How did SLI die?
Not so fast anonhttps://overclock3d.net/news/gpu-displays/dlss-5-sli-modder-runs-dlss-5-on-2nd-gpu-for-massive-fps-boost/
>>109760177Originally the idea was that pcie is good enoughTurns out that it wasn’t but by the time the next gen was cooking ai started becoming a thing so there was no way njudea was putting a proper interconnect on consoomer gpus. It’s a good thing, too, otherwise you’d be competing for every dinky 16gb 5070ti with small sized corpos willing to throw $100k on a local setup
>>109760207>communication via PCIE is SLIhuh?
>>109760177it never didthey just priced you out to not cannibalize their premium offering
>>109760210They actually already had it. NVLink allowed vram pooling and cards to work together. But it was deliberately disabled in drivers. Then NVLink was physically removed in Ada/4000 series. They Jewed the cards on purpose even before AI.
>>109760177SLI had to die so Jensen could charge $250k for DGX with nvlink.
>>109760177It was not very good, the only real use for multiple GPUs now is mocal LLMs as toys and they don;t care if you have nvlink or sli and work with multiple GPUs well anyway. For gaming what was the point a 1080ti does everything you need anyway.
>>109760177on the account that It was basically useless i'm hesitating between "naturally" and "obviously". maybe both is the way to go.
>>109760177>pay for 2 GPUs >receive 40% performance boost
>>109760371>Get xx90 performance for 60% of the cost of today's xx80Fixed
>>109760177consoomer retards thinking they get 100% more performance on one handdogshit developers unable to code for two gpus on the other handperfect subhuman combo for ending a technology
>>109760451You can do better than a 5090 for LLMs with any combinationn of older 24GB cards because really VRAM is all that matters. Fuck you can run local models on a 1080ti. I'm not into AI slop 'art' or crappy youtube scamming with one and don;t see the point in gayming given that there is nothing compelling for me to use a 5090 in. I don;t like the 5070/80/90 cards much as I do bits of work for a local compuyter shop and have now seen four with melted cables and one in a burnt out case that destoyed a bedroom. Just not worth the money or risk. I put more effort into making perfect metal XP and Win7 machines for gayming as I like their game libraries a lot ad have a 5060TI and two 1080tis.I went back from using a 5060 ti 16GB on an 15-14400 to a 1080ti on a i74790 with 32GB of DDR3 becasue really, it is in my experience better for anything I care about. Playing around with models even, it generates faster than reading speed which is all I care about. That's one of the problems something like an owner of a 5080 with 128GB of DDR5 has, there is not much use for it except LLM and generative AIslop and anyone who spends much time on LLMs will eventually figure out they are garbage anyway, there are no games or tech for that kind of machine that are in any way comelling or interesting and no matter how hard I tried I can;t give a shit about anything more than 60-100FPS at 1080p even if I want to use it on a 50 inch OLDED TV, they have been upscaling fine ever since the xbox360. People who have not really dug into LLMs will think 'if only I had a 5090 or massive vram or whatever' this will finally get good and it won't (maybe you will generate slop AI images or video faster but unless you are a jeet spamming videos on yourtube who the fuck cares), aside from that you will just have an LLM and theya re all shit from the latest cloud fronntier models down and they are shit because of the way they work, their design.
>>109760749whether old SLI setups that cost a fortune or modern 5090 setups that cost a fortune the market segment is too small for any game developer or media creator to target it.
>>109760306>NVLink allowed vram poolingIt did not. That was specifically blocked at the driver level. Even SXM doesn't make it so you can pretend all the linked GPUs share a unified memory space for tensors. Many tasks still need all the tensors loaded on a single GPU, and can only parallelize things at the inference or training step level.
>>109760177>How did SLI die?Horrendous frame pacing issues
>>109760242It's not specifically SLI in that link, but later AMD crossfire DID use the PCIe bus instead of a dedicated lower bandwidth link.
>>109760177I've tried it with multiple generations of nvidia GPUs and it always felt like an afterthought that they never properly polished.
>>109760177It still exists, it just doesn't work for real time workloads
>>109760177It costs twice as much but doesnt perorm twice as well, and it always had problems with microstutter.
>>109760177Frametime benchmarks exposed to the masses what SLI/CF users knew. SLI/CF was only good for big FPS avg figures that completely masks the shuttering.
>>109760177painfully
>>109760177insane latency therefore useless for anything realtime
>>109760177wouldn't surprise me if people were encouraged to use a secondary GPU for their AI post-processing needs while they rasterize on their primary in the future
>>109760918This. Doesn't anyone remember all the complaints about 'microstutter'? I remember people were pulling their second card out to get rid of the issue.
