DLSSG

[Doffy]

So i have a RTX 3050 mobile, is a really tough card actually being capable of offer 6,9TFs without overclock, and about 7,7TFs, with both ram and chip overclocked (from core: 1695MHZ to 1875MHZ and memory: 5500MHZ to 6600MHZ; 180+, and 1100 MHZ respectively.)

And in advance, yeah i know that "You CaNt beCOuse doNT hAvE opTicAnaL Flof on 3000s", thats bulshit, since the 2000s series this cards have OFA, just a 2X performance bedtime story every 2 years, as usual.
But the 4000s series don't have anyshit new, just a better node and a better L2 cache, they don't even increase the performance per W in the mobile cards, forcing the battery designers to use meth to deliver a new battery model with 200W+ capability that don't melt the entire rooms olner; or yet increase the bandwidth in the NEW cards, sounds lazy job for me.

As such, I PERSONALLY THINK, that to prof that given software cannot run, even delivering a 30% performance increase in the LAST GEN cards, a "ethical" company, should deliver that specific software to the consumers themselves test (I CAN SWALLOW INTEL OR AMD REMAKING AND USING NEW ARCHITECTURES AND NODES IN A TOTALLY NEW ARCHITECTURE, BUT YOU CHANGE A TRANSISTOR OR TWO AND SPELL THAT YOU REMAKE A SINGLE COMPONENT SO MUCH THAT WILL RUN LIKE A R3000 RUNNING CYBERPUNK, I DON'T!), so if anyone please has any means to get it, please share your software/hardware mage knowledge.

I'am open to discuss the AD1xx new lazy architecture, i think that making things smarter lead to evolution, but by any means i thing smarter = lazy, contrariwise i thing that the smarter work should work even harder than the average work to deliver smarter massive output, i know that the bonobos at npimba office will not get a bonus doing this but, if they get, by the personal and corpo side you should not carry this as a flag to be proud about it.

 

[Deleted member 182555]

read this...

NVIDIA Ada's 4th Gen Tensor Core, 3rd Gen RT Core, and Latest CUDA Core at a Glance

Yesterday, NVIDIA launched its GeForce RTX 40-series, based on the "Ada" graphics architecture. We're yet to receive a technical briefing about the architecture itself, and the various hardware components that make up the silicon; but NVIDIA on its website gave us a first look at what's in store...

www.techpowerup.com

 

[Doffy]

read this...

NVIDIA Ada's 4th Gen Tensor Core, 3rd Gen RT Core, and Latest CUDA Core at a Glance

Yesterday, NVIDIA launched its GeForce RTX 40-series, based on the "Ada" graphics architecture. We're yet to receive a technical briefing about the architecture itself, and the various hardware components that make up the silicon; but NVIDIA on its website gave us a first look at what's in store...

www.techpowerup.com

DLSS 3.5 isn't the same? just delivering the same rate frame instead of a "doubled"?

 

[ir_cow]

DLSS 3.5 is Ai Ray-tracing information generation, where as DLSS 3 was the entire frame. So it is still frame generation, but just for Ray-tracing data.

That's how the marketing material reads to me.

 

[Deleted member 182555]

DLSS 3.5 isn't the same? just delivering the same rate frame instead of a "doubled"?

I'm not an expert on this also because there is no information on how ofa is implemented. I assume they make use of less accurate but faster fp8 format to speed up calculations and inference.

 

[ir_cow]

The information is in the white papers. Need media press access to see it.

 

[Deleted member 182555]

https://images.nvidia.com/aem-dam/Solutions/Data-Center/l4/nvidia-ada-gpu-architecture-whitepaper-v2.1.pdf

https://images.nvidia.com/aem-dam/Solutions/geforce/ada/ada-lovelace-architecture/nvidia-ada-gpu-science.pdf

ofa = hardware piece

 

[GerKNG]

"DLSS3 frame sequencing wont be in Ampere and older as there is just so much wrong with it to try to do realtime frame interpolation using motion vectors and such. ADA takes one clock cycle to use the Tensor cores and then get data from the Tensor cores to the OFA while Ampere and older takes tens of thousands of clock cycles to do the same. Ampere and older cant get the Tensor data to the OFA after its done its calculations in the same clock cycle or without software help. The data also needs to be organized and blocked out which requires more software help and many more clock cycles. The OFA also prefers low fidelity data rather then high fidelity data when doing per frame sequencing and only ADA has low fidelity FPUs in their Tensor cores. ADA is also the only architecture to have a high enough Tensor throughput to do per frame sequencing. Last issue is with Turing, that is also just missing OFA "featuresets" which is described in the OFA SDK documentation"

 

[Dr. Dro]

While there is merit to the claim that DLSS frame generation would probably work in Ampere and Turing - within their performance limitations of course - and considering that FSR 3 will work, the RTX 3050 mobile is literally the weakest graphics card of the past generation. It seems that you are greatly exaggerating its capabilities, it's barely faster than the desktop GTX 1650 in terms of raw performance. It does put up a fight - my laptop has a 80-watt version of it and runs at quite a bit faster clocks than those that you have claimed your laptop can do (2100 core/1250 mem) but I wouldn't call it anywhere close to "tough" or "high performance", it's a really basic GPU for low-cost laptops. And again, teraflops mean absolutely nothing.

