NVIDIA GeForce RTX 50-series Blackwell Availability

Looking at some benchmarks from the latest releases it is interessing that more advanced engines cant scale anymore on pure compute and bandwith powerhouses.
UE5 and AC: Shadows are a great example of how advancing forward do not come with more efficiency. Older games or games with lower quality run much better on Blackwell like Atomfall:

The 5090 is more than twice as fast as the 9070XT. The same is true for UE5 game South of Midnight which doesnt use Lumen and Nanite: https://www.dsogaming.com/pc-perfor...-midnight-benchmarks-pc-performance-analysis/

But when you look at The Talos Principle Reawakened with Lumen and RT-Lumen the performance on Blackwell is abnormal: https://www.pcgameshardware.de/The-...ened-Test-Release-Preis-Benchmarks-1470097/2/

Power consumption is under 350W on my 5090 OC. For example i get over 500W in Guardians of the Galaxy with Raytracing in 1440p. There is really something broken in UE5 and other engines.
slight correction, South of Midnight is Unreal Engine 4.27
 
Smartphones and computers are among many tech devices and components that will be exempted from reciprocal tariffs imposed by President Donald Trump, according to new guidance from U.S. Customs and Border Protection.
The guidance also includes exclusions for other electronic devices and components, including semiconductors, solar cells, flat panel TV displays, flash drives, and memory cards.

So no reasons to buy GPUs now because they'll become more expensive due to tariffs later then?
 
But when you look at The Talos Principle Reawakened with Lumen and RT-Lumen the performance on Blackwell is abnormal
So yeah this one is weird:
Everything seem to scale about as you'd expect - with two big exceptions being RDNA4 and Blackwell with the former outperforming its usual self while the latter underperforming for whatever reason.
5070Ti is generally within 2% from 4080S and here it's up to 10% behind (the gap is bigger in 4K which makes no sense to me).
9070 non-XT on the the other hand is beating the 7970XT here by some 5% while usually it's some 2% slower.
RDNA3 vs Ada seem to be as you'd expect though, RDNA2 vs Ampere also seems normal.
No idea what's going on with Blackwell in this title. It certainly isn't a problem of "scaling" the performance as such scaling happens just fine on the previous GeForce generation.
 
Not that interesting really. More complex graphics is more shading limited while less complex graphics is more memory bandwidth limited. Which means that more complex graphics tend to be limited by math flops while less complex graphics will be limited by GB/s.

AC Shadows is a weird one though, no idea what this one is limited by but it does underperform on Nvidia h/w in general. Same is probably true with TPP remake.
I dont see that modern engines are math limited. Going from Ampere to Blackwell there is more from everything but modern engines do not scale.
Compute performance is the best way to spend transistors. But game engines are so inefficient that they cant use the compute performance at all. And there are engines like UE5 who do lightning in a way which introduces overhead and underutilization. Pathtracing games scale better than UE5 games. And i remember a time when people said that raytracing is a GPU unfriendly workload.

For nVidia it is just better to use wafers for other markets.
 

So no reasons to buy GPUs now because they'll become more expensive due to tariffs later then?

The situation is very fluid but the overarching direction being communicated hasn't actually changed.

"All those products are going to come under semiconductors, and they're going to have a special focus type of tariff to make sure that those products get reshored. We need to have semiconductors, we need to have chips, and we need to have flat panels -- we need to have these things made in America. We can't be reliant on Southeast Asia for all of the things that operate for us," Lutnick told "This Week" co-anchor Jonathan Karl.

He continued, "So what [President Donald Trump's] doing is he's saying they're exempt from the reciprocal tariffs, but they're included in the semiconductor tariffs, which are coming in probably a month or two. So these are coming soon."


I'm just going to say it's the same overall problem for the GPU since crypto/covid. There's a lot of extenuating factors influencing supply and demand that can change quickly. As such you're taking a risk effectively either way, buy later and risk higher prices, buy now and risk prices dropping. This the problem basically, GPU prices haven't been stable and predictable since 2020.
 
In his 5060 Ti preview video GN Steve mentions that bulk purchases of 5090s from manufacturers are in the $2500-$3000 range. So basically no chance of seeing 5090s at retail anywhere close to MSRP after distributor and retailer markups.
 
Europe seem to have ample stock of 5060Tis available even on a second day after launch (I'd ignore the clickbait in the title about 8GB being the reason for that as 16GB cards are also clearly available):

US though seems like a wasteland again:

In fact 5080, 5070Ti, and 5070 especially all have better availability in the US at this moment.
5090 is MIA but I think that this is to be expected at this point - the card will have very low volumes.

And for comparison: 9070 is MIA as well while 9070XT has some availability.
B580 also appeared in some stores apparently - was essentially MIA for the last couple of months.
 
Last edited:
Are the 5060 ti selling near MSRP? Haven't seen any actual store numbers with cards in stock or maybe just missed them but if anyone knows it'd be appreciated.

I figure the 8GB should get cheap soon, also no one should buy one.
 
They seem to be readily available on local retailers here, limited to one per customer but no need to buy in bundles. The MSRP for 5060 Ti 8GB is NT$13,790 (~US$420) and there are cards from various IHVs selling at such prices. Of course there are also some "OC" versions which are slightly more expensive. The 16GB version is NT$3,000 (~US$90) more expensive.
 
Are the 5060 ti selling near MSRP? Haven't seen any actual store numbers with cards in stock or maybe just missed them but if anyone knows it'd be appreciated.

