News Nvidia's latest DLSS revision reduces VRAM usage by 20% for upscaling — optimizations reduce overhead of more powerful transformer model

[Admin]

Nvidia's latest version of the DLSS SDK provides VRAM optimizations to the transformer model, reducing its memory footprint by 20%.

Nvidia's latest DLSS revision reduces VRAM usage by 20% for upscaling — optimizations reduce overhead of more powerful transformer model : Read more

 

[palladin9479]

Slowly chiping away at the ungodly VRAM requirements for DLSS / MFG.

 

[fiyz]

Eh? I'm still adamant about giving us realistic audio. Visual effects is how the wizard controlled oz, but the audio effects gave him away.

 

[usertests]

Eh? I'm still adamant about giving us realistic audio. Visual effects is how the wizard controlled oz, but the audio effects gave him away.

Audio has long been a pain point in game development. It's harder to market than fancy graphics. We see things like AMD TrueAudio, or Tempest in the PS5, but we don't hear about them. 😉

Ideally, sounds could be simulated in real time (e.g. below 5 milliseconds). For example, if you throw a metal wrench at a granite surface at 2 m/s, at a certain angle, it starts clattering, and the sounds are accurately reflected or absorbed off nearby walls, objects, the hallway around the corner, etc. Sounds difficult, maybe it isn't anymore but I don't know.

The only games that can easily justify spending a lot of attention to sound design are stealth games. Games like Thief come to mind, but RPGs like The Elder Scrolls also use stealth. VR games are also likely to benefit from realistic audio, but it's still a small market.

 

[hotaru251]

give me the stats on their 60 tier base model.

dlss/MFG use more vram than normal...so can the base model of their "entry" gpu even play games well w/ it given their gimped vram/bandwidth in AAA 2025 games?

 

[TheyStoppedit]

It's easy to see what's happening in the R&D department over at NVidia. See, the problem here is that.... AI GPUs require tons and tons of VRAM. VRAM is finite, so they are trying to put as little VRAM as possible on GeForce cards so there is more for AI cards, because that's where all the money is. So they are trying to find a way to make it so games can use less VRAM. So..... fake frames it is!!!.... You're gonna see more and more different variants and addons to DLSS/MFG. It will get to the point where GPUs will just stop increasing VRAM and do very little rendering, and it will all be AI/MFG/DLSS based and optimized to be VRAM efficient. What this means is that gen over gen uplifts in VRAM will be very little.

By the time you get to the RTX 9000 series, VRAM won't be a whole lot better than the 50 series. VRAM Capacities on RTX 9000 cards in 6 years from now:

9090 = 48GB
9080 = 32GB
9070Ti = 24GB
9070 = 20GB
9060Ti = 16GB
9060 = 16GB
9050 = 12GB

You read it here first. That's 4 gens ahead of what we have now. It looks awful for a 4-gen jump, but that's the way it will be. Why? Because VRAM will become scarce as it all gets absorbed by the AI industry, and DLSS/MFG will be designed to be optimized to use very little to no VRAM for games. This means less VRAM on GeForce cards for more VRAM on AI cards. All your frames will be fake frames by this point. Why? Because AI is taking over and that's where the VRAM is needed. Let's jump ahead even further........

15090 = 64GB
15080 = 48GB
15070Ti = 40GB
15070 = 32GB
15060Ti = 24GB
15060 = 24GB
15050 = 16GB

That's 20 years from now. Some of us reading this will probably be dead by then. Look at those numbers, those are close to 2x what we have now. We will have MFGx120 & DLSS 14.0. Before you call me crazy with these predictions, just remember, the 2060 comes in a 12GB variant and the 5060Ti we have now comes in a 8GB variant. That means in some VRAM-heavy games, the 3 gen old GPU thats 1 tier lower will perform better.

And then when customers cry about the 15090 that they paid $10K for with so little VRAM, NVidia will just come out and say "64GB is enough for todays games, we dont need anymore than that" ..... that's because we aren't actually rendering/rasterizing the games anymore. Its all fake frames and AI.

