Vega 64 GPU vs HBM temperatures

allen5055

[H]ard|Gawd
Joined
Mar 25, 2012
Messages
1,553
I have two XFX Vega 64's with EKWB waterblocks in a custom water loop. When Ethereum mining they run at 40c for the cores. However, I've noticed that curiously, the HBM runs 10c hotter. One card runs at 53c HBM and the other runs at 50c.

I noticed when I was putting on the waterblock that the XFX cards have the "non-molded" HBM, which means that it sits a bit lower, and consequently has more thermal paste inbetween the HBM and the waterblock. No doubt this increases temperature. I wonder if I did a shitty job on the thermal paste for one card which is why its HBM is 3c higher than the other card. I used a cross application with Grizzly Thermal Kryonaut paste.

Does anyone know if the "molded" Vega 64 cards have lower HBM temperatures? Is everyone else seeing higher HBM temperatures than GPU core temps?
 
I don't think brand matters if card has molding or not since all reference cards made by AMD. I have an XFX and Gigabyte Vega 64 black versions and both have molding. I put EKWB on one and see the same temp difference during GTA5 benchmark 40-44c for GPU and 50-53c for HBM.
 
Mine stays within a degree of each other... Is mine accurate? Seems so as every bench or game I play it stays with in a degree or 2 C.
 
I don't think brand matters if card has molding or not since all reference cards made by AMD. I have an XFX and Gigabyte Vega 64 black versions and both have molding. I put EKWB on one and see the same temp difference during GTA5 benchmark 40-44c for GPU and 50-53c for HBM.

I wonder if me and you both have a 10c difference because our vega 64's are watercooled with EKWB blocks. Is everyone else that has their HBM and GPU core within a degree or two using an AMD aircooler or watercooler?

The molding makes the HBM and GPU core the same height and should improve contact with the waterblock.

My temperature difference could also be due to the fact my memory is overclocked to 1100 whereas my core is downclocked to 1222 at 875mv. However even at stock settings my hbm still runs 10c hotter.
 
Last edited:
The molding makes the HBM and GPU core the same height and should improve contact with the waterblock.
The moulding shouldn't affect the height. In fact it's lacking on the unequal height parts as contact is otherwise problematic for any heatsink. That was my understanding at least.

I wouldn't worry too much about HBM2 temps unless they hit a level where temp is a concern. HBM2 stacks are just that, so there will be some insular effects as temp sensors should be in the base die. Heatsink contact on the opposite. Seeing a delta would be expected and not a hotspot to worry about. Within just the core is another concern.

One thing you guys with water blocks might test is dialing back the fan speed so it runs a little warmer. If undervolting that may help a bit and leakage at lower volts less a concern. Run it at 60 or 70C and see if there is any difference.

Still need to get myself a Vega to play with, but not in a rush on Linux. Soon.
 
I wonder if me and you both have a 10c difference because our vega 64's are watercooled with EKWB blocks. Is everyone else that has their HBM and GPU core within a degree or two using an AMD aircooler or watercooler?

The molding makes the HBM and GPU core the same height and should improve contact with the waterblock.

My temperature difference could also be due to the fact my memory is overclocked to 1100 whereas my core is downclocked to 1222 at 875mv. However even at stock settings my hbm still runs 10c hotter.
Mine is the watercooled version, again only a degree or 2 C difference.
 
The moulding shouldn't affect the height. In fact it's lacking on the unequal height parts as contact is otherwise problematic for any heatsink. That was my understanding at least.

I wouldn't worry too much about HBM2 temps unless they hit a level where temp is a concern. HBM2 stacks are just that, so there will be some insular effects as temp sensors should be in the base die. Heatsink contact on the opposite. Seeing a delta would be expected and not a hotspot to worry about. Within just the core is another concern.

One thing you guys with water blocks might test is dialing back the fan speed so it runs a little warmer. If undervolting that may help a bit and leakage at lower volts less a concern. Run it at 60 or 70C and see if there is any difference.

Still need to get myself a Vega to play with, but not in a rush on Linux. Soon.

The molding does affect HBM height. See the article from TomsHardware on molded vs unmolded.
http://www.tomshardware.com/news/amd-vega-package-problem,35281.html

My cards both come with Samsung memory.
 
Hm, so I've now been mining Ether for 12 hours. I checked my HBM temps again, they are 15-17c higher than the GPU core.

