rtx and multiple nvlink

[super gnu]

ive not seen a conclusive response to this question so im starting a thread.

in the past Vlado suggested that future gpu drivers would not allow multiple pairs of nvlinked cards to do the "double memory" trick. ive no idea if this is the case yet, as i dont yet have nvlink, but im considering an upgrade.

the latest version of vray requires the newest nvidia drivers to allow RT cores to be used.

so, do the RT compatible drivers also allow multiple nvlink pairs or not?

if not its a real shame

[Nejc Kilar]

Seems like the latest drivers disabled the option to run NVLink on more than a pair of cards (1 + 1). Whether that is inadvertently or on purpose I think is still unknown but some nice people already contacted Nvidia about it (check the nvidia forums) and they said they'll investigate it. So yeah, technically right now you can't run more than one pair but the jury is still out

[super gnu]

well it doesnt seem like its an accident. ages ago vlado mentioned he thought it was going to be removed as a feature in later drivers. which , if true is an unbelievably a**hole move by nvidia, since im sure more than a few cgi peeps have multiple (extremely expensive) pairs which have just had their functionality seriously reduced. seems like the old "softquadro" hackers should dust off their skillz and get cracking on some custom driver work.

[Nejc Kilar]

Yeah, I mean its a tricky situation. Product segmentation is not fun but at the same to an extent it makes sense because otherwise the numbers probably don't work Still, I think basing workstation builds on a workaround (NVLink dummy plugs to get 2 x 2 pairs working) was a risky move and unfortunately right now, seems like a not so good one. Let's see what Nvidia will say

[super gnu]

well if they didnt mean to allow multiple nvlink pairs to work with "consumer" cards (id hardly call the titan RTX consumer) then they should have disabled it in the driver at release, not ages after , when people may well have spent tons of money on multiple cards, since it was known to work fine. and i dont consider usiong dummy plugs a "workaround/hack" AFAIK its just how you do it with non tesla cards, nvlink or not.

[squintnic]

http://www.cgchannel.com/2020/02/rev...e-rtx-2080-ti/

looking at the tests here, particularly the high memory situations really shows VRAY GPU without NVlink isnt really a usable render engine for complex stuff. showing crashes on 2080Tis which reflects my frustration with it on everything i work on. great for teat pots - terrible for production.

save your money and get more CPU power if you do arch vis, vfx or simulation based work.