Tuesday, October 13th 2026

NVIDIA DLSS 5 Gets 2x Performance Boost with 20% Less VRAM Usage

NVIDIA's continuous optimization of software and its technology stack has yielded another great improvement. The latest DLSS 5 update, scheduled for October 16, will deliver another 2x performance boost, meaning that previously DLSS 5-optimized workloads will literally run twice as fast. Alongside this significant performance uplift, NVIDIA promises that in 4K gameplay with DLSS 5 enabled, the neural rendering technology will use 20% less VRAM, leaving more room for game textures to load without DLSS 5 interfering with capacity or hogging resources. NVIDIA updates its DLSS by running a continuous loop on top of an internal supercomputer that constantly optimizes DLSS technology for performance, efficiency, and visual fidelity. Specifically for this week's DLSS 5 update, NVIDIA states that the company used a combination of systematic pipeline optimizations and neural model refinements to achieve that 2x performance boost.

Originally in late August, we saw that the leaked DLSS 5 file was roughly a 148 million parameter model. Our analysis turned out to be accurate, and we see that, according to NVIDIA, the DLSS 5 October update shrinks the footprint of this model by another 24% to roughly 118 million parameters, meaning that our original estimation was correct. Consequently, the smaller model scales memory usage almost linearly and cuts VRAM usage by about 20%. Compared to the initial launch in mid-March, when the DLSS 5 was showcased on two GeForce RTX 5090 GPUs, we see that NVIDIA managed to scale the performance by 10x in the span of a couple of months.
You can see the DLSS 5 September versus October update below.

Source: NVIDIA
Add your own comment

25 Comments on NVIDIA DLSS 5 Gets 2x Performance Boost with 20% Less VRAM Usage

#1
lexluthermiester
This is not unimpressive. But will it work with non-5000 series cards?
Posted on Reply
#2
Quicks
Since when is going from 240fps to 312fps a 2x performance boost?

"will deliver another 2x performance boost, meaning that previously DLSS 5-optimized workloads will literally run twice as fast."

Let's wait and see benchmarks from other people.
Posted on Reply
#3
Youseethis
It seems to be better looking as well but there will be another article I guess. Nvidia showcased it in Hogwarts Legacy.
QuicksSince when is going from 240fps to 312fps a 2x performance boost?

"will deliver another 2x performance boost, meaning that previously DLSS 5-optimized workloads will literally run twice as fast."

Let's wait and see benchmarks from other people.
They are talking about its impact in ms and it's not the only thing running.
Posted on Reply
#4
FR3D1
Ninja Cat
I enjoy reading about news like this, and not just because it comes from NVIDIA.

However, the most important question remains unanswered:
lexluthermiesterBut will it work with non-5000 series card?
Who or what benefits?

Regards,
RTX. It’s On.
Posted on Reply
#5
MikOnTopGG
QuicksSince when is going from 240fps to 312fps a 2x performance boost?

"will deliver another 2x performance boost, meaning that previously DLSS 5-optimized workloads will literally run twice as fast."

Let's wait and see benchmarks from other people.
Well, 40 to 80 FPS is a performance boost right? (Of course unofficially.)

So, they are still busy I see. Good.

Hope to see some of these results on non RTX-50 Series GPU's.
Posted on Reply
#6
chrkeller
I legit worry developers are going to be lazy and just expect gamers to DLSS poorly optimized software.
Posted on Reply
#7
Dorek
6x MFG comparisons so dunno how thats supposed to translate to anything useful.
Posted on Reply
#8
wolf
Better Than Native
It will be interesting to see whether the discussion focuses on the actual improvements or the usual talking points and completely original ai sLoP/reshade filter comments.

I'm astonished at the rate of development and performance optimization, I wonder what modders will be able to eek out of this, its going absolutely bananas out there.
Posted on Reply
#9
JustBenching
QuicksSince when is going from 240fps to 312fps a 2x performance boost?

