More GeForce RTX 5090 rumors emerge following Nvidia's unveiling of its Blackwell AI GPUs

D

DragonSlayer101

Posts: 372   +2
Staff
Rumor mill: Nvidia unveiled its next-gen Blackwell GPU platform this week at GTC 2024. However, the company said nothing about upcoming consumer GPUs based on the same technology, leaving many gamers fuming at the company's newfound priorities. Thankfully, a noted leaker has filled in some of the gaps with alleged key details about the GB202 GPU that's expected to underpin the flagship Blackwell gaming card, the GeForce RTX 5090.

This latest bit of information comes from tipster @kopite7kimi, who claims that the GB202 GPU will use the same TSMC 4N foundry node as the GB100 designed for AI acceleration and computing applications. He further clarified that, contrary to some reports, the TSMC 4N node is derived from the company's 5nm EUV process, not the 4nm.

In the same update, the tipster also claimed that this new node would offer a 30 percent increase in transistor density compared to TSMC's mainstream 4NP node. The 4NP node itself marked a 6 percent increase in transistor density and a 22 percent enhancement in power efficiency over N5. If these reports are accurate, it suggests that Blackwell utilizes a variant of TSMC's 5nm node, akin to Ada.

In a separate, now-deleted post, Kopite also suggested that with Blackwell, Nvidia is enhancing the L1 caches in the streaming multiprocessors, indicating efforts to boost GPU performance at the SM level. According to the post, the flagship GB202 gaming GPU will feature a "significant improvement" in L1 cache size over its predecessors, AD102 and GA102, which both have 128 KB L1 caches.

For now, there is no additional information about the RTX 5090, but recent leaks have already offered a glimpse into its capabilities. An earlier report indicated that the GB202 GPU might come with a 512-bit memory interface, an upgrade from the 384-bit interface of the flagship AD102 Ada GPU. Additionally, the RTX 5090 is rumored to be equipped with 28Gbps GDDR7 memory.

If the recent leaks turn out to be accurate, it could mean the RTX 5090 will be a performance beast, and one that AMD could find hard to beat. These are all unverified rumors, and they should be approached with caution for the time being. Nvidia's GeForce RTX 50 series is only expected to debut in late 2024 or early 2025, leaving room for developments and changes in the interim.

Permalink to story.

https://www.techspot.com/news/102327-more-rumored-rtx-5090-details-arrive-after-nvidia.html

 
With these numbers is even a 50% improvement over the 4090 possible?
30% more dense, while being 20% more efficient and GDDR7 @ 28Gpbs. I guess it's possible but those 70% uplift seems doubtful with these numbers imo. Thoughts?
 
godrilla said:
With these numbers is even a 50% improvement over the 4090 possible?
30% more dense, while being 20% more efficient and GDDR7 @ 28Gpbs. I guess it's possible but those 70% uplift seems doubtful with these numbers imo. Thoughts?
Click to expand...
Says an increase in L1 cache so maybe that will get rid of some bottlenecks to the SM's.
 
“If the recent leaks turn out to be accurate, it could mean the RTX 5090 will be a performance beast, and one that AMD could find hard to beat.“

Even if the performance is the same as the 4090, AMD won’t be able to beat it…

AMD has already given up on the high end, they focus on low-mid and console contracts…
 
Performance is about 6% better and the density is around 30% more.

So basically, for the same die size, in the best case scenario, you would get 30% more performance.

I am not seeing where those leakers make their stuff up because on paper, the 5090 would barely be 30-40%.

People forget that Ampere was on Samsung 7nm and moving to TSMC give a major bump to performance alone.
 
A 50%+ boost to the 4090 would be a significant uplift to for the 5090, but Nvidia really needs to bring the gains down the stack. The 4060 is only 40% faster than the 2060. To put that in perspective, the 4090 is 140% faster than the 2080 Ti. The gap from enthusiast to 'mid-range' if you can still call it that, grew with the 30 series, but exploded with the 40 series. The 60 level GPU really needs to see a 75% + boost with the 50 series, otherwise the 60 level will just be irrelevant.
 
  • Like
Reactions: ddferrari
Theinsanegamer said:
A 30% increase in density over the 4090, with a 512 bit bus, is gonna be a monster no matter how you cut it.
Click to expand...
The problem is that 30% more cores would not result in 30% more performances.

Ampere doubled the Core counts and the 3080 was barely 30% faster than the 2080 TI.

You can put the BUS that you want, it will not change much beside not be limited by memory at higher resolution.
 
