• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Gamer Pairs RTX 3090 With RTX 3050 to Offload Frame Generation Work

erek

erek

8=D
2FA
Joined
Dec 19, 2005
Messages
18,192
“Could dual-GPU PCs make a comeback thanks to AI enhancement?

Although something like this is usually only worth it if you already have both cards, in this case, the user already had the 3090 and was able to buy the 3050 second-hand for a mere 80 euros. Although it's possible they could have achieved the same frames per second without spending that amount or by leveraging a second card, there's no telling how the 3090's native frame rate would have been affected, potentially degrading the overall experience.

If this marks a new age of multi-GPU setups, I'm here for it. With single cards being so expensive, leveraging a second one to improve performance would be great. Considering Nvidia even did this with its DLSS 5 demonstrations, maybe it's not so far-fetched a future after all.“

Source: https://www.pcmag.com/news/gamer-pairs-rtx-3090-with-rtx-3050-to-offload-frame-generation-work
 
Oh with Lossless Scaling

I thought they did it with official DLSS

Would be cool if you could get a 3090 for rendering per say, and a 5050 ect for later upscaling/DLSS methods not supported by the big card

But that also runs counter to a company incentivizing new (main/big) products and purchases 🤷
 
There are so many variables that go into having a good gaming experience that I bet this setup would be awful to actually live with. But cool concept.
 
This is one of the recommended ways to use Lossless Scaling. I guess slow news day at PC Mag.
 
I have an old 3090 laying around form when I upgraded to the 5090 I wonder how that set up would work.
 
tunatime said:
I have an old 3090 laying around form when I upgraded to the 5090 I wonder how that set up would work.
Click to expand...
You'd probably need another outlet/circuit to try it. Unless you're on 20a wiring. Pretty sure those two would pop a breaker otherwise. Ha.
 
SticKx911 said:
You'd probably need another outlet/circuit to try it. Unless you're on 20a wiring. Pretty sure those two would pop a breaker otherwise. Ha.
Click to expand...
Y'all didn't pull a dedicated 20 amp 220 for your PC yet? It's worth it just to get a tiny bit more efficiency out of the PSU. Also helps if you do it for a living and have spare wire and get tired of the 15 amp that feeds 2 bedrooms tripping lmao
 
tunatime said:
Y'all didn't pull a dedicated 20 amp 220 for your PC yet? It's worth it just to get a tiny bit more efficiency out of the PSU. Also helps if you do it for a living and have spare wire and get tired of the 15 amp that feeds 2 bedrooms tripping lmao
Click to expand...
I did but I probably should have done 2.
 
erek said:
“Could dual-GPU PCs make a comeback thanks to AI enhancement?

Although something like this is usually only worth it if you already have both cards, in this case, the user already had the 3090 and was able to buy the 3050 second-hand for a mere 80 euros. Although it's possible they could have achieved the same frames per second without spending that amount or by leveraging a second card, there's no telling how the 3090's native frame rate would have been affected, potentially degrading the overall experience.

If this marks a new age of multi-GPU setups, I'm here for it. With single cards being so expensive, leveraging a second one to improve performance would be great. Considering Nvidia even did this with its DLSS 5 demonstrations, maybe it's not so far-fetched a future after all.“

Source: https://www.pcmag.com/news/gamer-pairs-rtx-3090-with-rtx-3050-to-offload-frame-generation-work
Click to expand...
People been doing this with FSR frame gen and lossless scaling, for like a year or more.
You can use optiscaler to inject FSR framegen into a lot of games.
So, systems which dont support ML based framegen, can benefit. And using a second card helps maximize the base framerate.

I think Arrowlake's iGPU may be powerful enough, you could skip the 2nd DGPU, for FSR.

Intel's next desktop CPUs will support ML framegen. So, could be an interesting opportunity, there.
 
With PCIe 4/5 bandwidth, this might work really well.

PCIe is bi-directional simultaneously. PCIe4 can be transmitting 31.5GB/s and receiving 31.5GB/s at the same time.

And most videos where they investigate the impact of using older PCIe generations, it is usually 1.5% to 2.5% impact. Which is mainly pointing out that GPU's never use the full bandwidth of PCIe.
So, I think it could work.
 