"will deliver another 2x performance boost, meaning that previously DLSS 5-optimized workloads will literally run twice as fast."

Let's wait and see benchmarks from other people.
It's the DLSS 5 part that is twice as fast.
Posted on Reply
#10
Godrilla
chrkellerI legit worry developers are going to be lazy and just expect gamers to DLSS poorly optimized software.
Luckily the modding community isn't waiting for the gaming industry to get it's act together. Aaa titles are becoming more irrelevant by the quarters. With more aaa title sudios closing are being replaced by 100x indi developers. Also the modding community is putting these incompetent developers to shame using Vibe coding to fix their shortfalls. Poetic Justice is being served.
lexluthermiesterThis is not unimpressive. But will it work with non-5000 series cards?
I don't see why not if the dll* can be replaced and or modded thus far on all rtx cards and or Radeon gpus. It's a matter of time.
Like I said before everyone is eventually going to have a decent neural network renderer. Nvidia is racing to corner the market and need to maximize it efficiency as quickly as possible.
Gamers are adding neural rendering to even PS5 video pass through, capture card and midrange rtx cards as of late.



Gta 6 with Neural rendering is already a possibility.
Posted on Reply
#12
v12dock
Block Caption of Rainey Street
I picked up a N1X laptop I am excited to try this out.
Posted on Reply
#14
evernessince
lexluthermiesterThis is not unimpressive. But will it work with non-5000 series cards?
Nvidia numbers so a mountain of salt.

It should absolutely work on prio gen cards. The only feature that could improve performance on the 5000 series is 4-bit support and 4-bit support has never resulted in a 2x increase in any AI model. It's often 10% or less. Smaller models (like this) also see a larger quality drop when going from 8 bit to 4 bit.
Posted on Reply
#15
HisDivineOrder
The original model was built to sell those Super variants that were going to up the 2gb to 3gb for every card, which would have boosted their VRAM. Whoops. Those didn't get released and only now are they doubling back because, whoops, selling 32GB 5090's isn't exactly profitable anymore for Nvidia, so here's that 5080 Super card but probably with a 5090 original MSRP.

Given that, I'm not surprised they've redone the model now with less VRAM requirement because it was pretty silly given they're only selling 16GB or (mostly) less cards.
Posted on Reply
#16
FamHectix
evernessinceNvidia numbers so a mountain of salt.

It should absolutely work on prio gen cards. The only feature that could improve performance on the 5000 series is 4-bit support and 4-bit support has never resulted in a 2x increase in any AI model. It's often 10% or less. Smaller models (like this) also see a larger quality drop when going from 8 bit to 4 bit.
Please don't spread misinformation like that. Deepest r1 gained 300% performance improvement on fp4, flux image generation running locally gained 390% performance improvement on fp4 vs fp8, etc. Most of the clear indicators of fp4 are 2x performance improvement and smaller vram footprint.
This is probably gonna be locked to 50 series, but 40 series and older can still run fp8.
Posted on Reply
#17
lexluthermiester
FamHectixThis is probably gonna be locked to 50 series, but 40 series and older can still run fp8.
I'm curious specifically about my 3080. If it'll work, would like to try it out. However, a 5080 is tempting me just a little bit..
Posted on Reply
#18
evernessince
FamHectixPlease don't spread misinformation like that. Deepest r1 gained 300% performance improvement on fp4, flux image generation running locally gained 390% performance improvement on fp4 vs fp8, etc. Most of the clear indicators of fp4 are 2x performance improvement and smaller vram footprint.
This is probably gonna be locked to 50 series, but 40 series and older can still run fp8.
Let's see here:
  • You are a brand new account created today.
  • This is your first and only post.
  • Regurgitates Nvidia's market numbers
Nice to meet you JensenBot.

You are leaving out the fact that the scenario in which those gains are had is at higher interactivity values.

