NVIDIA GeForce RTX 2080 Ti SUPER Rumored To Feature 4608 Cores & 16 Gbps GDDR6 Memory

bill dennison said:
SLI/CFX was great when cards were 350 to 500 bucks each and you got 65%+ for adding a second card

very very few are going to spend 2400 USD on two cards plus 300+ on water blocks because two cards need water cooling .
so no developers will support it .

no normal SLI/CFX is dead .


now this is more likely to be a early MCM GPU needing a form of Multi-GPU support


https://wccftech.com/nvidia-hopper-gpu-mcm-leaked/
Click to expand...

MCM is still going to have to be supported by the developer. It's not going to be "Multi-GPU" as we know it. Unlike past iterations of dual GPU'S it's going to be like the Zen2 architecture.

My guess is you will have something like a 4 core GPU.

Core 1 - focus ray tracing or whatever.
Core 2 - focus AA methods.
Core 3 - focus shading
Core 4 - focus manches.

The memory will be pooled. Probably 24GB GDDR6 if not 7 or HBM ( doubt it ).
 
50% scaling is not impressive, especially not at the prices of GPUs now.

It's mediocre at best. There isn't much more room for improvement either. Both GPUs hitting 92%+ ..
 
Considered 80 percent scaling ideal; 65 percent average; 50 percent mediocre as well with Afr. Tiling should offer more smothness and less latency so afr and cfr are not apples-to-apples to me. Nvidia may use the gpu resources more efficiently as driver and feature matures.

I feel receiving 50 percent scaling with tiling without any DirectX 12 game developer support - respectable. It may breathe life and may resurrect Sli from a slow death. At least it is good to see some effort and not giving up.
 
Nunz said:
50% scaling is not impressive, especially not at the prices of GPUs now.

It's mediocre at best. There isn't much more room for improvement either. Both GPUs hitting 92%+ ..
Click to expand...

this

we used to get up to 90% and even 100% scaling in a few games


average was about 70%

when it worked and sometimes CFX worked best on some games other games it was SLI
 
Hapatingjaky said:
MCM is still going to have to be supported by the developer. It's not going to be "Multi-GPU" as we know it. Unlike past iterations of dual GPU'S it's going to be like the Zen2 architecture.

My guess is you will have something like a 4 core GPU.

Core 1 - focus ray tracing or whatever.
Core 2 - focus AA methods.
Core 3 - focus shading
Core 4 - focus manches.

The memory will be pooled. Probably 24GB GDDR6 if not 7 or HBM ( doubt it ).
Click to expand...

true but at that point both AMD and NV will be MCM most likely down to midrange cards so developer will have no choice

most likely have 2,3 & 4 core cards
 
Last edited:
bill dennison said:
this

we used to get up to 90% and even 100% scaling in a few games


average was about 70%

when it worked and sometimes CFX worked best on some games other games it was SLI
Click to expand...

The good old days when higher end gpu's were 500 and would garner 65-80 percent performance gains per generation cycle. Those days are hard to come by now. You're complaining strongly about 50 percent scaling yet paid 1200+ for similar performance gains over your 1080ti.


Off topic:

I asked Blaire about his subjective findings between cfr and afr, just waiting for a response.
 
Will do. I haven't talked to him in 2 months and that was a discussion about the quality of Nvidia 's raytracing denoising filters with a fast moving screen and more indepth discussions on how dlss handles artifacts in dynamic environments.
 
SIrPauly said:
The good old days when higher end gpu's were 500 and would garner 65-80 percent performance gains per generation cycle. Those days are hard to come by now. You're complaining strongly about 50 percent scaling yet paid 1200+ for similar performance gains over your 1080ti.


Off topic:

I asked Blaire about his subjective findings between cfr and afr, just waiting for a response.
Click to expand...

at 4k it was not a choice but a necessity the 1080 ti was craping out at 4k and I was having to turn off or turn down too much .


now with a 2080 ti strix SLI is a unneeded option at 4k
it would be nice but not worth the cost unless it worked in 80% of all games


now if SLI still worked well I would have just opted for a second water cooled 1080 ti and put it on top and the fan card on bottom
but that would have been around 900 at the time would have saved me about 450



……..

what would be great is if NV worked out a way to SLI a RTX 2080 ti with a RTX 3080 ti primary card when it comes out just for RT stuff

like using a old card for PhysX and give RT a boost
 
Last edited:
acroig said:
Post it when he replies to you.
Click to expand...

Here is his response:
Blaire said:
Hello Paule.

How are you? I was noticed while my work-break, you was online 3DCenter forum some time ago. I am not sure how are you interested in MultiGPU/SLI?
NVIDIA included "silently" a new render-mode "CFR" into the drivers.
I didn't get any official information about it because it's NDA, but since I discovered this mode myself in an official driver, it's also allowed to communicate, so it was ok for me.
It was very exciting for me, because it opens new possibilities and advantages for SLI and multiGPU in the future.
Better compatibility to TAA and post-process techniques. Less latency, won't have to pre-render any frames like it does with AFR.
Also Ultra Low Latency Mode finally got working too, i could measure less Driver-Latency output compared to SingleGPU in several games! Even in cpu-bounded games i have noticed the minimum frame rate are improved often , it could be the driver or API-overhead will be lowered with new CFR mode.

Hope also you like new Sharpness+Denoiser + Upscaling Resolutions directly in control panel.

Lets hear from you again.

Best Regards from Germany
Click to expand...
 
I may be talking out my ass:

One of the limitations of sfr and tile based render for multi-gpu was scaling for geometry, which afr does well in. What if Nvidia, through software engineering and nvlink have solved this limitation to some degree and can offer a very compatible cfr option that scales decently? Curious to know what kind of data is going through nvlink cfr?
 
Talking out ot my ass part 2:

Is it possible that Nvidia is thinking out-of-the-box with their cfr method by taking advantage of the mgpu frame pipeline ability and its potential flexibility? A wildly uninformed question but curious.
 
SIrPauly said:
I may be talking out my ass:

One of the limitations of sfr and tile based render for multi-gpu was scaling for geometry, which afr does well in. What if Nvidia, through software engineering and nvlink have solved this limitation to some degree and can offer a very compatible cfr option that scales decently? Curious to know what kind of data is going through nvlink cfr?
Click to expand...

would not a MCM with HBM do that better ?
 
I don't know but i think Nvidia is doing multi gpu frame pipelining, meaning the two gpu's are working on different workloads and abilities with nvlink supplying the speed, bandwidth, move closer to improve latency issues.
 
bill dennison said:
a lot of us didn't wait and paid sales tax

and will get the 3080 ti not the 3080 :hmm:
Click to expand...

I you can believe any of the stuff regarding the next gen cards NV are supposed to be reverting back to launching the 3070/80 variants first with a Ti version later so you might be waiting a long while. This could of course be all crap and even if true could change if AMD manage to produce a good high end card. Personally can't see AMD doing that :lol:
 
You must log in or register to reply here.
Back
Top