However, my GPU core is heavily undervolted, at 0.85v (0.3v drop), the GPU is underclocked. The HBM on the other hand is overclocked from 950 to 1100Mhz, and Ethereum mining intensely uses the HBM. I am using the August 23rd AMD blockchain driver. As soon as I stop mining, my HBM temperature instantly drops down as does the GPU core temperature, and the temperatures are identical. Instantly. As soon as I restart mining, the HBM temp instantly jumps up to 15c higher than the GPU core.

Can anyone that's done Ethereum mining weigh in on their HBM vs GPU core temperatures?
 
Last edited:
The molding does affect HBM height. See the article from TomsHardware on molded vs unmolded.
http://www.tomshardware.com/news/amd-vega-package-problem,35281.html

My cards both come with Samsung memory.
aHR0cDovL21lZGlhLmJlc3RvZm1pY3JvLmNvbS9HL00vNzA0MzI2L29yaWdpbmFsL1BhY2thZ2UtMDEuanBn.jpg
From the article you linked. The chips with moulding were always flush, they'd have to be without some sort of slope between the CPU and HBM, making it pointless. It's not changing the height as much as components from different manufacturers are different physical dimensions. Someone having shorter HBM stacks that prohibit the moulding, not elevating the HBM to make it flush. Haven't followed the story close enough to know if the GPU or memory is the culprit.
 
I have a XFX 64 and 56. Both Samsung memory mining ETH. 17.8.2 drivers memory at 900 and gpus 1000. Both cards are running at about 60 C gpus 2-3 C more at the hotspot and HBM temp at 70 - 75 C.
Both cards are black versions. Temps are perfectly normal. Water cooled cards will run much cooler. I can run both at 1125 on the HBM, but fans need to be cranked all the way up to 75%.
Both are undervolted to 0.850mv
 
I have a XFX 64 and 56. Both Samsung memory mining ETH. 17.8.2 drivers memory at 900 and gpus 1000. Both cards are running at about 60 C gpus 2-3 C more at the hotspot and HBM temp at 70 - 75 C.
Both cards are black versions. Temps are perfectly normal. Water cooled cards will run much cooler. I can run both at 1125 on the HBM, but fans need to be cranked all the way up to 75%.
Both are undervolted to 0.850mv

How'd you undervolt? I spent days undervolting my cards, and the only thing that worked for me was a registry tweak that someone else had prepared. However, I would love to clock my GPU core even lower and go for 0.8v but I don't know how. Wattman doesn't successfully change voltage for me using the 17.8.2 drivers.

This is the registry tweak that finally let me undervolt my cards. But... to change any more settings besides what he already encoded, I'd have to edit the matrix, convert values to hex, etc.
 
Get rid of the registry power table update and run the miner. Set your desired clock at the next to the last option in Wattman. Set as minimum. Wait 30 sec while the miner is running. Set your voltage next and repied for every change you make.
One change at a time seems to do the trick. Hope it helps Happy mining :D
 
Get rid of the registry power table update and run the miner. Set your desired clock at the next to the last option in Wattman. Set as minimum. Wait 30 sec while the miner is running. Set your voltage next and repied for every change you make.
One change at a time seems to do the trick. Hope it helps Happy mining :D

So set the desired gpu core under the p6 power state?
 
So set the desired gpu core under the p6 power state?

I figured out how to use the registry modded power states and make changes in Wattman to get my voltage down. If I set the power limit to -41%, and then change the GPU core clocks down from 1222 to lower, the voltage will go down to 0.8125v. If I don't set the power limit lower, once I change the GPU clocks down the voltage will shoot up. I was able to shave off 40w lowering the voltage and core clocks from 1222 core to 1050 core with voltage dropping from 0.85v to 0.8125v. My hash still remains around 44 per card. I'll eventually get around to tweaking the HBM memory voltage, but I'm leery of doing so since I have it overclocked to 1100mhz. If I raise it past 1100mhz it crashes almost instantly.


I've noticed that as I've underclocked and undervolted my GPU core even more, the temperature differential between the HBM and GPU becomes larger. Right now my GPU temp is 36c and my HBM temp is 54c. So I'd guess that my much lower GPU temps are a result of how much I'm downclocking/undervolting the GPU whereas my HBM stays hot because it's running full blast at 1100mhz without an undervolt.
 
Back
Top