  • Like
Reactions: Manoa
EdmondRC said:
A 50%+ boost to the 4090 would be a significant uplift to for the 5090, but Nvidia really needs to bring the gains down the stack. The 4060 is only 40% faster than the 2060. To put that in perspective, the 4090 is 140% faster than the 2080 Ti. The gap from enthusiast to 'mid-range' if you can still call it that, grew with the 30 series, but exploded with the 40 series. The 60 level GPU really needs to see a 75% + boost with the 50 series, otherwise the 60 level will just be irrelevant.
Click to expand...
They can't get an additional performance boost from passing from a gimp node to a great one. That's why the 4090 had a 20% extra performance available over the expected 40% uplift.

But this card has been played. Now they are passing to another TSMC node with a marginal performance increment and 30% better density.

There is no miracle to pull here.
 
Squid Surprise said:
“If the recent leaks turn out to be accurate, it could mean the RTX 5090 will be a performance beast, and one that AMD could find hard to beat.“

Even if the performance is the same as the 4090, AMD won’t be able to beat it…

AMD has already given up on the high end, they focus on low-mid and console contracts…
Click to expand...
2000$ is not the high end, it is the whale tier...
 
  • Like
Reactions: Squid Surprise
redgarl said:
They can't get an additional performance boost from passing from a gimp node to a great one. That's why the 4090 had a 20% extra performance available over the expected 40% uplift.

But this card has been played. Now they are passing to another TSMC node with a marginal performance increment and 30% better density.

There is no miracle to pull here.
Click to expand...
I think my main point was that the high-end GPUs really don't need a massive performance boost, it's the mid-range, especially the lower mid-range that need a big boost. Right now, there is a 3X delta between the 4090 and 4060. The 2080 Ti was ~1.8X faster than the 2060, the 3090 was ~2.1X faster than the 3060, the 4090 is ~3X faster than the 4060. If that trend continues, the 5060 will be irrelevant no matter how fast the 5090 is, the 5060 needs a huge boost, it needs to be at the minimum as fast as the 4070, but probably should be more like a 4070S in terms of performance with at least 12GB VRAM and a maximum $349 price tag. And that's not too much to ask, though it sounds crazy given the past few generations. Even with a huge performance jump like that, the 5090 would still be more than 2X the 5060 in performance, but it would at least reverse the trend of making mid-range GPUs irrelevant. The 4060 is a terrible desktop GPU, it's not a bad laptop GPU considering its low power consumption with some laptops running it just as fast as the desktop variant, but it's still a terrible desktop option.
 
redgarl said:
The problem is that 30% more cores would not result in 30% more performances.

Ampere doubled the Core counts and the 3080 was barely 30% faster than the 2080 TI.

You can put the BUS that you want, it will not change much beside not be limited by memory at higher resolution.
Click to expand...
That's because Ampere doubled the SHADER count per SM, not total core count.

Interestingly, The 4080 is faster then the 3080 despite fewer cores. Hmmmm.....

Unless the CUDA core per SM changes again, a 30% core increase is gonna give you a 30% increase in performance, clock rates none withstanding.
 
Theinsanegamer said:
A 30% increase in density over the 4090, with a 512 bit bus, is gonna be a monster no matter how you cut it.
Click to expand...

A monster? Compared to what a big 3nm chip would have been? Not at all. They are not getting 60-70% gen to gen improvement out of this for sure. Probably not even 50%

512 bit and GDDR7 means nothing when even 4090 is not really bandwidth starved to begin with.
Bandwidth and memory might deliver a 5-10% gain and only at 4K/UHD minimum, and only in some games.

It seems that TSMC is not ready for big chips with 3nm yet. I think Nvidia was forced to use 4nm again (or they would have used 3nm for sure on AI cards).

I bet this is why we never saw a 4090 Ti. Nvidia knew they had to use 4nm again. After all 4090 is 12% cutdown. AD102 has 18432 cores.

5090 could have been an absolute monster on 3nm, on 4nm (5nm) again, not so much. Same process as 4000 series already use and GDDR7 won't be a massive uplift from GDDR6X, especially not if Nvidia uses 28 Gbps modules instead of 32+

This probably means 5090 will be cheaper tho. Still like 6-9 months away.

Maybe 5000 series will be all about DLSS 4.x (which is mostly about Ray Tracing this time)

I wonder what Intel will do with Arrow Lake if TSMC 3nm is not really working properly for bigger dies yet. Apple used 3nm for like 1+ year now, but for smaller chips.

Also might be 2nm TSMC that is delayed which means Apple will be locked in on 3nm for longer. And Apple is king at TSMC.
 
Last edited:
  • Like
Reactions: godrilla
Theinsanegamer said:
That's because Ampere doubled the SHADER count per SM, not total core count.
Click to expand...

Typical replies from typical Nvidia fanboys...

