26 June 2020

Adventures in a Graphics Card Upgrade

Decision Time

As my HIS IceQ Radeon HD 7850 2GB was getting on in age, I decided it was time for an upgrade. I was trying to stay around the 130W TDP of the current card so that my folding did not increase my electric bill. The main options that I was considering was:
  1. Used GTX 1060 ~$135-150
  2. GTX 1650 ~$159
  3. GTX 1650 Super ~$200
  4. GTX 1660 ~$215
  5. GTX 1660 Super ~$230
  6. RTX 2060 KO ~$300
  7. Used RX 470 4GB (avoid mining only) ~$75
  8. Used RX 480 4GB ~$90
  9. RX 570 4GB ~$120
  10. RX 580 ~$155
  11. RX 5500 XT 4GB ~$159.99
My first elimination was anything over $200 as that was just too expensive to pair with my old Core2 Quad Q6600. I then eliminated the RX 570 and GTX 1060 as it was just overpriced compared to the RX 480 and RX 5500 XT. The RX 580 and GTX 1650 were also overpriced compared to RX 5500 XT. This narrowed my choices to:
  1. Used RX 470
  2. Used RX 480 4GB
  3. RX 5500 XT

The Purchase

I then scoured ebay for the used graphics cards. I placed the following price guidelines (including shipping) RX 470 for $70, RX 480 4GB for $80, RX 480 8GB for $90. It seemed that the 8GB models were pulling a large premium and the 470 and 480 4GB models seemed to be about the same price. I also wanted a backplate on the GPU to try to avoid GPU sag.

I was able to score a MSI RX 480 Gaming X 4GB model for ~$81 with shipping. When it arrived, it was in great physical shape and looked clean. I gave it a blast of compressed air to ensure maximum airflow. I installed the GPU after having to remove a fan mount in my old Antec 900 to increase the room for it. It then booted up without any issue. When I went to run benchmarks is when the bad news came. Under any kind of load the GPU spiked to 90C under 100% fan.

The Fix

To try to resolve this I tried:
  1. Reducing the Core Voltage
  2. Checking for a firmware upgrade
  3. Tightening the 4 screws that maintained mounting pressure between the cooler on the card a quarter turn (be careful as over tightening can damage the chip)
However, none of those helped. It was finally time to try the last thing that I could think of, replacing the thermal paste. Here are the steps that I followed:

  1. Removed the back plate taking care to keep track of which screws came out of where
  2. Removed the two screws on the pcie bracket
  3. Removed the 4 spring loaded screws that maintain the mounting pressure
  4. Unplug one 4 pin connector between the PCB and cooler which allowed me to open it up like a book
  5. This gave me a good look at the largely dried and cracking gray thermal paste
  6. Carefully cleaned this off the GPU and the heatsink
  7. Applied some Silicone thermal paste that I had leftover from my Q6600 build
  8. Reattached the 4 pin connector
  9. Put the 4 spring loaded screws back in, in a X pattern slowly tightening each one to maintain even pressure
  10. Screwed the 2 screws on the pcie bracket back in
  11. Screwed the back plate back on
Then the for the test. I reinstalled the graphic card and booted up the system. Thankfully, the card now worked like new!

Undervolting

The card's defaults are 1303MHz at 1150mV. I turned this down to 1250MHz at 1065mV using MSI Afterburner. This cost me about 4% theoretical performance for a savings of ~15% power (because Voltage is squared), which put the max GPU core/memory power usage at about 105W. This should keep me in the ~130W total board power range that I was looking for. The card is now doing about 65C with 40% fan.

I am still verifying the stability in folding, but at 1060mV it was just barely coming across an occasional error. I am hoping that bumping to 1065mV will eliminate those errors.

Update 2020-07-01: The additional 5mV have resolved the occasional error. I have seen some work units spike to 110W and hit 67C with 50% fan.

Update 2020-07-12: I still saw an infrequent error. I pushed to 1070mV in state 7 and 1055mV in state 6.

Update 2020-09-11:
Here are the default clock speeds and voltages:
1303MHz @ 1150mV
1235MHz @ 1143mV
1191MHz @ 1087mV
1145MHz @ 1037mV
1077MHz @ 973mV
910MHz @ 850mV

Default Memory: 1750MHz @ 975mV

Here is what my final stable underclock is:
1250MHz @ 1077mV
1235MHz @ 1075mV
1191MHz @ 1075mV
1145MHz @ 1037mV
1077MHz @ 973mV
910MHz @ 850mV

Memory Clocks/Voltages unchanged

GPU power usage (not total board power) drops from ~120Watts to ~100Watts