This is happening on Nvidia Cards also - 1080ti can not do anything once the models load (actual work for viewport) hits the card. Usually in changing the viewport to show shadow (in my sequence) will put into freeze. Depending on registry TDR time setting is usually equal to the time of freeze before crashing & if you stack the calls times the TDR Delay you will come up with your length/severity of freeze. Even this beast of a card can not handle the 100% load due to TDR (even at over 90).
I know the reader will probably say 'wth is he talking about?' here --- > If any setting on comp calls for 'hardware acceleration' and a GPU gets a call for this, weather you are browsing web, using GIMP, or any other 3rd party program, this same error will occur. (this kicks on/reqs the OpenCL to handle the task in some manner is why it is mentioned) It will report to windows 'Hardware Error' 'Not Reported' and call its code 141, which is in reference to GPU driver issues. If anyone is having this same error, it may also come with some pixilation & 'rain' dots that are inverse pixel specs on the screen. This is even outside of blender. So when I have trouble shot this issue the past week, I notice Turing architecture will not fail, but the older architecture will fail. It is an issue with OpenCL file acting as terrible hinder vs. helper.
My post doesn't have any of your specs / stats and all that stuff, so It's OK to remove it & call it outrageous /ridiculous as nobody knows me, & I have no details due to the fact that it is kinda hard to screen-shot a frozen screen & system stats are kinda redundant when probably over 5K people (in NVidia forums) are complaining of same issue, except on a certain game, and i can tell you it is doing the same exact behavior, as another one did years ago with a different name.
And the issue is the OpenCL file will somehow work like 'hardware acceleration' works, but it only hurts things creating the TDR to timeout and windows calling it a hardware failure. But this comes from the command from the programs to hand all to OpenCL in some manner. I do not code python or blender, but i code other things, and as being semi-literate in how things work, but maybe you might wish to look outside of this tiny post and this tiny program for a bigger scale of this happening.
Sorry for not posting my system specs and a recorded emulation of a frozen screen that causes a crash.... But this is more of a generalized message to someone who has the brain-power and authority that could actually DISABLE the OpenCL 'acceleration' that freezes things, that might be a temp fix.
I have no idea if this is possible, but if so, and if you can, please send the code for temp disabling of OpenCL to NVidia as they have post after post after post of user complaining of this as well. All using the 10-series cards. If i were using my coding skills, I would probably set an argument to grab the time & send fake message to windows TDR that the delay is not exceeded while the GPU sorts out the load with OpenCL. And I really think if OpenCL could be stopped for a moment (without stopping entire progress of desired action), there would be no need to worry about TDR as that is what windows classifies the error to stem from on error logs.
I have spent 2 days on it and somewhere there is a glitch on the path mentioned. And it is not just AMD cards. It is all cards that will somehow send a 100% full load to the card, kicking in TDR and depending upon registry settings, Blender may Freeze, crash, or windows in its entirety will crash. And this issue has happened already in blender 2.81 a little less than 2 years ago. I do not know what fixed it, but i am sure some of the Guru's here can recall the issue.

Delete my post if you want, as I'm sure it's all gibberish, ill-formatted, no-recorded freeze to reference, & non-sense to the one who rarely step outside of Blender, or the web for that matter - but maybe someone know what I am saying, and know more of the inner workings better than I do - & maybe secretly it can be used to figure this out by an insider.  And please do not pick apart my attempt to help and quote me on shit. I'm just trying to help and have been working on this for a couple of days myself. And as i re-read the data above, it looks like little has been done in the last year to to closer to fixing.... just the endless needing the prove and re-prove and re-prove the that the problem exists. Well, why doesn't someone with a bigger Blender/GPU brain than myself work on a fix. It has been over a year it looks like. Maybe we could work on fixing it instead of just the endless back and forth on weather or not it exists...   I would like to use Blender again someday.

**NOTE:: It takes time to type all this out. If you just call is shit and say it's not more blender hard-data 'proof' and remove the connection I have presented (which is partially presented above), then maybe one is not interested in finding solution, but still trying to prove weather it exists or not a year down the line....