NVIDIA GeForce RTX 5050, RTX 5060, and RTX 5060 Ti Specifications Leak

[theglaze]

I'm fuelled by coffee: what is the TPU measured difference between 8GB and 16GB?

RTX 5060 Ti is likely aimed at solid 1440p performance across most titles

From 5070 FE review: Ray Tracing performance at 1440p is 73.1 FPS (averaged all results in tables), so working that backwards:

• 4060 Ti 16GB = 47.5 fps
• 4060 Ti 8GB = 36.5 fps

-23% reduction in performance for 8GB

Several of those RT benchmarks are in the 30fps range, even the 16gb model, so I'd argue most gamers will not accept that.

So there's three options:

1. Don't use Ray Tracing. Average jumps up to 70 FPS at 1440p and we get these results: just a1% difference between 8GB and 16GB with 4060 Ti.
2. Drop resolution to 1080p with RT enabled. 16GB will have a 3% advantage over 8GB.
3. Deploy DLSS. But I can't find TPU performance data comparing 4060ti 8GB to 16GB with DLSS 3.5 or DLSS4.

Obviously there are lots more games out there, plus the mods people do, and most gamers run background apps that can consume VRAM...

But TPU data is clear: 16GB only offers advantage in unique (RT) situations, and not across the spectrum of performance analysis. I don't expect that to be different with 5060 Ti variants.

 

[Lew Zealand]

I think the above analysis is fine but I'll chime in as the owner of a 4060 Ti 8GB. Some games I play:

• Minecraft with Path Traced Shaders and a realistic texture pack (SEUS PTGI and Patrix 64 or 128) - this will fill the 8GB buffer quickly and severely degrade FPS and frametimes. So I keep Path Tracing and revert to the default textures as Path Tracing is so goddam the future, even in MC. RT looks like regular lighting in CP2077 so I find it useless, but Path Tracing in CP2077. Effing fantastic.
• Hogwarts Legacy - this game wants more than 8GB and fails back gracefully when only given 8GB, but many textures will load slowly which is distracting
• Ark SA - this game wants more than 8GB but will play mostly OK at 8GB with texture pool set to low. But with flaws— OK amazingly with even more flaws than the typical sh**fest that Ark always has been. So maybe not the best example but I do play the thing and it needs more VRAM.

The 4060 Ti is in the kid's PC now and doesn't do VRAM-heavy games so it's perfect, but then I use a 3070. Which is functionally the exact same card with 50% more power use, so all the experiences are the same. But with a warmer room!

I would enjoy all 3 of these more with 12+GB and my B580 does the latter 2 very well. However those Path Traced MC shaders only work well on Nvidia so I have zero cards that can truly let those shine with a good texture pack, sigh.

For this reason alone I may find myself buying the 16GB 5060 Ti at some point when I find a tolerable price.

 

[Macro Device]

16GB only offers advantage in unique (RT) situations, and not across the spectrum of performance analysis

Despite being painfully close, Ada and Blackwell are different architectures. VRAM speed has also been vastly increased for the latter.

4060 Ti at 8 GB is bad not because of 8 GB but because of how much it costs.
5060 Ti at 8 GB will be bad for BOTH reasons and not because it's 2025 and 8 GB is questionable but because 5060 Ti has enough muscle to meaningfully enable more VRAM. This means we will have more than 1% difference between 8 and 16 GB SKUs even without any RT whatsoever. This means we'll have two inherently terrible SKUs: one hampered by its low VRAM capacity (and questionable price); another twisted by its beyond insane pricing.

Both NVIDIA and AMD want to make a whole circus of all this. BUY ABSOLUTELY NOTHING, UNLESS YOU REALLY, REALLY NEED THIS OR THAT EXCRUCIATINGLY OVERPRICED PIECE OF HOT SILICON TURD.

 

[Dawora]

My 3060 ti stays relevant longer then ... the castrated 128 bit bus is just insulting...

Who cares about 128bit
it gives good memory bandwidth 448 GB/s

Just like 3060Ti whit 256Bit = 448 GB/s

What's insulting is that this has still less RAM than a plain 3060 that came out 4 years ago. Even more insulting for the 8 GB Ti because the 3060 was a lower price tier.


Actually the 192-bit bus would at least give us 12 GB of RAM, which would be the minimum decent amount for the $400-$500 Ti models.
(Sure, they could also do 12 GB with the 128-bit bus now that 3 GB modules exist, but I assume that they're too expensive)

5060Ti = 16GB vram so what is a problem

u get insulted if there is 8Gb GPUs even u dont need to buy those (Buying is optional) if u didint know

And 5060Ti 16Gb is more than 12Gb

16-12 = 4 so 4Gb More Vram

Let's see how long the "there's no bad card just bad prices" argument still holds for 8 GB cards.

