Hello folks,

I hope you are doing well.

i have got my hands on this RX6900XT nitro+ from Sapphire, the card was mostly, baked in an oven from the previous owner, fortunately, the card is still alive. but I'm trying to connect the dots..

As you see on the attached picture I have removed two memory chips, which caused the memory rail not to startup at all, and also the 0.75 rail is not present, I have traced it down to the 1.8V which basically signal a powergood to the 0.75_EN rail. The strange thing with this, is that the powergood signal goes to the MPQ8633A for the 0.75_EN pin, it also goes to other places, but will keep the rest for later, the powergood from the 1.8v rail once its connected to the MPQ8633A 0.75_EN (through a resistor 0ohm) pin it acts differently, it acts as if its an PWM signal, while it should not be like that, I mean it should be a steady 3.5v (at least thats how I understood it) but what I see as a result is the second attached image, which shows the signal being on average at 0.6v with Vpp off 3.5v

Now the 1.8v is up and running, the 0.75 is up and running (at 0.44v) but there is no powergood signal coming out of it (its 0.17v with vpp of 2v) if no powergood from 0.75 then there is no GFX or SOC rails

my question, what should be the circled capacitors voltage (I think that is the MVDDC capacitors), I guess it should be at 0.9v which I see it when I run it at the moment, once I make a steady 3v to the 0.75_EN pin then all rails are up and running at their correct voltages, except the circle capacitors rail, which changes from 0.9v to 1.8v which is too high for the memory, but I maybe wrong, and the correct voltage there should be 1.8v, thats why I'm asking for reference voltage of these capacitors .

All my best,

I just realized, I still have a short from VPP to VDD_MEM (The circled capacitors in my first post should be VDD_MEM). This short could be a GPU short or a memory short, the maximum voltage on these pins is 2v so I think I have to feed VPP with 2v and see if the main die heats up or some other memory chip? are there any ideas?

All my best,

UPDATE:

After looking at the schematic, I could not find VPP going to the main die, so it must be a memory chip that has shorted both rails, I have a feeling its the one on the bottom :D but I need to verify..

OK the problem is found, see the attached picture, its C373 (22UF/6.3V(0805) X6S) its the VPP main capacitor, this was the reason of shorting several memory chips, as well causing all this trouble most probably. as I always hear, its a bad cap.

Now VPP resistance to ground is around 6k and VDD is setting at ~60 with, 5 memory chips removed.

The challenge just started I would say :D

OK. replaced 5 memory chips, but the 0.9v rail for vddc_mem did not come up, so i assume the 0.9v on that rail was actually coming from the 1.8v vpp short I described above, so in theory this rail never came up..

RIght now all GFX/SOC rails are up and running, and the GPU is hot but there is no vddc_mem, the vr controller on the back which is IR35271 seems to be dead, i found a short on the 1.8V decoupling pin, so i replaced it.. Now with a working IR35217 I can see the vddc_mem is coming up for a short while, but then the IR35217 seems to shut it down for another reason..

just a correction, the three missing phases are the vddc_mem and vddci which are controlled with ir35271, they all come up for a short while, like a split of a second and then goes down..

Update:

I think at this point as I can see the rails coming up and then shutting down, I can say the IR35271 needs to be programmed, which is a new chapter :)

OK now i have plugged my esp32 in, and I can see all the I2C addresses :D the problem is I think finding which one is my device, then figuring out what to communicate..

I think the IR35217 address is the 0x72, of course if anyone has an idea or two, feel free to share.

I managed to get to the point where all rails are up and running the device manager can at last see the card, bur it does not yet put a picture, i programmed the IR35217 with the IR PowIRCenter, i just reprogrammed the USER section to do 1+1 loop and set the vmax for loop 1 and 2 according to the screenshots here https://elmorlabs.com/forum/topic/6900x ... it-tricks/

the memory rail picked up and all voltages are within the correct ranges, 1.3 for memory and 0.87 for Soc/Aux.

So more work apparently still needed, this could be a memory issue, as I had to replace 6 chips on the card, which make it a bit hard to investigate without proper software, if you guys know where to find the tCore engine for NAVI21 I could not find anything other than NAVI10 for tserver I'm referring.

ok i found a nice tool on this forum, named memtune, it was able to actually see the 6900xt and test its memory, the card was not visible for any agt or tserver command as I can't find the 6900xt 21tCore engine, but it seems this memtune is doing something, I already had a suspicion on G and F and also H, but H is not reported so I will start with G and F and see if memtune reports something else,

I used the retrainingcheck with memtune, which reported the attached image.

apparently this memtune is lying, I replaced F and G chips, and made sure they are working chips as I got it from a working/dead board.. so it must at least report something different this time, but it reported the same channels as before so something is not working right..

I still cant find the tserver tests for navi21, trying something like dmgg.py crashes the kernel as apparently some security measures are being taken, from linux kernel side of things (IOMMU) most probably related.. So it would be using a software that wraps around all the pitfalls.

Any ideas?

All my best,

I can still at least see it in PCIE and apparently memory is ok from what GPU-Z is showing, but GPU-Z could be wrong anyway :D

Good news,



The memtune is not lying, but I did not know which is E from F from H, I was doing the old style anti-clock wise from the gpu, this is outdated..