It's all because of AI. AI is where the money is and that's where the VRAM is needed. Those values I gave for the 15XXX series, with true, full native rendering should be 4x those values in 20 years. But they won't be. Real rendering/rasterization will be history and it will all be MFG/DLSS/AI based

 

[bolweval]

It's easy to see what's happening in the R&D department over at NVidia. See, the problem here is that.... AI GPUs require tons and tons of VRAM. VRAM is finite, so they are trying to put as little VRAM as possible on GeForce cards so there is more for AI cards, because that's where all the money is. So they are trying to find a way to make it so games can use less VRAM. So..... fake frames it is!!!.... You're gonna see more and more different variants and addons to DLSS/MFG. It will get to the point where GPUs will just stop increasing VRAM and do very little rendering, and it will all be AI/MFG/DLSS based and optimized to be VRAM efficient. What this means is that gen over gen uplifts in VRAM will be very little.

By the time you get to the RTX 9000 series, VRAM won't be a whole lot better than the 50 series. VRAM Capacities on RTX 9000 cards in 6 years from now:

9090 = 48GB
9080 = 32GB
9070Ti = 24GB
9070 = 20GB
9060Ti = 16GB
9060 = 16GB
9050 = 12GB

You read it here first. That's 4 gens ahead of what we have now. It looks awful for a 4-gen jump, but that's the way it will be. Why? Because VRAM will become scarce as it all gets absorbed by the AI industry, and DLSS/MFG will be designed to be optimized to use very little to no VRAM for games. This means less VRAM on GeForce cards for more VRAM on AI cards. All your frames will be fake frames by this point. Why? Because AI is taking over and that's where the VRAM is needed. Let's jump ahead even further........

15090 = 64GB
15080 = 48GB
15070Ti = 40GB
15070 = 32GB
15060Ti = 24GB
15060 = 24GB
15050 = 16GB

That's 20 years from now. Some of us reading this will probably be dead by then. Look at those numbers, those are close to 2x what we have now. We will have MFGx120 & DLSS 14.0. Before you call me crazy with these predictions, just remember, the 2060 comes in a 12GB variant and the 5060Ti we have now comes in a 8GB variant. That means in some VRAM-heavy games, the 3 gen old GPU thats 1 tier lower will perform better.

And then when customers cry about the 15090 that they paid $10K for with so little VRAM, NVidia will just come out and say "64GB is enough for todays games, we dont need anymore than that" ..... that's because we aren't actually rendering/rasterizing the games anymore. Its all fake frames and AI.

It's all because of AI. AI is where the money is and that's where the VRAM is needed. Those values I gave for the 15XXX series, with true, full native rendering should be 4x those values in 20 years. But they won't be. Real rendering/rasterization will be history and it will all be MFG/DLSS/AI based

Conjecture!

 

[BTM18]

Still a hard pass on fake frames and snake oil from these clowns.
Stuff your fake frames up your green a**.

 

[NightKnight1337]

Slowly chiping away at the ungodly VRAM requirements for DLSS / MFG.

"running DLSS at 4K increases the transformer model's memory footprint by 3.5X, consuming 307.37MB of VRAM."

300mb is ungodly? wtf

 

[palladin9479]

"running DLSS at 4K increases the transformer model's memory footprint by 3.5X, consuming 307.37MB of VRAM."

300mb is ungodly? wtf

That's in addition to anything else that is running and only for the upscaler. MFG consumes about 1.5GB of VRAM currently.

 

[Amdlova]

Vram for consumer it's just a dream... why the nvidia will looser profit over quadro cards.
If you want vram pay for it.

 

[palladin9479]

Vram for consumer it's just a dream... why the nvidia will looser profit over quadro cards.
If you want vram pay for it.

There is a misunderstanding amongst consumers on how VRAM works, I blame the YT tech and other sites like Toms for just complaining without explaining.

GPU VRAM is not like system RAM where you can stick an arbitrary number. Each GDDR chip requires a dedicated 32-bit bus connection to run at full speed. 256-bit bus means eight chips, 192-bit bit means six, 128-bit bus means four and so forth. GDDR6 maximum is 2GB (16Gb) per chip, GDDR7 started at 2GB per chip and it looks like Samsung is making 3GB (24Gb) available now, though it might be rather expensive. Nobody has made 4GB (32Gb) GDDR7 chips yet. This means that unlike system RAM, which is run serially, GPU VRAM runs parallel with each chip able to be accessed independently from the rest. More VRAM means more chips means more memory bandwidth means more performance means more expensive GPU die.


