Looks like it might have just been a lucky streak for me, as they finally ran out a couple of hours ago. Still, the more crunched, the better...

What sort of GPU are you getting those errors on?

That surprises me. There's a known problem with some of the slowest integrated iGPUs, but discrete GPUs like those should be OK. Are you sure that some other application, like Folding at Home, isn't competing for the resource? That's known to have been a problem, because sometimes their apps run at a higher process priority level than BOINC apps.

@Crystal Pellet
How many simultaneous WUs are you running in your GTX970?
I'm running 4 in mine.
My last page (15 Wus) of Results Status for it shows they average about 0.6h on the current batch 0044xxx WUs, range 0.35h - 0.95h.
Divide the 0.6h by the 4 WUs done at the same time = 0.15h = 8 min each.
That's reasonable for a 970 I think, which sits at 9730 G3D Marks on the benchmarks list at https://www.videocardbenchmark.net/high_end_gpus.html .
4 simultaneous WUs may not be optimal in a GTX 970, but the variability in its diet of WCG WUs and the stop-start flows have not allowed me the optimise that number.
Other members' experience would be interesting.

In another machine I have a GTX1650 Super (9904 G3D Marks) doing 3 Wus per 0.4h = 8m ea, plus an AMD RX570 (about 7000 G3D Marks) doing 2 Wus per 1.0h = 30m ea (snore). I'm allocating a full CPU core to all types of GPU, including Intel, (<cpu_usage>1</cpu_usage> OR <avg_ncpus>1</avg_ncpus>), as that prevents the CPU parts of the GPU WUs having to share a CPU with another running task such as a CPU WU. The CPU speed probably affects the GPU total WU times. (The machine with the RX570 has only a 2600K at stock 3.4GHz, so I need to hasten that. The others have 3770Ks at 4.2GHz.)

Yet another machine is running an RTX3060 (16700 G3D Marks) , doing 5 WUs per 0.35h = 4.2m ea.
Hope this is of interest ...