Nvidia Geforce RTX 5090 reviews

"To sum it up, while the Transformer model for Super Resolution runs at very minor performance cost of a 5% on an RTX 3060, running Transformer Super Resolution and Ray Reconstruction at the same time will significantly drop your performance by 25% compared to DLSS SR + RR on the CNN model, which is quite massive. Some could argue that an RTX 3060 is struggling with ray tracing anyway, thus Ray Reconstruction isn't really a thing on this GPU, but we also tried to run the same test on an RTX 3080, a more powerful Ampere GPU and very capable of running ray tracing—and the results were the same, indicating that the quality of Ampere's tensor cores is the issue, not the quantity."

 
"To sum it up, while the Transformer model for Super Resolution runs at very minor performance cost of a 5% on an RTX 3060, running Transformer Super Resolution and Ray Reconstruction at the same time will significantly drop your performance by 25% compared to DLSS SR + RR on the CNN model, which is quite massive. Some could argue that an RTX 3060 is struggling with ray tracing anyway, thus Ray Reconstruction isn't really a thing on this GPU, but we also tried to run the same test on an RTX 3080, a more powerful Ampere GPU and very capable of running ray tracing—and the results were the same, indicating that the quality of Ampere's tensor cores is the issue, not the quantity."

i'll be damned.
 
"To sum it up, while the Transformer model for Super Resolution runs at very minor performance cost of a 5% on an RTX 3060, running Transformer Super Resolution and Ray Reconstruction at the same time will significantly drop your performance by 25% compared to DLSS SR + RR on the CNN model, which is quite massive. Some could argue that an RTX 3060 is struggling with ray tracing anyway, thus Ray Reconstruction isn't really a thing on this GPU, but we also tried to run the same test on an RTX 3080, a more powerful Ampere GPU and very capable of running ray tracing—and the results were the same, indicating that the quality of Ampere's tensor cores is the issue, not the quantity."

How does Ada fare? They mention it being better on the 4080 but some results would be sweet.

Does the new TNN model use FP4 or something to make it run better on Blackwell?
 
It would be interesting to see the overhead in terms of latency and VRAM per generation. Going from some data I'm wondering if the newer models leverage FP8/INT8 (or even FP4) which makes them relatively lighter on Ada and Blackwell over Ampere and Turing.
 
That's a really good idea. Wonder if I can get Cyberpunk running on my Fedora rig? My EVGA 3080ti is over there folding right now...
 
From what benchmarks we have so far it seems that Ada is about as good as Blackwell while Turing and Ampere are getting a higher performance hit.
But this is true for RR only it seems. SR is about the same on all GPUs.

I think turing and ampere are way behind in terms of "tensor ops". Could be that it's enough to handle SR on its own, but if you enable RR or both, you hit a big bottleneck.
 
From what benchmarks we have so far it seems that Ada is about as good as Blackwell while Turing and Ampere are getting a higher performance hit.
But this is true for RR only it seems. SR is about the same on all GPUs.

I seem to recall a clip posted by I believe @DavidGraham (?) that showed SR (or maybe SR + FG) with the new model using more VRAM on a Ampere GPU? While it seems to be using less VRAM on Ada and Blackwell from the numbers from say TPU. Not sure if there is more corrobating this.
 
The more I think about it, the less likely I am to buy a 5090 anytime soon. I've never cared enough to "camp" in preparation to buy a thing, no matter what the thing might have been. I'm also not willing to camp virtually on top of the Best Buy app in order to buy a card in the first 1.2s before every scalper on the planet buys all of them to resell at $3999 USD.
 
The more I think about it, the less likely I am to buy a 5090 anytime soon. I've never cared enough to "camp" in preparation to buy a thing, no matter what the thing might have been. I'm also not willing to camp virtually on top of the Best Buy app in order to buy a card in the first 1.2s before every scalper on the planet buys all of them to resell at $3999 USD.
I'm wondering if you'll ever be able to go on Newegg and get a 5090 for MSRP.
 
Same way you were never able to get a 4090 for MSRP either, insofar as I'm aware.
There was maybe a brief period where you could find at least one model 4090 in stock at Newegg for around MSRP. Going by memory since the price history trackers I looked up are highly questionable. I don't recall the 4090 ever going for $400.

But yea the 4090 was effectively an $1800+ card, and spent much of its life well above that. It's why I was shocked by the $2000 5090 MSRP. NVIDIA is leaving money on the table, but maybe the goodwill is worth more to them. The missed profits are a drop in the bucket to them.

P.S. I'm cracking up thinking about a $2000 GPU generating goodwill :mrgreen:
 
From what I remember the RTX 4090 was relatively accessible during the first half of 2023.

It was really in the second half things became rather problematic as multiple market shocks ended up happening. Speculatively it was due to the rise in AI popularity combined with the trade situation in China (eg. they were said to diverted shipments into China before the first embargos) that initially sparked the issue.
 
From what I remember the RTX 4090 was relatively accessible during the first half of 2023.

It was really in the second half things became rather problematic as multiple market shocks ended up happening. Speculatively it was due to the rise in AI popularity combined with the trade situation in China (eg. they were said to diverted shipments into China before the first embargos) that initially sparked the issue.
That sounds right. IIRC the 4090 went poof when the China restrictions were announced.
 
From what I remember the RTX 4090 was relatively accessible during the first half of 2023.

It was really in the second half things became rather problematic as multiple market shocks ended up happening. Speculatively it was due to the rise in AI popularity combined with the trade situation in China (eg. they were said to diverted shipments into China before the first embargos) that initially sparked the issue.
Hmm, that might be true, I honestly can't remember.

I actually bought my ASUS Tuf Gaming 4090 directly from the ASUS store almost exactly a year ago (Jan 21st, 2024) and paid $1819.99 for it with free shipping, alas I had to pay my local 9.75% sales tax :( Anyway, point being, I was able to find a 'reasonable' priced one, albeit the qualified statement being I had to buy an aftermarket one for a higher than FE-MSRP price. I'm OK with that, and I suspect this is how I'll end up with my own 5090 in the eventual future.
 

My memory is coming back and @arandomguy is definitely correct. The 4090 was available for just about MSRP until the ban which caused it to immediately shoot up and then up some more until it became impossible to purchase outside of bundles.
 
One of the larger PC retailer in Taiwan is hosting a "launch party" for 5090 and 5080 on 1/30. Availability of 5090 is very limited, with only two varieties available:

- ASUS ROG-ASTRAL-RTX5090-O32G-GAMING (NT$89,990)
- MSI RTX5090 32G GAMING TRIO OC (NT$83,990)

5080 are more available, with prices range from NT$35,990 to NT$45,990.

(US$1 ~ NT33)

need to purchase with selected motherboard. Prices are tax included.
 
Back
Top