VRAM underclock "fixed" this GPU, but I want to know WHY

[zakuradyte]

(TL:DR How did underclocking VRAM fix a broken card)

 

Hello all,

Today I traded a 1660 ti for an RX 570, $100 and a "broken" Gigabyte Windforce 2070 from a friend at work. He isn't very computer savvy and just wanted to upgrade from the 570 he was using, so I figured I could do something to make the 2070 that broke on him work again, given what he described as the problem. I get home, plug in the 2070, launch Heaven and get a ton of rainbow artifacts and black screens, and the PCI connector LED were blinking the whole time (apparently this is a warning for "abnormal power supply"). Not knowing when to quit, I opened up afterburner, underclocked GPU and VRAM to as low as they went, undervolted, and set my power limit to 75%. Ran Heaven again and got what appears to be the lowest performing 2070 I could possibly have, but the artifacts and black screens disappeared, with only a dropped frame or two every 20 seconds or so (didn't run a benchmark set to get AVG and 1%, but noticed the stutter). Not satisfied with that, I pulled the cooler off and repasted the GPU, checked the thermal pads on all components were making good contact, and tightened all the screws down tighter this time. Running heaven again, no crashes, no artifacts, even fewer framedrops. I changed the clock frequency back to default and ran it again, no problem (PCI LEDs still blink, but no issue otherwise). Finally, I started bumping up memory frequency bit by bit until finally the black screens and artifacts returned, somewhere in the -330 range on the VRAM. As per the TL:DR and given this information, can anybody give me an idea on what component failed on this card/ if or how it can be rectified. Mainly I'm trying to figure out if I have a bad GDDR6 module or if the power delivery to the VRAM is to blame. If me or any of my friends can't fix it (I'm an electrician and work in the company of electronics technicians), can anyone else do it for money in the commercial space? If not, i'm just gonna leave it as is and live with the performance loss from underclocking the VRAM. Thanks in advance, this forum has always been a great place for advice. 

 

Pic for proof, specs are as follows (it's my GPU test rig):

 

ASrock x300 PC 

Ryzen 5 4650g 

16GB DDR4 3600MHz

Windforce 2070 8G running from m.2->PCI X4 Adapter W/ 250W PicoPSU as supply

 

 

[Helpful Tech Witch]

memorys bad obviously, it cant operate at stock speeds anymore. You could try a bios mod to make it permanant, ig but I dotn know how long it will last

[InstantNewt]

Do you have another PSU you can use? A 2070 pulls about 215w under load, and with a 250w picoPSU it’s possible the load is overwhelming the PSU if it’s not the best quality. Undervolting the card would reduce its power draw which could explain the “fix.”

[zakuradyte]

3 minutes ago, InstantNewt said:

Do you have another PSU you can use? A 2070 pulls about 215w under load, and with a 250w picoPSU it’s possible the load is overwhelming the PSU if it’s not the best quality. Undervolting the card would reduce its power draw which could explain the “fix.”

That was my next step, I have a system with a 750W 80+ gold PSU that is currently running a Tesla M40 in "fake laptop" mode to play games, just gotta swap it to try. 

[InstantNewt]

1 minute ago, zakuradyte said:

That was my next step, I have a system with a 750W 80+ gold PSU that is currently running a Tesla M40 in "fake laptop" mode to play games, just gotta swap it to try. 

Try that and maybe try a full x16 connection if you can. I believe x4 is usually limited to 25w and the 2070 will pull about 50w over its PCIe slot. 

[zakuradyte]

Just now, InstantNewt said:

Try that and maybe try a full x16 connection if you can. I believe x4 is usually limited to 25w and the 2070 will pull about 50w over its PCIe slot. 

will do! i'll update in a bit once it's swapped.

[zakuradyte]

30 minutes ago, InstantNewt said:

Try that and maybe try a full x16 connection if you can. I believe x4 is usually limited to 25w and the 2070 will pull about 50w over its PCIe slot. 

Confirmed with my 5800X system, stock clocks cause a near instant crash. Now i just wish i knew if it's a memory module issue or a VRAM power delivery issue. 

[InstantNewt]

14 minutes ago, zakuradyte said:

Confirmed with my 5800X system, stock clocks cause a near instant crash. Now i just wish i knew if it's a memory module issue or a VRAM power delivery issue. 

Probably not a power delivery issue since you’ve confirmed it on a better test bed. You could try to pin it out with a multimeter, but I agree with @HelpfulTechWizard, memory is probably dead. 

[zakuradyte]

21 minutes ago, InstantNewt said:

Probably not a power delivery issue since you’ve confirmed it on a better test bed. You could try to pin it out with a multimeter, but I agree with @HelpfulTechWizard, memory is probably dead. 

That's the big sad. Off topic, how long do you think it would last as it is now?