And still 8Gb vram Gpus is most used, ppls are not like us in forums who whining and crying prices,vram and whatever they see..

 

[sLowEnd]

The die sizes of the chips on these cards is probably similarly small to their Ada predecessors, so the bus width isn't really a surprise. Because of how the memory controllers are laid out along the edge of the dies, it's less of an added cost to fit a larger bus on a big die than it would be if trying to shoehorn one onto a small die.

 

[AusWolf]

And still 8Gb vram Gpus is most used, ppls are not like us in forums who whining and crying prices,vram and whatever they see..

Oddly enough, those are the people who tend to keep their hardware the longest.

no u cant

Sorry, you're right. The 7600 is £10 more expensive (239 vs 229). My info was a couple of months old.

 

[mate123]

Who cares about 128bit
it gives good memory bandwidth 448 GB/s

Just like 3060Ti whit 256Bit = 448 GB/s

Yeah, great stagnation here... actually my 3060 TI with gddr6x has 640 GB/s but that is not the main issue here
A less complex chip will be sold for more $ than the one over 4 years ago...

 

[Prima.Vera]

…that’s not even close to how things work. Like, by that logic high-end card nowadays should run, what, 4096-bit buses? 8192-bit?
Oh, and no, no those cards won’t choke in 1080p. Overall bandwidth still should go up due to GDDR7.

That's exactly how it should have been! The HBM2 professional video cards are exactly that, with 4096-bit buses. Those are the real progress and advancement. Of course now we have GDDRX7 that mittigate some of the bandwidth issues, however 128-bit is still half of everything compared to a 256-bit bus.
Off course you can run a game in HD resolution using all the tricks those cards have, including DLSS and frame gen, and for some people is good enough. Especially for game engines that are not resources hungry.
For those that are.... I really cannot wait for some benchmarks using the latest UE5, CryEngine or Unity engine games.

 

[AusWolf]

That's exactly how it should have been! The HBM2 professional video cards are exactly that, with 4096-bit buses. Those are the real progress and advancement. Of course now we have GDDRX7 that mittigate some of the bandwidth issues, however 128-bit is still half of everything compared to a 256-bit bus.

If you want a consumer card with HBM, look at the Radeon R9 Fury, or the Vega 56/64. They were expensive to make, and not much faster than the competition, if any. There's more to a GPU than VRAM bandwidth.

Edit: The R9 fury, for example, uses 512 GB/s. Even a 7700 XT doesn't have that much bandwidth, only 432 GB/s. Guess which one is faster.

 

[Macro Device]

Not to forget nor underestimate the advantages of better compression and prediction. Maxwell, for example, had much less raw VRAM bandwidth than Kepler (e.g. GTX 960 had 112 GB/s versus 192 GB/s in GTX 760) but it still didn't seem to struggle more.

Caching is improving, latencies are going down. Also you can overclock GDDR7 obscenely high. 128 bit bus is not a problem at this level of performance. 160 at 10 GB would've been better and probably a sweet spot but we all know how NVIDIA works: provide as little as you can, AMD won't bother outclassing you anyway (9070 series is another one proof: their MSRPs are so fake it's even more fake than the NVIDIA's so instead of becoming the value king, AMD once again did NVIDIA minus 50 bucks (this time, even less than 50) by not flooding the market with reference models).

Also, @Prima.Vera, not to discard HBM longevity issues. HBM chips only can do a limited amount of rewrites and become e-waste thereafter. And it's way too limited for it to be fine for gaming in this day and age when your high tier GPU becomes unusable more than a decade after release. Dying after five years would be problematic. Like, 2080 is seven years old. Is it unusable? No, it's still a great 1080p card. Imagine 2080s massively dying throughout 2025...27 because of HBM limitations. Not cool. It's also much more expensive than classic GDDR.

 

[AusWolf]

Caching is improving, latencies are going down. Also you can overclock GDDR7 obscenely high. 128 bit bus is not a problem at this level of performance. 160 at 10 GB would've been better and probably a sweet spot but we all know how NVIDIA works: provide as little as you can, AMD won't bother outclassing you anyway (9070 series is another one proof: their MSRPs are so fake it's even more fake than the NVIDIA's so instead of becoming the value king, AMD once again did NVIDIA minus 50 bucks (this time, even less than 50) by not flooding the market with reference models).

Huh?

These are the cheapest cards currently listed at Scan UK (the rest of them are either more expensive, or don't have their price listed).