I remember this one pretty well because Kopite and the other leakers never though Nvidia would double the core counts and they all were fed false information until the launch day where Jensen said Ampere was having 2 times more cores than Turing.

cores.jpg
 
Last edited:
redgarl said:
Typical replies from typical Nvidia fanboys...

I remember this one pretty well because Kopite and the other leakers never though Nvidia would double the core counts and they all were fed false information until the launch day where Jensen said Ampere was having 2 times more cores than Turing.

cores.jpg
Click to expand...
Leakers also got Pascal launch totally wrong.
What we have been hearing for the past year is that Nvidia will prioritize Ai over gaming. Maybe the 3nm process is reserved for ai. Also we heard that Nvidia is going the MCM route what other way to introduce the most dense MCM gpu other than smaller chiplets from poor yields on newest node.
In the end Nvidia could just dedicate the whole lot to Ai and give the middle finger to gamers.
I personally doubt they will do that and rather do paper launches with fat margins and scale accordingly but then again I am biased towards gaming.
 
  • Like
Reactions: Theinsanegamer
redgarl said:
Typical replies from typical Nvidia fanboys...

I remember this one pretty well because Kopite and the other leakers never though Nvidia would double the core counts and they all were fed false information until the launch day where Jensen said Ampere was having 2 times more cores than Turing.

cores.jpg
Click to expand...
Do you....think before you type? You just proved my point, which was that claiming that increased core count =! performance because AMpere doubled counts isnt a slam dunk, because Ampere double cuda count per SM, not total SM count. As Blackwell is NOT doing this, a 30% core count increase will likely correlate with a 30% performance increase.

And I've talked about my all AMD main build, my AMD laptop, and my AMD GPU in my media pc here before, but sure, I'm an nvidia fanboi. LMFAO XD
Disiw said:
A monster? Compared to what a big 3nm chip would have been? Not at all. They are not getting 60-70% gen to gen improvement out of this for sure. Probably not even 50%

512 bit and GDDR7 means nothing when even 4090 is not really bandwidth starved to begin with.
Bandwidth and memory might deliver a 5-10% gain and only at 4K/UHD minimum, and only in some games.

It seems that TSMC is not ready for big chips with 3nm yet. I think Nvidia was forced to use 4nm again (or they would have used 3nm for sure on AI cards).

I bet this is why we never saw a 4090 Ti. Nvidia knew they had to use 4nm again. After all 4090 is 12% cutdown. AD102 has 18432 cores.

5090 could have been an absolute monster on 3nm, on 4nm (5nm) again, not so much. Same process as 4000 series already use and GDDR7 won't be a massive uplift from GDDR6X, especially not if Nvidia uses 28 Gbps modules instead of 32+

This probably means 5090 will be cheaper tho. Still like 6-9 months away.

Maybe 5000 series will be all about DLSS 4.x (which is mostly about Ray Tracing this time)

I wonder what Intel will do with Arrow Lake if TSMC 3nm is not really working properly for bigger dies yet. Apple used 3nm for like 1+ year now, but for smaller chips.

Also might be 2nm TSMC that is delayed which means Apple will be locked in on 3nm for longer. And Apple is king at TSMC.
Click to expand...
30% over the current market leader, which can manage 4k60 in many modern titles, is a pretty monstrous chip. They wouldnt give it a 512 bit bus for no reason.
 
Theinsanegamer said:
Do you....think before you type? You just proved my point, which was that claiming that increased core count =! performance because AMpere doubled counts isnt a slam dunk, because Ampere double cuda count per SM, not total SM count. As Blackwell is NOT doing this, a 30% core count increase will likely correlate with a 30% performance increase.

And I've talked about my all AMD main build, my AMD laptop, and my AMD GPU in my media pc here before, but sure, I'm an nvidia fanboi. LMFAO XD

30% over the current market leader, which can manage 4k60 in many modern titles, is a pretty monstrous chip. They wouldnt give it a 512 bit bus for no reason.
Click to expand...
4k60 is kinda meh tho, when 4k240 is the new black.

I think 5090 will be mcm, like the blackwell ai cards.
 
You must log in or register to reply here.

Similar threads

D
Nvidia RTX 50 series Blackwell GPUs tipped to use 28Gbps GDDR7 memory with 512-bit interface
2
Replies
28
Views
706
mbk34
M
Daniel Sims
Graphics card makers believe the RTX 5090 and 5080 will ship in Q4 2024
Replies
24
Views
281
godrilla
godrilla
Daniel Sims
Nvidia's upcoming RTX 5000 memory interface boost might be exclusive to (very expensive)...
Replies
14
Views
289
Disiw
D

Latest posts

Back