dual 3090 ti SLI/ nvlink early benching

Can you tell me a little more about the performance in Star Citizen. Im about to NVLINK mine and Im curious about the performance you are getting. TIA.
 
I am getting between 68-119FPS at 4k, with most max settings. I also load straight from the exe and not use the client.
(Its not SLI )
 
How were you able to get your hands on 4 founders edition 3090 Ti's at/near launch… were they just easy to get at local BestBuy's?

Nice rigs and results! That Titan V CEO edition still looks good as well.
 
Last edited:
How were you able to get your hands on 4 founders edition 3090 Ti's at/near launch… were they just easy to get at local BestBuy's?

Nice rigs and results! That Titan V CEO edition still looks good as well.
No magic powers. And I don’t have 4 of the Ti s

One Ti bought from Best Buy
One, Ti I had my fiend get from Best Buy (yes I payed him)
one Ti I bought new from eBay for about 200 more than retail from someone that was waiting for a Best Buy order fulfillment because I thought I’d just put that in my wife’s Pc but changed my mind (so 3 Ti s)

I had 4 x 3090 FE, sold two in advance to help finance the Ti s (almost a wash because they were sold 2 months in advance)
So I have 2 3090 Ti and 2 3090 FE in wife pc

So I have a spare in box 3090 ti FE , 2x 3090 ti FE in my rig, and 2x 3090 FE in my wife’s rig - which is complete overkill for her.

When I do my DL, NN. I share workload across both rigs and all 4 cards, -the wife would really not know what to do with her 2x 3090, she generally plays cyberpunk, Detroit, bio mutant and Sherlock Holmes. Her main role is editing my dissertation ;)

She’s a pro photographer so most of her pc is about photoshop, Gimp, and web design.


There was a moment where I was going to put all 4 cards in my pc and have it converted to a mining open case to accommodate it, but running DL,ML, and NN across both PCs was working and seemed easier

Hope that clarifies the two rigs with a left over 3090 Ti FE
 
not sure what "negative lod bias" has to do with RDR2 version 1436
It's why some prefer to use DLSS over Native at high resolutions when it improves distant image quality despite the tradeoffs that DLSS introduces.
 
Interesting read! I recently obtained two 3090's since the price is right in the secondhand market and I am interested in finally getting into machine learning. I recently finally got both cards into the same rig once i was able to upgrade my PSU, man that was a lot of work since I had to do it sprawled/hunching over the computer instead of at a desk since it also has like 14 hard disks inside... --- These GPUs do get a little bit of use for stable diffusion but overall I've still yet to truly take advantage of them. The two cards that I have are "incompatible" for nvlink though, since one is an FTW3 and the other an XLR8. So although the nvlink fingers line up horizontally, they do not line up in height. I estimate around 30mm of 180degree x16 riser could do the trick, then I would just need to fashion a way to mount it there, and then nvlink could reach. But I can't justify even continuing to think about that further, if little real world benefit is even to be had from nvlink.

From time to time I wonder about "SLI"ing them (the proper term for these appears to be NVLink them) for games. It sounds like RDR2 is a candidate for this, which is music to my ears as I haven't finished that game yet and getting a graphical upgrade would be nice, however my 4K120 TV is in the living room and I will likely never actually be able to physically bring these two devices within range of one another...

I'll give gravitymark a spin with this, I'll probably get a kick out of that. But I wonder if, since you are clearly knowledgeable in this area, if you could comment on, beyond games, whether NVLink connectivity can make a tangible difference. It seems that the majority of software is not written with such topology in mind sadly. Do you benefit from NVLink in the workloads involved with your dissertation? I'll guess not.

Even for stable diffusion the easy way to get it set up is to run a separate instance for each GPU and just juggle them manually. It is okay, but far from ideal.

If I had the kind of free time that I'll never have again, I would love to get back into building some physics engine stuff, with CUDA. The amount of compute we have on tap is truly mind boggling these days, we're about to enter the petaflop-under-your-desk era...
 
Interesting read! I recently obtained two 3090's since the price is right in the secondhand market and I am interested in finally getting into machine learning. I recently finally got both cards into the same rig once i was able to upgrade my PSU, man that was a lot of work since I had to do it sprawled/hunching over the computer instead of at a desk since it also has like 14 hard disks inside... --- These GPUs do get a little bit of use for stable diffusion but overall I've still yet to truly take advantage of them. The two cards that I have are "incompatible" for nvlink though, since one is an FTW3 and the other an XLR8. So although the nvlink fingers line up horizontally, they do not line up in height. I estimate around 30mm of 180degree x16 riser could do the trick, then I would just need to fashion a way to mount it there, and then nvlink could reach. But I can't justify even continuing to think about that further, if little real world benefit is even to be had from nvlink.

