NVIDIA GeForce RTX 5090 Founders Edition

[10tothemin9volts]

Even with same random seed?

Not me asked, but I think seed (or temperature for that matter) doesn't matter. It just changes/randomizes the output and was matters is the (VRAM) memory bandwidth. For my previous post text-to-text suggestion, if you do the benchmark, please in tokens per second, as indeed the answer can change, but to have the same answer every time reduce the temperature to the lowest value possible. Rough tokens per second speed can be calculated: Bandwidth / LLM size = 1792GB/s / 27GB (for the Qwen2.5-32B-Instruct-Q6_K.gguf suggestion) = 66 token/s. Of course, context size (=input text amount) matters, but as long as it's less than a couple of sentences, the VRAM consumption won't be higher than 32GB.

PS: Power scaling benchmark would be nice too.

 

[N3M3515]

35% more performance with 25% higher power draw qualifies as "Highly Energy Efficient"?

Whatever you say, I guess...

And 25% higher price

 

[Legacy-ZA]

NVIDIA GeForce RTX 5090 Founders Edition Review - The New Flagship

NVIDIA's GeForce RTX 5090 is the fastest graphics card ever released. It comes with a whopping 32 GB VRAM, and support for multi-frame generation, which achieves hundreds of FPS easily. We also managed to disassemble the card, and our review includes performance testing without upscaling or...

www.techpowerup.com

Ah, missed the latencies, thank you, I still want to see some image comparisons and a short video with MFG enabled.

 

[swirl09]

Getting old sucks. Never thought a gen could be so bad I would so easily decide to skip it.

Its been frequently parroted with each gen (especially since the bad price creeping with Turing and onward) that flagship buyers have no sense and dont considering anything - they just buy it anyway. I never really thought this was true, and always gave consideration to what was on offer. I like tech, I like upgrading, the money involved is not nearly as egregious as some make out (seriously, try doing a hobby - they all cost!)

But whats being offered here, is not good enough. Another price jump, frankly too much power, for gains that more or less only go up along side those. This is just a bad deal.

2027 then....

 

[N3M3515]

Indeed, it is as I suspected too, the cooler has amazing performance. Oof, and the FPS is much desired. All I know is, I want one, but I can't buy one, envy, indeed.

The RTX5080 is going to look like a turd compared to this card.

The 5080 won't even beat the 4090

 

[Shtb]

Again it's 35% more rasterization over the 4090. You can't subtract 74 from 100. You have to divide.

When A is bigger than B by such and such an amount, you subtract.
When A is so much a part of B, you divide.

А = 100.
В = 74.

A is bigger, than B, by 26.
But A is 135% of B.

p.s. Sorry for my English, it's not my first language, i hope i made it clear.

 

[N3M3515]

Spoiler: hidden because irrelevant

Was the 4090 Really 2400 $ or is it a special edition?
View attachment 381150

Edit: apparently it is! (current price) Thanks for letting me know.

And the 5090 will be at the very least $2500

 

[Krit]

RTX 5080 will be very interesting review for nvidia users indeed.

RTX 5090 claimed boost was 33% and it's actually is! TPU 35% faster on average, HWU 27% faster on average.
RTX 5080 claimed boost was 15%
RTX 5070 Ti claimed boost was 20%
RTX 5070 claimed boost was 20%

 

[TSiAhmat]

And the 5090 will be at the very least $2500

yes and probably >3000€ in my region. But I never planned on buying one. So it won't affect me.

Good Luck to everyone eyeing the 50 Series cards, this time it might be rough (especially the 90 card)

 

[Solaris17]

@W1zzard

NVIDIA GeForce RTX 5090 Founders Edition Review - The New Flagship

NVIDIA's GeForce RTX 5090 is the fastest graphics card ever released. It comes with a whopping 32 GB VRAM, and support for multi-frame generation, which achieves hundreds of FPS easily. We also managed to disassemble the card, and our review includes performance testing without upscaling or...

www.techpowerup.com

went up by [FIXME]°C,

 

[W1zzard]

@W1zzard

NVIDIA GeForce RTX 5090 Founders Edition Review - The New Flagship

NVIDIA's GeForce RTX 5090 is the fastest graphics card ever released. It comes with a whopping 32 GB VRAM, and support for multi-frame generation, which achieves hundreds of FPS easily. We also managed to disassemble the card, and our review includes performance testing without upscaling or...

www.techpowerup.com

I'm putting my card back together right now, writing directions in reverse Will update the temp once the card is back together, and it works .. wish me luck

 

[wNotyarD]

When A is bigger than B by such and such an amount, you subtract.
When A is so much a part of B, you divide.