Now I have a different problem with only three CHIPS D-E-G

F0/1 is working, which was reported at first as a failure, I can tell you this is 100% working software to test the memories..





use memtune, if its a memory problem it will mostly tell you the truth

I just ordered the stencil for the NAVI21 chip, after getting to the point where only G channel is not being trained. I replaced it, and then I got a dead 0 ohm on memory rail, I had to remove three memory chips and the main chip to find the short, and it was a small dust under the memory, when i removed the main chip, i have broken the BK47 address, which is a PCIE related pin, doing something apparently important, but the broken pad, had still a small thread coming from pcie, so i have connected the pad and i can probe the pcie lane again from the BK47..

As I have three memories to reball and the GPU itself, i will take a break from posting, at least it will be three weeks until the stencil arrives, but then I think it should work :D but maybe not.. I thank the Lord of heaven and earth for whatever comes the way...

All my best,

Ok some progress here :)

I have managed to reball this guy after I have thought that I have a problem with the GPU balls as I have been getting a picture with it and with 0 memory failures according to software, but the picture i was getting is kind of different than what I expected, I should see a single screen, but I was seeing multiple reflections of the screen.. So after reballing, I still have the same problem as before. Now with memtune i have 0 failures all fbcheck and trainings are successful but I still face the same problem again after the reballing..

Should I plant a new GPU here or there is something else I can do to test it (software or hardware).. I can boot with it into Linux and run all the test suits.

Any advices are appreciated..

I'm proceeding with this, I can't think of anything else at this point, everything seems to be fine, but I guess there is a problem with the chip

OK I replaced it with a working known GPU chip, and I got the same exact three mirrored screens again, any idea, have anyone experienced something similar?

ok some pictures might tell better the case :)

same card issue for me
i will try and replay it

bargo


Thank you for sharing your story, would love to get this card sorted out :) I think it could be a pcie issue of some kind, although it seems like a memory issue, but i dont think its a memory issue i have a feeling some channel is corrupted on the pcie lines.

All my best,

ok here is some nice photos of this card partially running, but with no memory issues at all after reballing the gpu the first time i get the same picture then after replacing the whole gpu with a new one i get the same image, even after replacing the chips for almost three times for each one of them. I still can't get what is wrong. but also the tcoreengine i currently have does not work blaming something about secure OS load and unable to load ucode stuff.. I'm updating my current version of memtune which ends with .94 with the one that ends with 0.98 which may help find the issue, which i doubt.*zes also 0.98 gives the same successful results_


Ok i managed to get my margintest running but it fails for another reason below:

[323:000] (Error) |KmdLib| PspTrainMemory2460CPSP11AutoLoad::PspTrainMemory() - Timeout waiting for bootloader ready.
[334:000] (Error) |KmdLib| PspBootloaderLoadSw CPSP11AutoLoad::PspBootloaderLoadSw() - Timeout waiting for bootloader ready.
[334:000] (Error) |KmdLib| CNavi21MicrocodeLoader::LoadMicroEngineUcodes() - SMC firmware load failed: bResult: 0x0


people are dying, but all people will die.. The one Lord will stay alive and he will never die..

Now I'm still not able to run any margintests or to loadmeucode from agt on 6900xt cards

it fails with:

Wait: Secure OS load.
SMC FW load failed.
Failed to set PowerPlay mode 1
AMD GPU TOOL version 3.0.0.23E, Copyright (c) 2020 Advanced Micro Devices, Inc.

[info][Framework] vbios.version: 020.001.000.071
[error][Framework] Exception during attempt to execute test: TcoreSiliconTarget: Failed to load microcode using kmdlib on Navi21 PF
[error][Framework] UMC321.025
[error][Framework] errorid: - UMC321.025PA
Test Count: 1
Total Pass: 0
Total Fail: 1
Total Skip: 0

Part is locked. Can not continue Margin Test


Have anyone got successful to run the margintest or loadmeucode on any rx6900xt or 6800xt or newer cards?

memtune reports no failed channels and fbcheck is successful mctraininfo returns all channels have trained successfully, so i dont think memtune is reliable enough as I thought before.

After trying all the Navi21 images available on this forum, I could not find one that runs a memfa boardtest successfully.

Had any one succeeded in Navi21 to run memfa would be great to list what was used:

1. Motherboard type
(I tried gigabyte b450m arous, asus b550E, old am3 msi grendae)
2. Kernel version
(I tried 4.4.31, 4.15.46, 5.12)
3. Kernel boot params (I tried all of them with iommu=on or off)
4. tserverlite version
(20.10.04)
5. agt version
(3.0.0.30E)


Update. Ok now i found the kommunist 12345 stuff I will check it out




Memtune trainingcheck and memtune fbcheck returns all successful no memory failed training



The card boots and runs in both legacy and UEFI boot, but the screen shows some static artifacts.

GPU and all memories have been replaced more than once, with the same results.

you did check outputs, e.g. DVI, HDMI and they all show the same ?

Thanks for reading. Yes, I did. But they all show the same exact symptom.

If i just can get the tcoreengine to work well i could then run all the suite tests. But the problem tcoreengine is failing due to smc firmware load failure. I have a feeling that the Sapphire Nitro+ 6900xt Special Edition (SE) have a special smbus address for the memory VR, if anyone have a similar card and can share the details of the I2C smbus loop1/2 addresses it would be great.

My current I2C address for the Memory VR controller. Is 32 and the loop1 address is at 71 and loop2 at 72 all hex of course.

the loop1/2 address is 0x72 and 0x73 not as i mentioned above