From time to time I wonder about "SLI"ing them (the proper term for these appears to be NVLink them) for games. It sounds like RDR2 is a candidate for this, which is music to my ears as I haven't finished that game yet and getting a graphical upgrade would be nice, however my 4K120 TV is in the living room and I will likely never actually be able to physically bring these two devices within range of one another...

I'll give gravitymark a spin with this, I'll probably get a kick out of that. But I wonder if, since you are clearly knowledgeable in this area, if you could comment on, beyond games, whether NVLink connectivity can make a tangible difference. It seems that the majority of software is not written with such topology in mind sadly. Do you benefit from NVLink in the workloads involved with your dissertation? I'll guess not.

Even for stable diffusion the easy way to get it set up is to run a separate instance for each GPU and just juggle them manually. It is okay, but far from ideal.

If I had the kind of free time that I'll never have again, I would love to get back into building some physics engine stuff, with CUDA. The amount of compute we have on tap is truly mind boggling these days, we're about to enter the petaflop-under-your-desk era...
Sir,

IF I was going to buy 2 3090s I'd shoot for identical to avoid nv link misalignment -obviously I know you know this. IF getting a swap / return done for identical is not possible, you could mount them both in a small mining case (50 bucks on amazon) and force the alignment with brackets or use spacers in your rig and a pcie 4.0 riser cable . Unfortunately there are no flex / soft nvlink cables.

to answer your primary questions:

a large benefit is gained from the 3090s being abled to be SLIed / NVLink together. While there are some games / apps that can use separate GPUs, without SLI, there are more you cam use and or force to use. Mixed results, but in most cased a positive.

Overall, I've seen the benefit of SLI/NvLink , but one has to work at it. It requires using inspector, coolbits, and learning what to put into the profile. While this may not be an appealing reasons here are mu rounded outcome specs as of lately:


about 20% of games I use that are less than 5 years old, work in SLI well with huge performance boost without too much fiddling

about 40 % of games I use that more than 5 years old, work in SLI well with huge performance boost without too much fiddling.

D11 and below has a high success rate
OpenGL has a high success rate

Vulkan and DX12 are a more complex learning curve with mixed results.

Fiddling, profile editing, experimenting etc - results in getting mediocre performance into a significant performance increase in about 1/3 of the attempts.

on 3DGuru a user called BugBear has an extensive thread with many many SLI settings for many games and detail to those inspector profile settings and how well they work


Probably not all the answers you want but figured just lay it out

I'm currently using dual 4090s and have been experimenting in the no nvlink bridge world and doing some multiGPU projects, so I'm more fresh on those than the NvLink / SLI froma few years ago.

My SLI rig was handed down to my wife and I was able to get some of her games to work well.

MGPU benefit games she is playing:

RDR2
SOTTR
ROTTR
Borderlands

I was able to get cyberpunk to benefit, but I really had to roll up my sleeves and work on that profile to get worth out of it. Had the same issue with Witcher 3 and Dying Light 2

As a family we play multiplayer Borderlands, Quake, Doom, UT3, Killing floor and all those work well with SLI.

Other titles of benefit in SLI are
ID software games,
Ashes of Singularity (both SLI and separate GPU config)
UE4 and UE3 games
Capcom engine games (resident evil 7 etc)
Square Einix (sp) Batman series, Tomb raider series, Deus Ex
Metro series

I have had bad luck with Unity engine games and sli



WE don't STEAM,EA, Epic, etc so not all titles out there have been tried. We like our stuff to be real and not rented (steam) so I may not have the latest things to test. Usually we get it from GOG, not GOG galaxy, and use the DRM free download full file set. If a title is only available on steam like platform, I will buy the license for it but end up using a "DRM free variant". My conscience is clear since I bought it, but don't want Denuvo, Steam, Epic, EA etc as a requirement. I also buy the box game whenever its available. I like having the games and not renting them. Sad state of affair when a boxed game/cd is nothin more than as team link. Hence GOG is better if a title is available.

with all that said, dual 3090s is an awesome set up. However to get the most out of them they would need the nvlink bridge.

Check out BugBear on the forum at Guru3D


Hope that helps
 
Back
Top