youtube Nvidia CES 2025 event today

[UltraMega]

 

 

Expecting a 5000 series reveal.

 

[Slaughtahouse]

Cliff notes:

 

$549 5070

$749 5070 Ti

$999 5080

$1999 5090

 

5070 = 4090 “performance”.

DLSS 4 = More frame gen, exclusive to 50 series*

 

January availability.

Edited January 7 by Slaughtahouse

[firedfly]

9 minutes ago, Slaughtahouse said:

January availability.

 

January 30 for the 5080 and 5090.  5070 and 5070ti are February (no specific date)

[Sir Beregond]

2 hours ago, Slaughtahouse said:

Cliff notes:

 

$549 5070

$749 5070 Ti

$999 5080

$1999 5090

 

5070 = 4090 “performance”.

DLSS 4 = neural network features, like neural materials. Think high res assets, textures but at a lower performance cost.

 

January availability.

I'm actually pleasantly surprised by this. I fully expected them to go full nGreedia mode activated with a $2499 5090 and $1199 5080.

[UltraMega]

Quote

The new GPUs

The new RTX 5090 flagship comes packing significantly more hardware over its predecessor. Not only does this GPU use Nvidia's new Blackwell architecture, but it also packs significantly more CUDA cores, greater memory bandwidth, and a higher VRAM capacity. The SM count has increased from 128 with the RTX 4090 to a whopping 170 with the RTX 5090 – a 33% increase in the core size. However, GPU boost clocks have decreased slightly, from 2.5 GHz to 2.4 GHz.

The memory subsystem is overhauled, now featuring GDDR7 technology on a massive 512-bit bus. With this GDDR7 memory clocked at 28 Gbps, memory bandwidth reaches 1,792 GB/s – a near 80% increase over the RTX 4090's bandwidth. It also includes 32GB of VRAM, the most Nvidia has ever provided on a consumer GPU.
 

With Blackwell still built using TSMC 4nm technology, the board power rating has risen to accommodate these cores and memory, increasing to a mammoth 575W – 28% higher than the RTX 4090.

The RTX 5080 uses a much smaller GPU configuration, based on an entirely different GPU die, with just 84 SMs – four more than the RTX 4080 Super from the last generation.

 

 

 

 

 

Ram size is still going to be insulting to the lower end, but at least they're establishing the 90 cards as a true Halo product against the rest of the stack this time. 

Edited January 7 by UltraMega

[UltraMega]

1 hour ago, Sir Beregond said:

I'm actually pleasantly surprised by this. I fully expected them to go full nGreedia mode activated with a $2499 5090 and $1199 5080.

Nvidia is not magic. They can't afford to charge a premium for being the closest to the wall forever. Everyone is about to hit the wall face first. 

[Slaughtahouse]

Small correction since I checked out of the live stream after it was AI, AI, AI for 30mns straight. DLSS 4 is apparently simply more frame gen stuff. Exclusive to 50 series to no surprise.  Also, they’re now calling DLSS (super resolution upscaler) a beta?

 

 

 

Just a moment...

VIDEOCARDZ.COM

 

Edited January 7 by Slaughtahouse

[Sir Beregond]

8 hours ago, UltraMega said:

Nvidia is not magic. They can't afford to charge a premium for being the closest to the wall forever. Everyone is about to hit the wall face first. 

Speaking of a wall....I suspect any traditional raster improvement over 40-series is going to look minimal outside of the 5090. Ada gen got a lot of its improvements simply from the move from Samsung 8nm (a 10nm process) to TSMC N4 (a 5nm process). Given Blackwell sticks to TSMC 5nm (but I think the rumor is moving to N4P), most of it's improvements are going to have to come from architecture, clocks, power draw, and larger dies...and that only gets you so far. We at least also have the move from G6X to G7 memory which should help a lot as well. But otherwise? AI AI AI AI AI. Presumably the RT performance in the very specific scenarios Nvidia used is where you'll see the biggest improvements, otherwise...

 

But I also fully expected nGreedia mode because AMD simply is not going to compete at many levels here with RDNA4...they have stated as much.