All in all, either expect a reaction from Nvidia backporting these to older generation GPUs, or wait for FSR 3 and be thankful that it can run that at all.

 

[oxrufiioxo]

While there is merit to the claim that DLSS frame generation would probably work in Ampere and Turing - within their performance limitations of course - and considering that FSR 3 will work, the RTX 3050 mobile is literally the weakest graphics card of the past generation. It seems that you are greatly exaggerating its capabilities, it's barely faster than the desktop GTX 1650 in terms of raw performance. It does put up a fight - my laptop has a 80-watt version of it and runs at quite a bit faster clocks than those that you have claimed your laptop can do (2100 core/1250 mem) but I wouldn't call it anywhere close to "tough" or "high performance", it's a really basic GPU for low-cost laptops. And again, teraflops mean absolutely nothing.

All in all, either expect a reaction from Nvidia backporting these to older generation GPUs, or wait for FSR 3 and be thankful that it can run that at all.

I can confirm a 3050ti at 1080p barely runs diablo and you can run that on a potato.

 

[ir_cow]

I don't see DLSS 3X ever running on 20/30 series.

 

[oxrufiioxo]

I don't see DLSS 3X ever running on 20/30 series.

Nvidia did a terrible job naming all this technically Upscaling, FrameGen, and Ray reconstruction all fall under the same umbrella it's just the FrameGen part that Ampere, Turing cannot do, the upscaling portion is already at DLSS 3.5.x


As far as Ampere/Turing ever getting a form of frame generation I think that depends on how good the AMD version is if it's pretty good I wouldn't be surprised if Nvidia figures it out with a miraculous newspost with how all their hardwork and determination lead to them being able to yada yada yada.....

 

[ir_cow]

It's just like CUDA versions, but for Tensor cores.

 

[Mazer]

I can confirm a 3050ti at 1080p barely runs diablo and you can run that on a potato.

Diablo IV sure..

OG Diablo/II without any problems.

 

[oxrufiioxo]

Diablo IV sure..

OG Diablo/II without any problems.

Yes it can play 20+ year old games I can confirm.....

 

[Dr. Dro]

Diablo IV sure..

OG Diablo/II without any problems.

It's a champ for what its market seems to address: League of Legends and other F2P eSports, with a tiny bit of Apex if you can play with a 80-100fps target (Apex requires a beefier PC than other eSports games). These games you mentioned are quite old by now so of course they run well. It'll do light gaming just fine, just don't hope for Starfield on Ultra on it (which was the OP's tone).

I don't see DLSS 3X ever running on 20/30 series.

The only DLSS subfeature that requires 40 series is frame generation, 3.5's ray reconstruction works on Turing and Ampere as well. NV may or may not backtrack on FG exclusivity, but perhaps what I was getting at is that @Doffy needs a reality check, it's a 3050 we're taking about. Even if made compatible, that "tough" GPU of theirs is but a mere notch above AMD's strongest integrated graphics.

 

[ir_cow]

The only DLSS subfeature that requires 40 series is frame generation, 3.5's ray reconstruction works on Turing and Ampere as well.

How do we know it works on Turning and Ampere?

 

[ViperXTR]

The thing about framegen is, the slower the card the less ideal it becomes (ideally you need base framerate of 60+ to minimize latency). I already knew when it was first demo and tested by the first users but the latency is real when i actually experience it when i got my ada card. If you expect playing 30fps boosting to 60fps with framegen and be done with it, sadly it will still feel like 30fps, this is more noticeable on fast paced twitchy games fighting games or first person shooters. Things like nvidia reflex and its AMD counterpart hyper-rx helps but if you are used to high framerate gaming you will still feel it.

 

[ir_cow]

Not how math works. 60fps is 16.6ms. 30fps is 33.3ms between frames.

 

[ViperXTR]

Not how math works. 60fps is 16.6ms. 30fps is 33.3ms between frames.

yeah, im just setting up a random example of sorts

 

[ir_cow]

Okay I think I get what you're saying. If the source frame is 33.3, The input latency will be the same?

 

[oxrufiioxo]

Okay I think I get what you're saying. If the source frame is 33.3, The input latency will be the same?

slightly higher but similar.

 

[ViperXTR]

yeah slightly higher but some users will immediately notice the variance. Im not a pro gamer but i do like fast paced games and i can feel it. Not game breaking but can be a matter of win or lose for some pro gamers.

 

[oxrufiioxo]

It's way more noticeable on mouse and keyboard than controller.

 

[Dr. Dro]

How do we know it works on Turning and Ampere?

Nvidia said DLSS 3.5 RR is supported on 20 and 30 series. Just FG isn't