Discussion Nvidia Blackwell in Q1-2025

[Win2012R2]

I'm pretty sure Nanite already does meshlet combination already.

This looks like hardware acceleration of BVH to make ray tracing a lot quicker, they'll incorporate it into UE5 branch, which is nice, but Epic really needs to solve this problem of people not being able to easily upgrade engine version - game dev takes years, and game popping in now with versions 5.1, it's a load of BS that they can't get newer versions for shipping - even if the game does not take advantage of new things, but stuff like better CPU multithreading should be in all UE5 releases even if they started few years ago.

 

[Win2012R2]

Short summary: "MLID: Everything is proceeding as I have foreseen..."

Wot about 60-70% raw perf in 5090?

Well DLSS was an upscaling tech that needed to work with the entire game. Those shaders are only needed for some specific shaders (effects) in some fixed parts of the world, so not entirely the same.

Not the same, no - but puts again a lot more onus on developers, which I think in this case was unavoidable, but we do know that developers don't have much time, especially that they have to get stuff working first on PS5 and then making PC version much better is a tad risky as Sony might be displeased...

 

[Win2012R2]

If the 5090 is indeed a 2 slot card and if it is at least 40% faster

It appears to have 25% more theoretical FP32, which is about what Far Cry 6 was getting and it scales well with GPU as long as CPU allows, maybe it's 40% faster when heavy ray tracing is involved, but CP2070 does not seem to show that either. Looks like it's worse than 2080 TI

 

[Timmah!]

Personally i think 5080 might be on the cheaper side, given similar number of cores in comparison to 4080, so the additional perf is gonna come mostly from arch improvements and all that AI shenanigans, and faster RAM. So imo 900 - 999 is more likely than 1200 - this not taking tarrifs into consideration.
5090 i can see then +200 to 400 compared to 4090, so 1800 to 1999, to account for the chip size and RAM capacity increase. I dont think they go up to 2500 (retailers might), but greed in unlimited, so you never know.

So my guess turned out to be right.
5090 gonna cost me 2500 EUR damn, and that is the lower estimation, cause some models will be sure up to 3000....even if i sell 4090 for 1000, it will still be hefty sum to pay.
Gonna have to find a way to buy it again without need to pay VAT, this time moreso than ever.

Did not watch the show, were there any performance numbers aside of those DLSS ones? I mean one related to other things than games?

 

[Win2012R2]

were there any performance numbers aside of those DLSS ones?

Like one, Far Cry 6 - around 20-25% faster, which is close to its FP32 increase over 4090.

 

[jpiniero]

I suspect what happened is that they went with a denser design iso-node that killed the clocks. To the point where the game clocks should be lower on Blackwell.

 

[insertcarehere]

It appears to have 25% more theoretical FP32, which is about what Far Cry 6 was getting and it scales well with GPU as long as CPU allows, maybe it's 40% faster when heavy ray tracing is involved, but CP2070 does not seem to show that either. Looks like it's worse than 2080 TI

Does it though? 4k ray tracing is as GPU bound a scenario as possible, a 14900k versus a 9800x3d won't change much when even chart topping 4090 is under 120fps..

Furthermore FC6 RT results are just... Not representative of how Nvidia and AMD cards normally fare with RT. Given the low FPS penalties involved, as well as the relative performance of RDNA keeping up with Ampere/Ada. 6800XTs and 7900XTs do not normally keep up with RTX 3080s and RTX 4080s under RT..

 

[Win2012R2]

Does it though?

Sorry, I did not mean to say 5090 improvement is worse than 2080 ti, but in the same ballpark - disappointingly low to upgrade from prev gen, 25% just does not cut the mustard.

Furthermore FC6 RT results are just... Not representative of how Nvidia and AMD cards normally fare with RT

True, but they also included CyberPunk 2070 which totally favour Nvidia tech, if you look at how low it is without DLSS you will see that it's not far off 4090 results (both trash unusable low)

 

[mzocyteae]

I'm looking through various vendors now, the thinnest I've found so far is:

asus tuf 5090: 348x146x72
asus ast 5090: 358x149x76

msi vent 5090: 325x139x67
msi trio 5090: 359x149x70
msi supr 5090: 359x150x76
msi vang 5090: 357x151x76

the photo of gigabyte windforce 5090 looks like it is around 3-slot (thinner than 70mm 4090)

 

[SolidQ]

CyberPunk 2070 which totally favour Nvidia tech

Raster favore to AMD, rest to NV

Black myth Wukong. Multifake Frames

 

[jpiniero]

Compubase claims that the 5070 does have 5.0x16. I was kind of expecting it to be x8.

