Radeon VII (2nd gen consumer Vega GPU)

[paulunderwood]

Originally Posted by Prime95

Rocm 2.9 here -- if that makes a difference.

 ./gpuowl 
2019-10-25 16:41:04 gpuowl v6.11-11-gfaaa2f2
2019-10-25 16:41:04 Note: no config.txt file found
2019-10-25 16:41:04 90901739 FFT 5120K: Width 256x4, Height 64x4, Middle 10; 17.34 bits/word
2019-10-25 16:41:04 Exception gpu_error: DEVICE_NOT_FOUND clGetDeviceIDs(platforms[i], kind, 64, devices, &n) at clwrap.cpp:70 getDeviceIDs

 /usr/sbin/dkms status
amdgpu, 2.9-6, 4.19.0-5-amd64, amd64: installed
amdgpu, 2.9-6, 4.19.0-5-amd64, x86_64: installed

uname -a
Linux honeypot9 4.19.0-5-amd64 #1 SMP Debian 4.19.37-5 (2019-06-19) x86_64 GNU/Linux

[preda]

XFX Radeon VII available on newegg for $570, seems a good deal in US

https://www.newegg.com/xfx-radeon-vi...82E16814150820

[paulunderwood]

Originally Posted by preda

XFX Radeon VII available on newegg for $570, seems a good deal in US

https://www.newegg.com/xfx-radeon-vi...82E16814150820

These cards are £570 on Amazon right now. A good buy considering they are over 6 times as fast as a 4 core i7 desktop at PRP-3 for Mersennes., So long as you have a suitable PSU and a PCIE-3 slot...

[preda]

I'd like to share my personal experience with the RadeonVII TIM (thermal interface material, i.e. the thing between the cooler and the core), as a warning:

- the "carbon patch" that the GPU comes equipped with originally is very good. Best not to mess with it at all, as it's very hard (in my experience) to improve upon it.

- the cooler contact surface is indeed concave, as others have reported. This does not matter with the built-in "carbon patch" which is thick and thus mitigates the concavity. But the concave shape seems to produce poor contact with traditional thermal paste (which is much thinner)

- lapping the cooler to flatten the contact surface is tedious and difficult to do properly

What I did do with some success though was removing the "RADEON" LED logo, which opened-up the letter holes for some additional air output in that region.

[kriesel]

Originally Posted by preda

What I consider a hardware design fail on the Radeon VII is the large red "Radeon" LED logo placed on top of the GPU that covers the air exaust that would have been there (moving the hot air "up" from the GPU) otherwise. As a result, most of the hot air is directed downwards under the GPU.

Yes. If they'd decentered it and narrowed the solid wall above and below the letters, they could still have a sizable air exit slot below the text, instead of the mere ~1.5mm present on the XFX. Marketing won out over engineering.

[paulunderwood]

I had a problem running ROCM under linux in order to get a desktop under a newly installed debian buster. Unless I pass nomodeset at boot the screen just goes blank. I can run gpuOwl fine through ssh (without nomodeset). I am using the port nearest the motherboard with an adapter from a VGA plug. Any ideas?

[preda]

I think I cracked the mistery of poor RadeonVII thermals: it's not the TIM, not the screws pressure, not the concave cooler, not even the logo.

It's the [bad] airflow:
Most of the hot air exits "downwards" (i.e. towards the PCIe connector), in front of the connector. From there, meeting the motherboard, the hot air is directed towards the front of the GPU. *if* there's anything sitting in front of the GPU (like another GPU or the PSU), the hot air hitting that obstacle moves up, in front of the fans intake, where it's sucked into the GPU and repeats the cycle.

So the designers managed to create a "hot loop", where the hot air keeps looping back into the GPU instead of being removed. The problem is that the usual solution of increasing the fan speed does not help much: a faster fan just makes the loop stronger.

The simptoms: empirically, the indicator of this effect taking place is a very hot GPU body, that can be felt by hand on the GPU case. Also "hot" air exiting the GPU can be a side-effect of "hot" entering the GPU.