I wished they brought back SLI and Crossfire. Two RX 9070 XT's in Crossfire would easily beat an RTX 5090 24GB in performance in 4K, but at 1/4th the price...
 
chameleoneel said:
I think Arrowlake's iGPU may be powerful enough, you could skip the 2nd DGPU, for FSR.
Click to expand...
I kinda doubt that. I played with one when I was building my rig. I figured I'd built it without a vid card, make sure everything worked, then move my card over. Ran a couple games before installing my 3090. 4 Xe+ cores don't have enough grunt to really even do upscaling. Felt like XeSS was kinda capped at ~30fps. I actually ran into a couple of instances where upscaling lowered fps slightly. Lunar Lake with 8 Xe2 cores would be a lot better, then Panther Lake with 12 Xe3 is even better.

Nova Lake is supposed to have some big iGPU versions, so that might work. OTOH I bet you won't be able to get all the features on one proc. Big iGPU might be laptop models only, and even if it isn't I bet they won't offer the big cache and big iGPU on the same proc. Then the dual compute tile version probably won't be available with the big iGPU either.
 
El_Capitan said:
I wished they brought back SLI and Crossfire. Two RX 9070 XT's in Crossfire would easily beat an RTX 5090 24GB in performance in 4K, but at 1/4th the price...
Click to expand...
And that right there is the real reason leatherjacketman killed SLI.

If you look at the texture fill rate to estimate raster only performance, it takes a minimum of a RTX 4080 to beat a pair of Titan XPs in SLI it with a single card.

If you only play older stuff and don't care about ray tracing, the used prices of 1080TIs and TitanXps are very attractive.
 
TXE36 said:
And that right there is the real reason leatherjacketman killed SLI.

If you look at the texture fill rate to estimate raster only performance, it takes a minimum of a RTX 4080 to beat a pair of Titan XPs in SLI it with a single card.

If you only play older stuff and don't care about ray tracing, the used prices of 1080TIs and TitanXps are very attractive.
Click to expand...
Careful what you wish for. With all this ai stuff they are running on cards now I wouldn't be surprised if we end up with two cards with one doing all the AI work. Also don't forget sli sucked in a lot of games. Worked amazing in others and barely made a difference in some. Also looking back at it it felt more fake than the AI fps nowadays feel. If I remember right you needed more then twice to performance for the fps to feel smooth. I had a 8800gtx sli rig 40 or 50s some ati in cross fire and to this day I remember none of them feeling amazing even though the fps might go form 45-70 it sometimes felt worse l.
 
Crossfire and SLI worked great for upping graphics quality without sacrificing FPS. For flat-out higher frame rates, that's where it didn't help too often.

At the time, bigger numbers and charts were king, and quality and effects weren't popular or even relevant metrics.
 
No doubt SLI has its pitfalls. Coupled with the SpecialK injector, SLI works extremely well in Trainz 19 at 4K 50Hz. In this application, latency doesn't matter much and 50 FPS is enough. Absolutely smooth and 2x the performance of a single card. SLI has always had a big YMMV attached to it, but it does not always suck.

My post was more of a criticism of how poorly the 4090 and 5090 scale with the older generations - and we add in cost, they look really, really poor. I'd prefer to be able to buy a single 5080 for a reasonable $500-$600.

I really do think Nvidia dumped it to hide the relatively poor raster improvements in the RTX lineup.
 
SLI died because why would Nvidia support someone buying an older card to boost their performance?

If in an alternate universe, Nvidia supported and optimised Multi-GPU so it worked flawlessly, they would essentially make it so that a 3080 owner today could buy another 3080 on eBay and get the performance of a 5080,and Nvidia wouldn't see a dollar of that.

And with Nvidia being 90% of the GPU market, their biggest competitor is themselves from the last generation. Supporting someone buying an older card is literally supporting their biggest competitor.

And because AMD Radeon is a snivelling cuck they just followed suit.
 
KazeoHin said:
SLI died because why would Nvidia support someone buying an older card to boost their performance?

If in an alternate universe, Nvidia supported and optimised Multi-GPU so it worked flawlessly, they would essentially make it so that a 3080 owner today could buy another 3080 on eBay and get the performance of a 5080,and Nvidia wouldn't see a dollar of that.

And with Nvidia being 90% of the GPU market, their biggest competitor is themselves from the last generation. Supporting someone buying an older card is literally supporting their biggest competitor.

And because AMD Radeon is a snivelling cuck they just followed suit.
Click to expand...

https://hardforum.com/threads/why-did-sli-and-crossfire-die-off.2007466/

Dan_D said:
In the past, SLI or Crossfire support in games was primarily up to NVIDIA and AMD to implement in their drivers via profiles. With changes to DirectX 12, its no longer up to the GPU makers. The game developers have to implement the feature in the games themselves. Given that most games on PC are ports of their console versions and given how few people utilize multi-GPU configurations the developers never bother to implement it. It's to the point now where AMD and NVIDIA cards don't even have the physical capability in hardware anymore. I think the only current high end card that does is the RTX 3090. It does via an expensive and proprietary bridge, but all the same limitations apply to it as well. Basically, it works in a few games but the scaling isn't what you'd want or hope for. It's just not worth doing.

