, I'm back as I promised with good news 1.
I finally found the source of advanced power efficiency demonstrated by Stilt's MLU and mining BIOSes. Yes, it's the lowered frequency
! The Stilt refused to disclose the details when I asked him, but I knew there's no other way...
So, how it works (my card is MBA reference):
-when the card power's up, the VRM controller reads the frequency value from the MTP, and it's initial frequency is set to 500kHz
-then, in a couple of seconds the value changes - it seems the Video BIOS finishes loading and executing and it overrides the 500kHz value with 480kHz
(any any stock BIOS) or 290kHz
-after that, the frequency can't be changed any further
, it only slightly deviates under load (it seems to me because of the input +12V rail voltage drop under load, my PSU in the lab is weak, and I couldn't test with a tough one).
Please note that even at reboot, the VRM frequency is not overridden anymore, it's like a Write-Once register at first power-up!!! D*mn ATIWinflash, I tested the reference BIOS with 290kHz and couldn't understand why there's no difference to MLU while it must be...
So, playing with the frequency (or maybe other VRM settings), don't forget to power-cycle the card after every BIOS change!!!
When I spoken about the losses in the VRM last time, I missed the inductor ferrite loss due to high current ripple at high frequency (it applies to very high-current VRMs and inductors and rarely else), and particular this loss seems to be reduced greatly with lower VRM frequency.Thinking about lowering the VRM frequency as a way to save power, I can't say it's bad, maybe we get 2 side-effects:
a) higher voltage ripple on VDDC under load,
b) higher requirement to the bulk capacity (silver alu-polymer caps in line) to prevent VDDC drop under significant load-step (undershoot) or VDDC overvoltage at load-release (overshoot).
I'll try to make some estimations and maybe release them later...2.
I've observed to effect of VRM phase shut-down. When there's no driver loaded, the VRM uses all it's 5 phases, and inly in Windows, when the driver is loaded, I could see how the card moves from 5ph-mode to 1ph-mode.
It seems to me the Dynamic Phase Control (DPC) feature seems to be turned off in MTP (see p.41 of 3565 datasheet), while the card uses only Power Saving States (see p.40), with only 1ph or 5ph options available through the driver request, while the intermediate power states could be really
useful with light load (up to 50-60A) like browsing, video playback, and even light 3D.As for me, from the hardware point of view, the VRM is designed well, and it could use all the available functions. But the software side is awful - when the desktop is totally static, I see 1ph active, but you just move the mouse - all the 5ph begin to work... What the h*ll! There's no load to the GPU, but the power state has already changed!! And vice versa, when everything is silent, you wait...wait...wait, until the card switches to 1ph. It seems to me there's a common problem with floating clocks and the power state changes, and in fact AMD is not interested in any tuning/improvement because the largest part of default users doesn't need anything but FPS in games...3.
I've observed the ZeroCore power feature (or how is it called??), when you set the display turn off timer to 1 min for example, and after a minute the display goes off, the card turns off the VRMs, completely stops the fan and the green LED on card bottom side turns on, indicating the card is off now. This feature works both with MLU and stock, and it's really cool 4.
Finally I would like to mention one more thing regarding the idle efficiency. When 290X card appeared in my desktop (Core i7-4790K), I noticed it switches the PCIe link from x1 Gen1 in idle to x16 Gen3 under load (clearly seen in GPU-Z).
I used Catalyst 14.12 and everything was fine until I moved to Catalyst 15.7. Since that time the lowest idle state became x16 (!!) Gen1. I opened a topic in AMD support but they didn't understand me and said "it's okay, nothing to worry about"
I tried to reproduce the problem during lab experiments, but I couldn't see x1 Gen1 mode with any drivers (including 14.12).
I have a specific MB here, mini-itx ZOTAC with GF9300 chipset, and maybe it doesn't support switching to x1 Gen1...
So please, guys, try to check the PCIe link params with old (14.12 or earlier) and new drivers and report back
It's very interesting, it's my fault or AMD S/W engineers had broken something in PCIe link management since 15.7...
It's all for the moment, maybe I find something new and report it here again