The fix: either have plenty of space in front of the GPU. Or better, place the GPU far from the motherboard (with PCIe risers/extenders). The cooling efficiency of the externally-located GPU is shocking!

Anyway, I'm still amazed by the design idea of pushing the hot air "down and front" towards the motherboard and in front of the the PCIe connecter, thus towards the fans.

Originally Posted by preda

I'd like to share my personal experience with the RadeonVII TIM (thermal interface material, i.e. the thing between the cooler and the core), as a warning:

- the "carbon patch" that the GPU comes equipped with originally is very good. Best not to mess with it at all, as it's very hard (in my experience) to improve upon it.

- the cooler contact surface is indeed concave, as others have reported. This does not matter with the built-in "carbon patch" which is thick and thus mitigates the concavity. But the concave shape seems to produce poor contact with traditional thermal paste (which is much thinner)

- lapping the cooler to flatten the contact surface is tedious and difficult to do properly

What I did do with some success though was removing the "RADEON" LED logo, which opened-up the letter holes for some additional air output in that region.

[preda]

On the topic of PCIe risers, I tried (Linux, ROCm) both USB-style risers and PCIe x16 ribbons and they seem to work fine without visible performance impact on GpuOwl. The esential thing I had to do to get those to work was to set the PCIe speed to "Gen1" (2.5 GT/s) in BIOS (the default usually being Gen3 (8 GT/s) which is too fast for the poor connectors).

The difference between USB-style risers and x16-ribbons is that the USB-things require one addional PCIe power input, which is annoying. (the ribbons OTOH simply forward the power from the motherboard).

OTOH the USB-style risers can aslo fit in the small PCIe slots on the motherboard (usually x4 or x1).

[chalsall]

Originally Posted by preda

So the designers managed to create a "hot loop", where the hot air keeps looping back into the GPU instead of being removed.

This is analogous to an electrical "short-circuit".

Without meaning to sound lecturing, but there is a lot to be learnt from studying the Hot aisle / cold aisle design considerations of machine rooms.

Even individual computers' internals should be thought about similiarily.

That heat has to go somewhere...

[kriesel]

Originally Posted by preda

I think I cracked the mystery of poor RadeonVII thermals: it's not the TIM, not the screws pressure, not the concave cooler, not even the logo.

It's the [bad] airflow

It seems to me you're forgetting about the case fans and the third axis.
The circular effect from the gpu fans you describe is only part of the picture.
Add the front to back effect of the case fans influence, and what you get is a helical flow near the gpu.
This amounts to ensuring the in-case air is well mixed and well heated near the gpus, and the heat removal is improved. At 300W max on the Radeon VII, it's optimistic to think you could squirt all that heat out its own 2-slot back bracket. Having the front-to-back flow thermally stratified from tower case top to bottom or desktop case left to right is not desirable. We want all the air passing through the case to contribute cooling, not only a little of it through a stripe here and there.

[M344587487]

Originally Posted by preda

...
It's the [bad] airflow:
Most of the hot air exits "downwards" (i.e. towards the PCIe connector), in front of the connector. From there, meeting the motherboard, the hot air is directed towards the front of the GPU. *if* there's anything sitting in front of the GPU (like another GPU or the PSU), the hot air hitting that obstacle moves up, in front of the fans intake, where it's sucked into the GPU and repeats the cycle.

So the designers managed to create a "hot loop", where the hot air keeps looping back into the GPU instead of being removed. The problem is that the usual solution of increasing the fan speed does not help much: a faster fan just makes the loop stronger.
...

My setup was a 20cm case fan mounted adjacent to the card on the same plane as the riser central to the two left GPU fans (the hot spots). Works pretty well and it's not so loud given that large fans can push much more air per rpm than small ones. Most of the fan pushes air into the other fans, the periphery pushes the warm air expelled by the GPU away.