Also, @Prima.Vera, not to discard HBM longevity issues. HBM chips only can do a limited amount of rewrites and become e-waste thereafter. And it's way too limited for it to be fine for gaming in this day and age when your high tier GPU becomes unusable more than a decade after release. Dying after five years would be problematic. Like, 2080 is seven years old. Is it unusable? No, it's still a great 1080p card. Imagine 2080s massively dying throughout 2025...27 because of HBM limitations. Not cool. It's also much more expensive than classic GDDR.

Wow, I didn't know that. I already thought HBM for a consumer card was a bad idea because of its price, but this... unacceptable.

 

[Macro Device]

UK

It's leagues worse in the rest of the world. We have 5070 Ti and 9070 XT going for $1200+ and $1000+ respectively which makes both of these products too expensive to be any relevant.

AND I ASSURE YOU. RUSSIA DOESN'T HAVE IT THE WORST.

 

[AusWolf]

It's leagues worse in the rest of the world. We have 5070 Ti and 9070 XT going for $1200+ and $1000+ respectively which makes both of these products too expensive to be any relevant.
View attachment 388990

I agree - both are way too expensive. I find even the current UK price questionable. But the point is, it's not Nvidia -$50.

 

[Macro Device]

it's not Nvidia -$50.

Sure. It's NVIDIA (MSRP) + $250.

 

[AusWolf]

Sure. It's NVIDIA (MSRP) + $250.

Did you look at my post? Even a mid-range 9070 XT is a good £150-160 cheaper here than a base model 5070 Ti. Then you yourself mentioned $1000 vs $1200 respectively.

 

[Macro Device]

Even a mid-range 9070 XT is a good £150-160 cheaper here than a base model 5070 Ti.

Still zero MSRP GPUs.
Still horrible situation outside US and UK.
Still barely existent stock.

If it's not AMD's worst fuck up of this century I don't know what is. Even Bulldozer wasn't that bad.

 

[AusWolf]

Still zero MSRP GPUs.
Still horrible situation outside US and UK.
Still barely existent stock.

If it's not AMD's worst fuck up of this century I don't know what is. Even Bulldozer wasn't that bad.

Where is Nvidia's MSRP? Where is Nvidia's stock?
At Scan UK, there are two 5070 cards listed as available - The MSi Gaming Trio OC and Gaming Trio OC White for £710 and £720 respectively. Apart from them, the entire 50-series is nonexistent.

You're talking like AMD was alone with these problems.

 

[Macro Device]

You're talking like AMD was alone with these problems.

No? I was saying this paper launch of 5000 series was so bad you (as AMD) must try your hardest to be worse than them.

They managed. AMD were dead to me long ago, now they're innecromancible.

 

[AusWolf]

No? I was saying this paper launch of 5000 series was so bad you (as AMD) must try your hardest to be worse than them.

They managed.

It wasn't worse. There was more 9070 and 9070 XT stock at Overclockers UK ready for launch than all 50-series cards for the entire USA. The XT is gone now because the market is so starved, but several non-XT cards can still be found starting from £620.

No? I was saying this paper launch of 5000 series was so bad you (as AMD) must try your hardest to be worse than them.

They managed. AMD were dead to me long ago, now they're innecromancible.

Some more info on Microcenter USA (across the entire country):
Stock status at 2025-03-06 09:18:16 EST:

5070: 182 units,
5070 Ti: 1 unit,
5080: 4 units,
5090: 4 units.

9070: 3113 units,
9070 XT: 8589 units.

A day before, they launched the 5070 with 1374 units (again, in a country of 300 million people).

https://www.techpowerup.com/forums/threads/microcenter-gpu-stock-status.331118/post-5468360

Sure, AMD paper launch.

 

[maxli86]

Please don't come at me when I say this.
I hope Nvidia release a GTX5030 Low Profile card.
Seriously GT1030 for Low profile PCs as well as entry/basic use PC hasn't had a successor.
RTX 3050 is trash seriously so hopefully this new RTX 5050 can perform well.

 

[Lew Zealand]

I hope Nvidia release a GTX5030 Low Profile card.

RTX 3050 is trash seriously so hopefully this new RTX 5050 can perform well.

What? Those 2 statements are in opposition to each other.

The 3050 8GB and even 6GB cards are worlds better than the 1030, especially relative to their release dates. So why are you asking for a 5030? The best option will be a 70W slot-powered 5050 but seeing as these ones are targetting 130W, it seems unlikely. More likely we'll see a 70W 4050 (matching the laptop 4050's specs) released after the 5050, just like Nvidia did with the 3050 6GB.