I figure the 8GB should get cheap soon, also no one should buy one.
Yea I can't see a reason to recommend that card. Just get the regular 5060 and save some money. 8GB is gonna be tough for that level of performance in the years to come.

The 16GB 5060Ti could make some sense. It's not that far from the 4070 and has plenty of memory. It wouldn't interest me but I could see someone still rocking a midrange Pascal or Turing card giving it a shot. It's even a somewhat viable upgrade from the 3060 since you at least won't be downgrading VRAM.
 
The RTX 5070 is currently selling for just €589, significantly lower than the €649 price at launch. Yes, the price was quickly adjusted to €620–€630 due to exchange rate changes, but overall, it has dropped by about 10% since its release on March 7.

The RTX 5070 isn’t the only card with a lower price. The RTX 5070 Ti is now available for €844.90, down from its €879 launch MSRP. Unfortunately, we can’t say the same for the RTX 5080 and 5090—the latter has reached an unreasonably high price point.
 
So yeah this one is weird:
Everything seem to scale about as you'd expect - with two big exceptions being RDNA4 and Blackwell with the former outperforming its usual self while the latter underperforming for whatever reason.
5070Ti is generally within 2% from 4080S and here it's up to 10% behind (the gap is bigger in 4K which makes no sense to me).
9070 non-XT on the the other hand is beating the 7970XT here by some 5% while usually it's some 2% slower.
RDNA3 vs Ada seem to be as you'd expect though, RDNA2 vs Ampere also seems normal.
No idea what's going on with Blackwell in this title. It certainly isn't a problem of "scaling" the performance as such scaling happens just fine on the previous GeForce generation.
It isnt weird. It just abnormal inefficient software. Oblivion remaster and Clair Obscur: Expedition 33 are showing the same performance. Nanite and Lumen are just so bad. And UE5 can only exist because there isnt any competition. Optimizing you engine for modern GPUs? Na, blame nVidia for their hardware.

UE5 is just rendering geometry with an outdated lighting system from the last decade. Reflections? Nope. Direct lighting? Nope. Shadows? Maybe VSM or so. Alan Wake 2 and Indiana Jones running at the same frames in 1080p with Pathtracing. Doom and F1 will be the same. You can not make a performing UE5 game with modern lighting.

There is just no reason for nVidia to use their Wafers for GB203 and especially for GB202.
 
It isnt weird. It just abnormal inefficient software. Oblivion remaster and Clair Obscur: Expedition 33 are showing the same performance. Nanite and Lumen are just so bad.
This isn't an issue with either as we have a huge number of UE5 titles using both which do not show similar performance issues.
 
"Huge numbers"? Most UE5 games with Nanite and Lumen showing the exact same performance pattern. Maybe with nVidia's branch there are some optimizations but standard UE5 has not got any kind of performance improvements since the launch two years ago.
 
"Huge numbers"? Most UE5 games with Nanite and Lumen showing the exact same performance pattern. Maybe with nVidia's branch there are some optimizations but standard UE5 has not got any kind of performance improvements since the launch two years ago.

This would be an incredible claim to substantiate.
 
Immortals of Aveum with Lumen and Nanite run with 100 FPS on a 4090 in 1080p: https://www.dsogaming.com/pc-performance-analyses/immortals-of-aveum-pc-performance-analysis/
Clair Obscur: Expedition 33 delivers the same performance on a 4090: https://gamegpu.com/action-/-fps-/-tps/clair-obscur-expedition-33-test-gpu-cpu

Dont see huge differences in image quality between both games.

Now compare this to the progression with Raytracing in more tradional engines. I can play Guardians of the Galaxy with 160 FPS in 1440p and RT-Reflections. In fact i get 100 FPS in 4K with RT-Reflections with a 5090. Indiana Jones is brute force rasterizing and i get over 100 FPS in 1080p in the jungle section with Pathtracing.

Just a random fun fact: UE5 runs so bad that people cant use one of the first 120Hz 1080p display at full speed.
 
Maybe with nVidia's branch there are some optimizations but standard UE5 has not got any kind of performance improvements since the launch two years ago.

Unreal Engine 5.4:

"Render Thread performance is very often the limiting factor for UE titles. This is because some operations are restricted to this particular thread, even though current platforms and graphics API can do them in parallel. The goal is to improve performance by refactoring the Renderer Hardware Interface (RHI) API to remove these constraints and fully utilize the multithreading capabilities of the target hardware.

We shipped a major refactor of the rendering systems in 5.4 which allows for far greater overlap of work, which in turn permits lower target frame times. In our CitySample testing, the render thread performance is roughly 2x faster in 5.4 as compared to 5.0."

"We made substantial Improvements to hardware raytracing (HWRT). These improvements offer speed gains of 2x in the case of primitives and it helps to ship 60hz experiences which use HWRT."

Unreal Engine 5.5:


"Release 5.5 has made many improvements to the underpinning systems which all fall into the external-facing hardware raytracing category (HWRT). These lower level systems all impact the performance and capabilities of Lumen, Path tracing, and Light Baking. There are improvements to asynchronous evaluation of raytracing code, improvements to caching and better management of acceleration structures. The intent of these changes is to make it feasible to utilize HWRT at higher refresh rates on platforms for which there is hardware support."


Alex found the City Sample demo ran 42% better in CPU limited scenarios under UE 5.4.


I don't believe any games currently use UE 5.4+ apart from Fortnite.
 
Last edited:
Back
Top