Edited January 7 by Sir Beregond

[bridgypoo]

Not worth it to upgrade the 4090 imo

[Fluxmaven]

25 minutes ago, bridgypoo said:

Not worth it to upgrade the 4090 imo

 

You're just saying that because you're tired of rebuilding that rig 

 

And spending money of course

[Sir Beregond]

Edited January 7 by Sir Beregond

[Slaughtahouse]

4 hours ago, Sir Beregond said:

 

Yea… would love to see how the 5070 hits 4090 performance. Maybe if the 4090 has a partially melted 12pin cable and can only utilize half the power… maybe, just maybe, the 5070 will edge it out.

 

[UltraMega]

On 07/01/2025 at 08:08, Sir Beregond said:

Speaking of a wall....I suspect any traditional raster improvement over 40-series is going to look minimal outside of the 5090. Ada gen got a lot of its improvements simply from the move from Samsung 8nm (a 10nm process) to TSMC N4 (a 5nm process). Given Blackwell sticks to TSMC 5nm (but I think the rumor is moving to N4P), most of it's improvements are going to have to come from architecture, clocks, power draw, and larger dies...and that only gets you so far. We at least also have the move from G6X to G7 memory which should help a lot as well. But otherwise? AI AI AI AI AI. Presumably the RT performance in the very specific scenarios Nvidia used is where you'll see the biggest improvements, otherwise...

 

But I also fully expected nGreedia mode because AMD simply is not going to compete at many levels here with RDNA4...they have stated as much.

Looks like you were right. Native CP2077 performance went from 24 to 27. 

[Slaughtahouse]

https://videocardz.com/newz/nvidia-reveals-die-sizes-for-gb200-blackwell-gpus-gb202-is-750mm²-features-92-2b-transistors

 

New info highlights that consumer Blackwell is not on the further refined 5nm node, 4NP (5nm++). Instead, they keep the exact same node from Ada, 4N (5nm+).

[Sir Beregond]

7 hours ago, Slaughtahouse said:

https://videocardz.com/newz/nvidia-reveals-die-sizes-for-gb200-blackwell-gpus-gb202-is-750mm²-features-92-2b-transistors

 

New info highlights that consumer Blackwell is not on the further refined 5nm node, 4NP (5nm++). Instead, they keep the exact same node from Ada, 4N (5nm+).

Interesting. Explains the power increase somewhat.

[UltraMega]

1 hour ago, Sir Beregond said:

Interesting. Explains the power increase somewhat.

More fps for more power and die.

 

User meets wall.

 

 

I've been talking about the wall for a while now. Here is the latest example. 

 

we are 10-30 years out from getting around this. 

Edited Friday at 10:12 by UltraMega

[Slaughtahouse]

4 hours ago, UltraMega said:

More fps for more power and die.

 

User meets wall.

 

 

I've been talking about the wall for a while now. Here is the latest example. 

 

we are 10-30 years out from getting around this. 

 

What do you believe will disrupt this wall and why would it take 10-30 years? 

 

In the short term, I believe we all agree that 2 year cycle of GPU architecture / node swaps is dead. If GPU vendors want to release new GPU SKUs every year for constant cash flow and revenue, expect more iterative changes. Node changes likely will be 4 years+ at this rate considering the costs and yields. 

 

The latest roadmap we have from TSMC (Nov 2024) is linked below. I read it that it's likely consumer graphics wont be on 3nm until late 2026 or early 2027.

 

TSMC Technology Roadmap: 1.6nm Chip Manufacturing Set for 2026

WWW.GURU3D.COM

---

Taiwan Semiconductor Manufacturing Company (TSMC) has revealed its technology roadmap for the coming years at the Open Innovation Platform (OIP) 2024 conference in Amsterdam.

 

That would mean it would be approx. 5 years since Ada / RDNA 3 debuted with 5nm. Now I could be dead wrong and Nvidia launches their refresh of Blackwell on TSCM 3nm, but it seems unlikely when they're maximizing 3nm Blackwell for data centers. For context, the revenue from data center vertical was  10x that of consumer, gaming cards in the last quarter. $30B vs. $3B. 

 

NVIDIA Announces Financial Results for Third Quarter Fiscal 2025 | NVIDIA Newsroom

NVIDIANEWS.NVIDIA.COM

---

Record quarterly revenue of $35.1 billion, up 17% from Q2 and up 94% from a year agoRecord quarterly Data Center revenue of $30.8 billion, up 17% from Q2 and up 112% from a year ago SANTA ...