 

[techjunkie123]

NVIDIA Project DIGITS: The World’s Smallest AI Supercomputer.

Reserve yours today.

www.nvidia.com

Haven't seen this been shared yet. Nvidia Mac mini format AI computer with Blackwell GPU, Grace ARM CPU with 10x X925 + 10x A725 cores from MediaTek, 128 GB of unified memory and 4 TB SSD, for 3k$.

Probably a preview of their WoA chip too. That will presumably be smaller but same idea.

 

[jpiniero]

NVIDIA Project DIGITS: The World’s Smallest AI Supercomputer.

Reserve yours today.

www.nvidia.com

Haven't seen this been shared yet. Nvidia Mac mini format AI computer with Blackwell GPU, Grace ARM CPU with 10x X925 + 10x A725 cores from MediaTek, 128 GB of unified memory and 4 TB SSD, for 3k$.

Probably a preview of their WoA chip too. That will presumably be smaller but same idea.

Oh interesting. It's like Strix Halo. Says it uses NVLink to connect to the Arm chip.

 

[Hulk]

In retrospect that entire nVidia performance last night was very cringe-worthy.
I kind of felt bad for Jensen because his performance was terrible, he was like a fish out of water. But then I remembered he has 100 billions dollars and stopped feeling sorry for him. It was like a stand up comedian when joke after joke just falls flat even though the crowd is with them and they won't boo, it's just a weird silence at the punch lines, or a couple people trying to laugh/cheer.

Also, as far as pricing they are taking a page out of the old Intel Extreme (profit) Edition CPU's. When you own the high end you can make your fastest "regular" part just a little more expensive than the competition. For CPU's it was about $500. Then you could price an EE CPU at double that. Don't put anything in between, make people who want more performance go all the way.

Same thing here. The 5080 at $1000 is competitive with anything AMD will most likely have. Why put better performing cards at $1300 and $1650 when there is no competition there? Just make people pay $2,000 to get the next performance level. It's smart and I'm sure the profit if off the charts.

So far so good for nVidia, they are following the footsteps of Intel from 15 years ago. They should have about 5 years before they start the long slide into the abyss.

 

[SiliconFly]

My DLSS 4 prediction two weeks back...

 

[In2Photos]

So Nvidia claimed 5070 with 4090 performance (of course there was no details about what they used to measure that). According to Tech Power Up the 4090 is on average 71% faster at 1440p than the 4070. And on the Nvidia slide on their website the performance increase from the 4070 to 5070 is only getting maybe 20-30% on the 2 games that don't utilize MFG. Man I really hate these marketing tactics.

 

[CakeMonster]

I wonder if the claimed improvements to DLSS on the existing cards (Turing and up) will make some people reconsider upgrading. If you have a 11GB card or more and DLSS2 gets noticeably better/faster maybe some people will just hold out.

 

[In2Photos]

Also, on one slide they said Blackwell has 4000 TOPS, but he 5090 only has 3352. So are they going to release a 5090 Ti or Super down the road? Or is that reserved for the workstation cards maybe?

 

[jpiniero]

Also, on one slide they said Blackwell has 4000 TOPS, but he 5090 only has 3352. So are they going to release a 5090 Ti or Super down the road? Or is that reserved for the workstation cards maybe?

There is room for a 5090 Ti. The 5090 is 170 SMs of 192 plus they could give a 5090 Ti 48 GB GDDR7 at faster speeds.

 

[In2Photos]

There is room for a 5090 Ti. The 5090 is 170 SMs of 192 plus they could give a 5090 Ti 48 GB GDDR7 at faster speeds.

Was the 4090 the full die? If so, then the 5090 pricing shouldn't be compared to the 4090 pricing as the halo part. Good ole nvidia!

 

[jpiniero]

Was the 4090 the full die?

No

 

[Josh128]

5090 $1899 minimum, wouldnt be surprised at all to see $1999. $2600 street price.

Called this biotch!

 

[Heartbreaker]

It is, and Jensen is a total nerd, but you can’t argue with the results. Their FE cards specifically are the best GPUs out there. My 4090 is whisper quiet in all except the most demanding stuff.

That new FE for the 5090 looked really cool. Small for high power card, and notice that both fans are flow through with only a tiny PCB in the middle.

 

[In2Photos]

That new FE for the 5090 looked really cool. Small for high power card, and notice that both fans are flow through with only a tiny PCB in the middle.

Which makes you wonder how they are getting the ports to the back of the card? Some sort of extension cable?

 

[Heartbreaker]

Called this biotch!

You and everyone else. I've said $1999 minimum ever since Kopite7kimi leaked the specs. There was no way that beastly HW was going to sell for less than that.