А = 100.
В = 74.

A is bigger, than B, by 26.
But A is 135% of B.

p.s. Sorry for my English, it's not my first language, i hope i made it clear.

Wrong math, you must calculate by rule of three.

5090 performance	4090 performance
100	74
x	100

To get the 5090 performance using the 4090 as a baseline, x = 100²/74 ≈ 135.
So 135%, or a 35% uplift.

 

[csendesmark]

Agreed. I don't see a text-to-text LLM benchmark so I suggest a SOTA 32B model perfect for the 32GB VRAM: Qwen2.5-32B-Instruct-Q6_K.gguf (e.g. from here) (the Q8 (34.82GB) wouldn't fit anymore).

Personally I would test with ~16+ GB models since all relevant cards 16GB or more VRAM
Llama 3.1 was really great when it came out, but Llama 3.3 is 2 months old already.

Download and run lmstudio-community/Llama-3.3-70B-Instruct-GGUF in LM Studio

Use lmstudio-community/Llama-3.3-70B-Instruct-GGUF locally in your LM Studio

model.lmstudio.ai

With maximum layers what this GPU can carry.

Would suggest

DevQuasar/deepseek-ai.DeepSeek-R1-Distill-Llama-8B-GGUF · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

DevQuasar/deepseek-ai.DeepSeek-R1-Distill-Llama-8B-GGUF F16

And to see this

deepseek-ai/DeepSeek-R1-Distill-Qwen-32B · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

 

[Legacy-ZA]

The 5080 won't even beat the 4090

That is what I am afraid of, sigh, might go for an RTX5070Ti instead, looks like it will be the model that won't give you the middle finger to the uttermost.

 

[Tomgang]

4090 will solidly keep a nice position. 5080 will NOT improve much over the 4080(S) without MFG, so it won't get anywhere near the 4090. It won't be able to keep its inflated prices for new cards, though, if not for the scalped/scarce 5090's.

Yeah rtx 5080 dosent look impressive on paper compared to rtx 4080 super.

If i can still call rtx 4090 second or 3 fastest card in the world after over 2 years. That's pretty good i think. Un less amd present something amazing and faster than 4090. I'll say if i can keep the second or third fastest card in the world for native performance. I will be most satisfied.

Also according to Nvidia multi framegeneration could come to the 4000 series of card in the future and if that happens. It deffently not worth going rtx 5090 over 4090

 

[Krit]

Don't worry folks RTX 5090 is not a gaming card and it will cost around 2300-2500€ for gaming it makes no sense at all.

RTX 5080, RTX 5070 Ti, RTX 5070 will be for gaming......

 

[Vya Domus]

@W1zzard Out of curiosity what are the hotspot temperatures under load or is that sensor not exposed to software yet ?

 

[W1zzard]

@W1zzard What are the hotspot temperatures under load or is that sensor not exposed to software yet ?

Hotspot has been removed, I just updated the OC page to mention that fact

 

[WonkoTheSaneUK]

NVIDIA: 566.14 WHQL
Why isn't Nvidia releasing new drivers to reviewers?

They probably won't even be ready until the cards go on sale.

 

[Vya Domus]

Hotspot has been removed, I just updated the OC page to mention that fact

Yeah just saw the overclocking page, curious choice.

 

[Francoporto]

Now i want the same perf for 75W

 

[W1zzard]

NVIDIA: 566.14 WHQL
Why isn't Nvidia releasing new drivers to reviewers?

RTX 5090: 571.86 Press Driver, look one row up

566.14 was used in December when I spent like two weeks retesting all the cards in the test group. After that xmas, CES and this launch--no way to retest anything. Will retest in a few weeks of course, with new games etc

 

[Legacy-ZA]

Don't worry folks RTX 5090 is not a gaming card and it will cost around 2300-2500€ for gaming it makes no sense at all.

RTX 5080, RTX 5070 Ti, RTX 5070 will be for gaming......

I hate what that snake leather jacket man turned us into, we have to choose between models that won't screw us the most.

 

[SOAREVERSOR]

Im buying but only because I have no GPU at the moment. Hard not to feel like I have to subject myself to robbery to buy something nowadays.

Unless you also plan on doing creative or AI/ML/DL work this is not the GPU for you. If you do plan on doing that it's far from robbery. That is the catch with these. When you view them as a work product everything makes sense. When you view it as a gaming product it's using solid gold diamond encrusted nuclear bombs because of a mouse in your apartment.

 

[Enterprise24]

Any chance that custom cards will get more than 600w power limit?