 

All of this is speculation of course...

 

Edited Friday at 14:33 by Slaughtahouse

[Kaz]

On 06/01/2025 at 22:51, Sir Beregond said:

I'm actually pleasantly surprised by this. I fully expected them to go full nGreedia mode activated with a $2499 5090 and $1199 5080.

 

The 4080 for $1200 didn't sell well, but it was priced that high because they wanted to sell 30 series stock.

 

On 07/01/2025 at 09:08, Sir Beregond said:

Speaking of a wall....I suspect any traditional raster improvement over 40-series is going to look minimal outside of the 5090. Ada gen got a lot of its improvements simply from the move from Samsung 8nm (a 10nm process) to TSMC N4 (a 5nm process). Given Blackwell sticks to TSMC 5nm (but I think the rumor is moving to N4P), most of it's improvements are going to have to come from architecture, clocks, power draw, and larger dies...and that only gets you so far. We at least also have the move from G6X to G7 memory which should help a lot as well. But otherwise? AI AI AI AI AI. Presumably the RT performance in the very specific scenarios Nvidia used is where you'll see the biggest improvements, otherwise...

 

But I also fully expected nGreedia mode because AMD simply is not going to compete at many levels here with RDNA4...they have stated as much.

The big sales pitch is frame generation.  Supposedly they are only redrawing the parts of the screen that change, which allows them to boost FPS.  It feels more like an attempt to muddy the water and say they preform much better than the hardware does, but maybe the end results are that much better.

 

$549 5070 feels like a travesty for 192 bit memory bus.  This should be a $400 card, but I guess they are saving that for a 60 series where they skimp on ram even more.  Everyone is selling AI specific hardware, but I've yet to see consumer AI needs.  Is there an AI product that the general population is actually using?  I don't need my operating system to understand what I'm thinking before I think it.  If Microsoft cared about the end user experience they wouldn't make Windows 11 start menu suck with history turned off.  It's a blank page.  They intentionally make it crappy!  AI is useful for analysing mass amounts of data.  The only mass data that my computer has access to better than a server is my actions and movements.  We continue to grab more and more user data without securing the existing info structure we have.  How long until this Janga house falls down?  Is all this really about lighting in a game?

 

I'm reminded of a Marrowind quote.  "Confucius says, he who owns glass house, dress in basement."

[UltraMega]

5 hours ago, Slaughtahouse said:

 

What do you believe will disrupt this wall and why would it take 10-30 years? 

 

In the short term, I believe we all agree that 2 year cycle of GPU architecture / node swaps is dead. If GPU vendors want to release new GPU SKUs every year for constant cash flow and revenue, expect more iterative changes. Node changes likely will be 4 years+ at this rate considering the costs and yields. 

 

The latest roadmap we have from TSMC (Nov 2024) is linked below. I read it that it's likely consumer graphics wont be on 3nm until late 2026 or early 2027.

 

TSMC Technology Roadmap: 1.6nm Chip Manufacturing Set for 2026

WWW.GURU3D.COM

---

Taiwan Semiconductor Manufacturing Company (TSMC) has revealed its technology roadmap for the coming years at the Open Innovation Platform (OIP) 2024 conference in Amsterdam.

 

That would mean it would be approx. 5 years since Ada / RDNA 3 debuted with 5nm. Now I could be dead wrong and Nvidia launches their refresh of Blackwell on TSCM 3nm, but it seems unlikely when they're maximizing 3nm Blackwell for data centers. For context, the revenue from data center vertical was  10x that of consumer, gaming cards in the last quarter. $30B vs. $3B. 

 

NVIDIA Announces Financial Results for Third Quarter Fiscal 2025 | NVIDIA Newsroom

NVIDIANEWS.NVIDIA.COM

---

Record quarterly revenue of $35.1 billion, up 17% from Q2 and up 94% from a year agoRecord quarterly Data Center revenue of $30.8 billion, up 17% from Q2 and up 112% from a year ago SANTA ...

 

All of this is speculation of course...

 

 

It's pure speculation about what will disrupt this, but as we get closer and closer to the wall I'd imagine a lot of research will shift towards looking for new ways to get around current limits, and eventually we will get a breakthrough in something like maybe graphene or 3D transistors or something not yet imagined.