The reason you don't see dual-GPU graphics cards anymore is simple. These cards all relied on some form of internal Crossfire or SLI between the GPU's. How well these GPU's worked was entirely dependent on SLI/Crossfire profiles in the drivers. Some game engines and some architectures also benefited more or less from this than others. To say it was inconsistent is putting it mildly.


Most dual GPU cards down clocked their GPU's to keep the TDP's in a reasonable place to allow two to exist on the same board without drawing too much power. You could functionally do the same thing today so long as you never drew more than 75w from the PCIe bus. You would simply have double the amount of power connections going to the card. However, the big problem here isn't power as much as the cooling solution would be. Something like a dual RTX 3090 card would almost have to be watercooled or the clocks would have to be nerfed considerably.
Click to expand...

Dan_D said:
Not everyone was sensitive to microstutter and some of us would take it as a trade off to get overall better performance. This was especially important at higher resolutions. Even today, an RTX 3090 FE can't handle everything at 4K the way I'd like it to. DX12 is on the other hand a huge reason why multi-GPU died.

This is simply not true. SLI and Crossfire waned in popularity back when GPU availability wasn't a problem. Pascal and later architectures simply showed worse GPU scaling than some earlier architectures. You combine that with DX12 and developers that wouldn't put in the time to support it, and you have the reason why multi-GPU died out. I literally bought GPU's in pairs on launch day or near launch day all the way up through the GTX 1080 Ti without issues. It was never an availability thing.

As I said, I had multi-GPU solutions for each generation all the way from the 6800 Ultra days through the GTX 1080 Ti. Microstutter wasn't the problem. In fact, no one had heard of it until several generations after it had been on the market. Those of us who had it and used single- GPU setups were often aware of it (but had no name for it) but you take the good with the bad when you are trying to run a high resolution / multi-monitor setup with the latest games. It was the only way to do it. The 50% boost by the second card wasn't the issue though. It was really a matter of fewer titles supporting it. I would agree with you that it was only ever worth it on the ultra-high end as two GPU's were able to do what one GPU could not.

Multi-GPU is not hard to support from a hardware perspective. Motherboards today still support it at various price points. You literally only need to have two PCIe x8 slots (NVIDIA) or an x8/x16 + an x4 slot (AMD). Power requirements aren't really a barrier either since most enthusiasts who would consider running such a thing would make sure to have a power supply that could handle it first. I size my PSU's to do just about anything and could add a second RTX 3090 FE without even thinking about that. Support can't be done by the GPU designers in drivers. That's the problem. With DX12 and Vulkan, it was left 100% up to the developers to implement in their titles. They don't do it for obvious reasons. The market for it is almost non-existent and it takes time, testing etc. to implement and most games are muilti-platform these days. They simply have other priorities.

Indeed. There are still games out there that do support multi-GPU today. They are few and far between. It isn't just a matter of the engines supporting it, although you can often create a profile yourself (NVIDIA only) and get it to work reliably, although its not perfect. The game developer still always had to do some work to make multi-GPU behave correctly. That's simply not done even in cases where the engine would still support it.
Click to expand...
 
KazeoHin said:
SLI died because why would Nvidia support someone buying an older card to boost their performance?
Click to expand...
Nah, it was because it never worked all that well, and it only got worse as time went on. Even in the Voodoo days of "scan line interleave" there were issues. It could work good, I played Quake Team Fortress in glorious 1024x768 with SLI'd Voodoo2s but you could find plenty of games with issues or little to no speedup. It only got worse later with newer tech where you couldn't split it per scan line as they did and had to do tings like AFR which lead to lots of micro stutter. Devs also totally stopped supporting it because not enough people had it.

I think there's a bit of misunderstanding/cope to think it would be an easy task to make it work flawlessly. I don't think it would be possible. There's a lot of issues with multi-chip sync which is why it isn't done. Really the only area where it is going to work "flawlessly" is for SSAA. If you render the same scene, on multiple cards, with different subpixel offsets, you can easily combine that in to a finished image. That's great and all, but are you willing to double your cost for 2xSSAA, particularly given how good DLSS is? Same thing would be true with Path Tracing, you could cast two rays per pixel instead of one which will improve how things look but with literally double the cost.