This makes memory bus the big product delineator and nVidia has used malicious segmentation to screw with people. The 4060/5060 GPU's are entry tier 128-bit GPUs, the kinds normally found at the bottom of the product stack, similar to the RTX 3050 that used to be priced at $250. Instead of releasing them as 4050/5050, nVidia just changed the product name to 4060/5060 and did similar to the entire product stack. It was a ploy to justify raising product MSRP.

 

[utroz]

There is a misunderstanding amongst consumers on how VRAM works, I blame the YT tech and other sites like Toms for just complaining without explaining.

GPU VRAM is not like system RAM where you can stick an arbitrary number. Each GDDR chip requires a dedicated 32-bit bus connection to run at full speed. 256-bit bus means eight chips, 192-bit bit means six, 128-bit bus means four and so forth. GDDR6 maximum is 2GB (16Gb) per chip, GDDR7 started at 2GB per chip and it looks like Samsung is making 3GB (24Gb) available now, though it might be rather expensive. Nobody has made 4GB (32Gb) GDDR7 chips yet. This means that unlike system RAM, which is run serially, GPU VRAM runs parallel with each chip able to be accessed independently from the rest. More VRAM means more chips means more memory bandwidth means more performance means more expensive GPU die.


This makes memory bus the big product delineator and nVidia has used malicious segmentation to screw with people. The 4060/5060 GPU's are entry tier 128-bit GPUs, the kinds normally found at the bottom of the product stack, similar to the RTX 3050 that used to be priced at $250. Instead of releasing them as 4050/5050, nVidia just changed the product name to 4060/5060 and did similar to the entire product stack. It was a ploy to justify raising product MSRP.

Just to add some interesting info:
Each 32bit memory channel can be run in clamshell mode allowing it to connect to 2 gddr packages allowing for double the memory in each 32bit channel it is used in allowing for incremental memory increases. However in practice most graphics cards that use clamshell mode use it for the entire memory bus resulting in double the memory compared to a normal memory configuration.

 

[palladin9479]

Each 32bit memory channel can be run in clamshell mode allowing it to connect to 2 gddr packages allowing for double the memory in each 32bit channel it is used in allowing for incremental memory increases. However in practice most graphics cards that use clamshell mode use it for the entire memory bus resulting in double the memory compared to a normal memory configuration.

So clamshell mode itself has some nuance. GDDR memory requests are done in 16-bit chunks and each GDDR chip can handle two chunks per cycle. All clamshell mode does is physically disable one of those channels and wire those connections into a second chip. Instead of one 32-bit chip running two 16-bit commands per cycle, we get two 16-bit chips running one 16-bit command per cycle each, reducing their effective bandwidth by 50%. You are paying 2x the memory cost but only getting half the performance.

The import thing to remember is that GPU memory is parallel while system memory is serial. That is why we can have 64GB of DDR5-6400 at 104 GB/s, while 32GB of GDDR7 would be 1,792 GB/s. The DDR5-6400 is running four 32-bit channels (two per DIMM) with only one DRAM chip active at a time per channel, while the GDDR7 is running sixteen 32-bit channels at two commands per chip. You can then use clamshell to put 64GB of GDDR7 memory but are still limited to sixteen 32-bit channels and now have to run one command per chip. Half the pins on the memory chips for those 16GB 5060's are not even connected.

Clamshell mode is something that was made for datacenter / professional GPU's that needed very large amounts of memory and were willing to sacrifice memory performance for it. There should of never been any 16GB 5060's, but nVidia is going to be nVidia and screw over consumers with limited memory bus's. The 3060 came with a 192-bit memory bus allowing for six chips for a total of 12GB of VRAM, it was the 3050 that had a 128-bit memory bus for 8GB of VRAM. The 4060/5060 should of also had 192-bit memory bus's but instead nVidia crippled them with the same 128-bit bus as the 3050 to prevent them from competing with higher tier cards. This continues all the way up the stack with the 4070/5070 getting the same 192-bit bus as the 3060, 4080/5080 having the same 256-bit as the 3070. Only the halo products like the 4090 / 5090 were given full capacity memory bus's. The 4090 had 384 bit while the 5090 has 512-bit, in comparison the 3080 had 320-bit and the 3080 Ti / 3090 had 384-bit.