score is dropping after a certain clock frequency has been reached. The technical background of this issue is the "Bidirectional Application Power Management" (BApm). The performance starts to degrade when the GPU divider is dropped to 3,00 or below (equals 1266MHz or higher SCLK at 100MHz BCLK). It is because at this point the power envelope (44W on 6800K) for the GPU is reached. When the GPU exceeds it's power envelope, BApm forces the SCLK to throttle to get the power consumption back within the limits. During the throttle the SCLK frequency will oscillate between 304MHz and 1266MHz (at 100MHz BCLK). Eventhou the BApm has access to the real power consumption and temperature data they are ignored for the SCLK power limit. The BApm only weights changes made to the commanded UNB Vdd and the SCLK Dpm frequency. Since the voltage changes made from the bios are not visible to BApm, it only sees the SCLK frequency itself changing. Therefore the SCLK throttling activates always at the same point, no matter if the chip is running on air or LN2 cooling. Anyway I will post a tool to disable the throttling on Trinity and Richland APUs later today. Have to make one first, as BApm cannot be killed easily without breaking number of other things

What about if you add a "-f" switch after the command?

You should be able to hot flash it in pretty much any board with a DIP8 socketed bios. http://ra.openbios.org/~idwer/flashrom/dos/ Command "Flashrom -p internal -w biosname.xxx" In case that does not work, try: "Flashrom -p internal:laptop=this_is_not_a_laptop -w biosname.xxx".

The APUs are really not that fragile. It is usually the voltage spike (e.g. a PState change with a large voltage offset) or a motherboard VRM failing or shorting due excess load or condensation. It is naturally possible to kill the APUs with voltage. They are more fragile than the FXs for example. About half of the die is reserved for the GPU so the power density on APUs is massive when they are pushed to the limits. I would guess it is rather the substrate-die interconnects which might be failing rather than the die itself. The CPU can usually take up to 1.95V on LN2 (depending on leakage) when only one of the CUs is being loaded. The GPU can take up to 1.7V, after that it dies.

I'll post some pictures when I have rested a day or two. The "Slow Mode" is quite simple as you only need to connect a switch between PROCHOT and Vss. It was a struggle to find the soldering spot for it thou, took around 1.5 hours. I think the UD7 could do even more after the VRM switching frequency is raised. There is no public datasheet for the ISL6330 so I don't know which pin sets the fSW or what it currently is set to. Usually GB boards have a very low fSW, on 990XA-UD3 for example the VRM operates at 263kHz.

I calibrate the RdPtrInit etc values by hand. There is no difference in performance after that. The different DRAM Pll Multiplier + Divider combinations result different duty cycles, however they do not have effect on performance. Besides, my PSCs (the PCB) was damaged after running on UP4. It is not a very good idea to extract the modules when they are frozen solid into the slot... Previously DDR-2666 with 9-11-8 timings was not a issue, now around DDR-2500 is the maximum.

I know. My "efficiency" absolutely sucks.