Also, it would make things WORSE these days with the RAM crunch since each card needs a complete copy of all visual data. So if a game needs 16GB VRAM, and you want 2 cards, that's 2x16GB cards not 2x8GB.
 
Lots of software stuff to figure out to make it work. If only we had some sort of powerful technology that could do software for us. Maybe even make that technology go faster by adding more graphics cards.
 
KazeoHin said:
SLI died because why would Nvidia support someone buying an older card to boost their performance?

If in an alternate universe, Nvidia supported and optimised Multi-GPU so it worked flawlessly, they would essentially make it so that a 3080 owner today could buy another 3080 on eBay and get the performance of a 5080,and Nvidia wouldn't see a dollar of that.
Click to expand...
While NVIDIA did indeed market SLI this way, this isn't how it was primarily used. People would often buy two mid-range cards to equal or exceed the performance of higher end cards because they could buy the cards one at a time, or they would get higher frame rates with both cards. What they failed to realize is that those mid-range GPU's often had limitations with resolution due to a lack of VRAM compared to higher end offerings. They didn't just have less VRAM, but it was often slower or had a slower bus. Guys like me did things differently. We bought two of the fastest cards available at the time to get what was effectively next generation GPU performance today. Despite its issues, it normally worked well but as time went on it seems like the latency issues got worse. Not only that, but GPU prices climbed very high, power consumption, heat, etc. went up and support for it in games died off. This was due to developers doing "multi-platform" releases for everything and prioritizing consoles. They often didn't test SLI given the general adoption rates weren't that good, so it never worked right on some games. Some engines didn't scale particularly well with it either.

Then Microsoft changed the game, shifting the burden of this from the GPU makers who had to build profiles for games to the game developers themselves. Game developers had to leverage things within DirectX like explicit multi-GPU etc. to make it work. Frankly, its a shame developers never did this as there were different options within DirectX for multi-GPU that could have made it better, but they never wanted to put the work in. Frankly, it's understandable on the basis that the percentage of multi-GPU systems had been falling for years, and never saw widespread adoption in the first place.

But that's what killed multi-GPU. Not this idea that NVIDIA was against people buying older GPU's to run them in pairs to equal more modern GPU's. This was frankly never really a concern as the older GPU's would be phased out of the retail market almost immediately, making it difficult to use that as an upgrade path. Sure, there is a secondary market for older GPU's, but people buying them for that purpose was never a huge issue. Not only were there various concerns in doing that on a technical level, but you were stuck on older technology. Technology NVIDIA made sure to market the hell out of on newer GPU's. So you could go ahead and buy a second older GPU, but most people sold that GPU to help fund the purchase of a newer GPU. If you had SLI from the beginning, selling both cards at the right time paid for the first of your new ones.
 
Also to add to what Dan said: Those generational upgrade were often not trivial. Lately things have slowed down a bit, both in terms of new technology with generations and how fast new generations come out but for most of the time when SLI was a thing, new cards came fairly often that brought big feature upgrades. Getting a second nVidia 7 series anything wasn't going to give you the massive DX10 changes the 8800s did so the idea of "buy another card later to speed up performance" wasn't the best idea since you would get left out on new graphics.
 
Sycraft said:
Also to add to what Dan said: Those generational upgrade were often not trivial. Lately things have slowed down a bit, both in terms of new technology with generations and how fast new generations come out but for most of the time when SLI was a thing, new cards came fairly often that brought big feature upgrades. Getting a second nVidia 7 series anything wasn't going to give you the massive DX10 changes the 8800s did so the idea of "buy another card later to speed up performance" wasn't the best idea since you would get left out on new graphics.
Click to expand...
Indeed.
 
I wasn't messing around with SLI at the time, but didn't the way it used to go is within 2 generations the fastest new card that was reasonably priced was faster than the fastest cards in SLI from 2 generations back? If true, a single 3080TI should have beat a pair of 1080TIs. The post below seems to confirm this:

Sycraft said:
Also to add to what Dan said: Those generational upgrade were often not trivial. Lately things have slowed down a bit, both in terms of new technology with generations and how fast new generations come out but for most of the time when SLI was a thing, new cards came fairly often that brought big feature upgrades. Getting a second nVidia 7 series anything wasn't going to give you the massive DX10 changes the 8800s did so the idea of "buy another card later to speed up performance" wasn't the best idea since you would get left out on new graphics.
Click to expand...

I know I was never interested in SLI until running into performance limits and the crazier and then crazier graphics card pricing over the last 10 years. The background information posted above looks interesting and I'll read through it.