Nevermind that the full DeepSeek R1 model is 1.36TB in size. DLSS5 5 is 150 MB. They aren't in the same universe, let alone the same type of models.

Any claims of Flux Dev getting a 390% speedup are complete BS. Stop regurgiating Gemi answers which cites Nvidia marketing material of all things as it's source.

VRAM savings going from 8 bit to 4 bit at this model size are basically irrelevant. I'd much rather take the higher qualty over 20-50 MB of saved space.
Posted on Reply
#19
lexluthermiester
evernessinceLet's see here:
  • You are a brand new account created today.
  • This is your first and only post.
  • Regurgitates Nvidia's market numbers
Nice to meet you JensenBot.

You are leaving out the fact that the scenario in which those gains are had is at higher interactivity values.

Nevermind that the full DeepSeek R1 model is 1.36TB in size. DLSS5 5 is 150 MB. They aren't in the same universe, let alone the same type of models.

Any claims of Flux Dev getting a 390% speedup are complete BS. Stop regurgiating Gemi answers that are based off Nvidia marketing material.

VRAM savings going from 8 bit to 4 bit at this model size are basically irrelevant. I'd much rather take the higher qualty over 20-50 MB of saved space.
Aww, that was cute. The fact that they're correct doesn't matter at all does it...
Posted on Reply
#20
Hecate91
lexluthermiesterAww, that was cute. The fact that they're correct doesn't matter at all does it...
It doesn't matter since it's all Nvidia marketing speak, it'll be more interesting to see how exaggerated these claims are in real testing.
lexluthermiesterI'm curious specifically about my 3080. If it'll work, would like to try it out. However, a 5080 is tempting me just a little bit..
30 series users may have to settle for mods with FP8 mode, Nvidia has only said 40 series will get DLSS5 at some point.
But a 5080 is tempting, even at the 2 grand price tag in the current market?
Posted on Reply
#21
lexluthermiester
Hecate91But a 5080 is tempting, even at the 2 grand price tag in the current market?
It is. I can afford it. And with an increasing number of reasons to do so, an upgrade might be in my near future, even at the redonkulas prices..
Posted on Reply
#22
chrcoluk
QuicksSince when is going from 240fps to 312fps a 2x performance boost?

"will deliver another 2x performance boost, meaning that previously DLSS 5-optimized workloads will literally run twice as fast."

Let's wait and see benchmarks from other people.
It isnt, they really mean the DLSS element is twice as fast, but obviously marketing wont say that.
Posted on Reply
#23
ISSEnterprise
There won't be any RTX5090 or something more than RTX5080 24G in few years later. Whatever DLSS tech they create, should be worked on 5080 24G for premium performance and visual. I am curious why they still refuse to let Multi-frame-gen be available on RTX30/20 series when community mods had done that for months.
Posted on Reply
#24
evernessince
lexluthermiesterAww, that was cute. The fact that they're correct doesn't matter at all does it...
Nvidia's marketing data is technically correct in the scenario they present it. Does that mean it's the level of performance customers should expect? No, absolutely not. That's called missing the forest through the trees.
Posted on Reply
#25
igormp
evernessinceThe only feature that could improve performance on the 5000 series is 4-bit support and 4-bit support has never resulted in a 2x increase in any AI model.
4-bit support can totally improve performance by 2x, be it due to the way lower memory footprint (not really relevant here), or due to the doubling of FLOPs when hw support is present. It really depends on the arithmetic intensity of your model.
evernessinceSmaller models (like this) also see a larger quality drop when going from 8 bit to 4 bit.
That really depends on the model architecture, how it was trained and specially if it's over-parameterized. I've quantized many models down to INT4/INT8 to be used in µCs with less than 2% in F1-score drop.
With that said, I don't think they changed the data types of the weights for the model at hand, but rather just adjusted the model's architecture.
Posted on Reply
Add your own comment
Nov 22nd, 2026 16:44 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts