I have a result file e4s136_e1s570p1f505-ADRIA_AdB_KIXCMYB_HIP-1-2-RND8920_0_9 which is 519.996.884 bytes long.
The upload of this result file was stuck at 476.41 MB (96.1%) for ~5 times, but (to my surprise) it eventually succeeded.
The event log says:

Started upload of e4s136_e1s570p1f505-ADRIA_AdB_KIXCMYB_HIP-1-2-RND8920_0_9
Temporarily failed upload of e4s136_e1s570p1f505-ADRIA_AdB_KIXCMYB_HIP-1-2-RND8920_0_9:transient HTTP error

I'm aware that this could be an undocumented apache webserver limitation / bug (which couldn't be fixed by the project), or this could be another effect of the unknown DDOS prevention (which prevents us from accessing the GPUGrid website in rapid succession from the same IP address) but in this case the project should make the workunits shorter to prevent volunteers waste time / energy crunching over 12/21/52 hours in vain.
The present workunits are annoyingly long anyway.

This is caused by an (outdated) Apache configuration.

See thread 5261 - they're the same issue - and it could easily be cured by changing the limit in the configuration.

in this case the project should make the workunits shorter

It seems that your clear claim has been at last attended.
My first task of the new batch "_ADRIA_BanditGPCR_APJ_b0-" came out today.
It took for my overclocked GTX 1660 Ti graphics card 63368 seconds to complete.
This same graphics card was completing previous batch tasks in about 103000 seconds. So there's been a significant reduction.
I was also aware to suspend network activity when the mentioned task was next to finish.
The heaviest "_9" result file was 266,18 MB in size. This leaves a great margin for not reaching any oversize.
For the same reason, I think that a greater amount of tasks can be made simultaneously available without the risk of server's data handling become overcommitted.
I hope this results in an steadier task supply...

More details in Managing non-high-end hosts thread.

... This leaves a great margin for not reaching any oversize.
For the same reason, I think that a greater amount of tasks can be made simultaneously available without the risk of server's data handling become overcommitted.
I hope this results in an steadier task supply...

so what remains to get solved is that tasks can finally be crunched under Windows and Ampere cards, when cuda101. So far, this combination does not work yet.