NVIDIA RTX 50 SERIES - Technical/General Discussion

[Dicehunter]

Managed to get some specs on Rubin

Only relevant number for us is the fp32 tflops. It's 63% higher than Blackwell.

So potentially 60-70% higher tflops going from RTX5000 to RTX6000 for gamers

Rtx4090 to 5090 was only 35% increase in tflops, so potentially the 5090 to 6090 will be a much bigger upgrade this time

By the time it comes out in 2027 it'll likely cost in the realm of $5000... I wonder how many on this forum will buy it.

 

[Grim5]

Dlss 4.5 is out now via latest driver

 

[Zigwig]

Question: In the Nvidia app under global settings, do you leave DLSS Override - Models Presets as ‘Use the 3D application setting’ or do you change it to ‘Latest’

I understand what each setting does, but just curious as to everyone’s experience and what do you leave on as default. I can’t be bothered to sort game by game.

 

[parisv]

Managed to get some specs on Rubin

Only relevant number for us is the fp32 tflops. It's 63% higher than Blackwell.

So potentially 60-70% higher tflops going from RTX5000 to RTX6000 for gamers

Rtx4090 to 5090 was only 35% increase in tflops, so potentially the 5090 to 6090 will be a much bigger upgrade this time

it was always gonna be a bigger leap due to going to 2nm

 

[Livewire88]

Question: In the Nvidia app under global settings, do you leave DLSS Override - Models Presets as ‘Use the 3D application setting’ or do you change it to ‘Latest’

I understand what each setting does, but just curious as to everyone’s experience and what do you leave on as default. I can’t be bothered to sort game by game.

Copied from the NVIDIA site:

"Available Now: With today’s NVIDIA app beta update, all GeForce RTX users can experience the new DLSS 4.5 Super Resolution model in over 400 titles. Simply select “DLSS Override feature - Model Presets” in the Graphics tab, and choose “Latest”. This sets our new Model M across all modes. You can also try out Model L under “Custom”, which is optimized for Ultra Performance Mode at 4K. Ultimately, you decide which balance of image quality and performance is right for your PC."

 

[Nitefly]

Interesting vid from Linus on it:

 

[Yaayuh!]

Wrong thread - ignore

 

[Calin Banc]

I know, that's why I said AI GPus, these are not gaming but the previous schedule has been about 6 months from launch of server production to when gaming GPUs launch, so we'll probably get RTX6000 announced in 1st or 2nd quarter of 2027

Unfortunately everytime Nvidia launches new architecture these days they don't provide the juicy details that I'd like to see, so all the materials released today just has claimed performance but doesn't mention specs of the cards. Nvidia is claiming 2-5x higher AI performance compared to last gen but thisnis meaningless for gamers

2-5x higher performance, plus bringing products to market extremely quickly means that those 6 years of life for the cards is no where near reality, ESPECIALLY when your power limited / data center. CAPEX costs are crazy and this is just pushing the bubble towards popping sooner.

 

[Rawr]

Not massively excited for 6x frames as I barely used MFG at the previous multipliers with my current screens, hopefully the DLSS upgrade will be good though

 

[Grim5]

Dlss 4.5 from what I've seen is offering quite significant improvements in image quality

However there is a large performance hit using on some older cards because the older cards don't have FP8

There is also a hit to VRAM on all cards but twice as bad on RTX3000 and 2000 models. Enabling DLSS4.5 uses around 50% more memory for the DLSS component on RTX4000/5000 but uses 100% more on other cards. In nominal terms this means DLSS 4.5 uses around 200mb more VRAM on RTX3000/2000 than it does on 4000/5000


Here is a chart showing performance impacts comparing dlss 4.5 to dlss 4. For example it shows framerate on a 3080ti can drop by as much as 18% using 4.5 vs 4

 

[~>Dg<~]

so basically making older cards look redundant when using it. well played nvidia

 

[Yaayuh!]

PSU and 5090 in. No time to play with it properly today. Tomorrow I will be benching and taking a look at getting the most out of it with undervolting.

 

[~>Dg<~]

noticed a lot of places the 5060ti is now " out of stock " i bet we see big jump in prices soon for those too once they put them back on the shelves from their warehouses

 

[Rroff]

Here is a chart showing performance impacts comparing dlss 4.5 to dlss 4. For example it shows framerate on a 3080ti can drop by as much as 18% using 4.5 vs 4

Think that chart is a bit optimistic if it is maximum - I was seeing up to 11% drops on my 4080 Super depending on settings, game and where in game and so on. Though average was more like 3-5% or so.

 

[Calin Banc]

so basically making older cards look redundant when using it. well played nvidia

If you have the performance left on what you're running, you can activate that or some other Preset, like K and not use version 4.5

 

[Grim5]

If you have the performance left on what you're running, you can activate that or some other Preset, like K and not use version 4.5

I wonder how many users will even know about it though - once games start shipping with 4.5 as the native preset, users on a 3060 for example will just enable dlss and play the game, not knowing that dlss 4 would give them 30% higher framerate

 

[Calin Banc]

I wonder how many users will even know about it though - once games start shipping with 4.5 as the native preset, users on a 3060 for example will just enable dlss and play the game, not knowing that dlss 4 would give them 30% higher framerate

CB77 give you the option to chose between models, so there isn't really an issue from a dev side to implement this.

Moreover, Nvidia's recommended settings should make the necessary adjustments.

On the other hand, keep in mind FG. People didn't like that was missing for 2xx and 3xx series.

 

[Tinek]

it was always gonna be a bigger leap due to going to 2nm

When will journalists finally learn and stop peddling that rubbish - it's not 2nm, it's N2 process: about 20nm smallest metal component pitch, 45nm transistor gate pitch. And they only difference between this "2nm" and previous "3nm" is a tiny bit tighter space between components so they can stuff a bit more transistors in the same space. That's it. Any improvement from the process comes only with more transistors and more power use but the GPU.

 

[parisv]

PSU and 5090 in. No time to play with it properly today. Tomorrow I will be benching and taking a look at getting the most out of it with undervolting.

nice. here's a good guide to undervolt

 

[strumpusplunket]

Put my 5070 Ti into my PC yesterday, and I've been messing about a bit with 3DMark to make sure it's all benching as expected.

One thing that has been a surprise is that even running the tests for a while I haven't been able to hear my card's fans. I downloaded checked GPU-Z to track their RPM, and it looks like the card is maxing out at around 67C with only 40% fan speed.

I did not expect it to be so much cooler and quieter than my old 3080.

I haven't actually got a demanding game installed at the moment to try it out with, but I would have thought the Time Spy and Steel Nomad tests would stretch its legs a bit?