Just SWAP is inflexible, especially with the way Linux usually handles buffer/pagecache, memory allocation,and countless other things, if not explicitly configured otherwise. You'd probably get the most out of it if you'd use https://en.wikipedia.org/wiki/Zram setup to use that VRAM instead.
Except if it has a very weak CPU which struggles with the load caused by compression, or power constraints, because mobile? But even then the compression is configurable.
I once used something similar on an old Thinkpad 60p with an ATI FireGL which had some 32bit Intel Centrino, and had only 3.25GB usable RAM from the 4GB. The FireGL had 256MB of which I only needed 2, and let it have 6, just to be on the safe side. The other 250MB I gave to compcache pointed to the device created by some (experimental?) blockdriver I can't remember anymore.
This wasn't automatic, didn't work on the first tries, and produced crashes.
But was worth it once I had it right(¹), because that feeble thing ran much later into its limits. Noticeably. Not by some meaningless benchmarks.
(¹) much fiddling with many hexadecimal memory addresses and ranges, exclusions of the same in other configuration files, scattered over /etc/, mostly Xorg, but also other Drivers/PCI-Address spaces, to not have one BYTE of
overlap, but also not wasting any ;-)
Edit: Skimming the two links phram rings a bell.
It also ran Google Earth! Didn't really matter, on that GPU it probably did everything in software, and not something which the FireGL could deliver, or at least not at the times, with what MESA, XORG, DRI could do then with that card.
Except if it has a very weak CPU which struggles with the load caused by compression, or power constraints, because mobile? But even then the compression is configurable.
I once used something similar on an old Thinkpad 60p with an ATI FireGL which had some 32bit Intel Centrino, and had only 3.25GB usable RAM from the 4GB. The FireGL had 256MB of which I only needed 2, and let it have 6, just to be on the safe side. The other 250MB I gave to compcache pointed to the device created by some (experimental?) blockdriver I can't remember anymore.
This wasn't automatic, didn't work on the first tries, and produced crashes. But was worth it once I had it right(¹), because that feeble thing ran much later into its limits. Noticeably. Not by some meaningless benchmarks.
(¹) much fiddling with many hexadecimal memory addresses and ranges, exclusions of the same in other configuration files, scattered over /etc/, mostly Xorg, but also other Drivers/PCI-Address spaces, to not have one BYTE of overlap, but also not wasting any ;-)
Edit: Skimming the two links phram rings a bell. It also ran Google Earth! Didn't really matter, on that GPU it probably did everything in software, and not something which the FireGL could deliver, or at least not at the times, with what MESA, XORG, DRI could do then with that card.