If you read the documentation on Special K apparently a lot of games get the graphics pipeline wrong. Trainz 2019 does not work in SLI without this injector to force the graphics into hardware independent flip mode. Then there are a couple of DX10-11 SLI compatibility bits to set in Nvidia Profile Inspector. After that, the game runs in SLI and has 2x the graphics performance in parts of the map that are GPU limited, utterly smooth.

Then there are titles like DOOM2016 that don't work in SLI at all, but runs well at 4K on a single 1080TI anyway.

It's clear Nvidia is up against a performance wall, likely part of the reason for them going the ray tracing route and then AI revenue path is so much more alluring than this gaming business.
 
TXE36 said:
I wasn't messing around with SLI at the time, but didn't the way it used to go is within 2 generations the fastest new card that was reasonably priced was faster than the fastest cards in SLI from 2 generations back? If true, a single 3080TI should have beat a pair of 1080TIs. The post below seems to confirm this:
Click to expand...
It often did. Not only becuase of large generational improvements, but also because you rarely saw 2x performance. If it was a situation of "2x performance, all the time" then maybe it would have been of more interest. However games were often more like 1.5x, or even 1.2x (or works case no speedup at all).

Also something like this is where it is a good example of how even if it WAS 2x, it wouldn't be worth SLIing to try and make it up because of features. The 1080Ti was a beast and still is a good card, however it lacks RT cores and more importantly tensor cores. So no DLSS. Right there, that's a huge loss and would push the 3080 further ahead since it could effectively get more FPS (and generally better anti aliasing) using that. Likewise raytracing would run waaaaay slower, if at all, on the 1080s since they lack RT cores. So any RT title the 3080 would pull way ahead.

It just wasn't a great upgrade path, so not a lot of people used it like that. You could do it, some people did, but realistically it was better to just save your money, not get a second card, and get a newer card. That left only people like Dan that wanted to get the absolute best performance despite the cost. Well, not a ton of those out there. Too expensive, particularly since it came with downsides. Like with the 90 level cards, at least you always get the benefit. If you buy a 5090 it is 30% faster (or more depending on) than a 5080. There aren't some games that just don't support the extra power, it just is faster, all the time. There no micro stutter, no compatibly, it is just faster. Likewise you get double the RAM, you have 32GB of RAM available for insane mods or whatever, it isn't 32GB being split, it is all available to every title. Hence it is a more interesting upgrade for the "ultra high end" people than 2x an 80 level card.

So a very tiny market. I'm sure there would still be people who would do it, if they could, but it would be VERY few. Just not worth putting in the support for because there WAS a cost in supporting it. More hardware on the boards, more complexity in drivers, it wasn't free to add.

TXE36 said:
I know I was never interested in SLI until running into performance limits and the crazier and then crazier graphics card pricing over the last 10 years. The background information posted above looks interesting and I'll read through it.
Click to expand...
Again, it would likely be even crazier, because of the additional cost of the hardware. There's a cost in making a board, but the real cost is RAM and that doubles for no gain. The second chip you can look at as a cost that gets you something, more performance, potentially up to close to double. However because of the speeds involved, each board needs its own RAM and each RAM needs to have a copy of all geometry, all texture data, etc. You don't get any savings. So like I said: Game needs 16GB? Each card needs 16GB. You have 32GB of RAM, but only get 16GB of effective RAM, you are basically RAID-1ing your RAM. Not very cost effective, so getting performance level X with SLI is always going to be more expensive than getting it single card if that's an option.

TXE36 said:
It's clear Nvidia is up against a performance wall, likely part of the reason for them going the ray tracing route and then AI revenue path is so much more alluring than this gaming business.
Click to expand...
Everything is, lithography isn't scaling well anymore. Costs are going waaaaaaay up because of how hard EUV is to do, and even then, there's not a lot more steps down in size we can make on it, and we aren't sure if they are going to be commercially feasible. The big performance gains we used to get were by shrinking the size of transistors, without costs going up a lot, which meant throwing more transistors at a problem for a cost that was more, but not a ton more. That's just not happening anymore. Means chips are big and hot to get more performance, and we are reaching the limits of that too. Like, would you really want 2x 5090s if that were an option? Really? A system where you needed a dedicated 20a plug just to run the computer?

That's why there's a focus on things like DLSS: Figure out how to do more with less. Render less pixels, still get a good result. We can't just throw more rendering power at the wall and expect people to buy it. I mean even get rid of the AI craze: $2000+ for a GPU is a ton of money, way too much for the vast majority of gamers. If graphics are going to get better for the mainstream, if has to be done with less cost which means less silicon which means less raw horsepower.
 
You must log in or register to reply here.
Back
Top