Discussion Nvidia Blackwell in Q1-2025

[beginner99]

And it seems clear why new improved SM did not make any difference - they can't do FP32/INT at the same time now! AFAIK before Ampere/Ada they could do FP32+INT or FP32+FP32, so now the only thing that changed is that they can do INT+INT too, which as we can see does not do a lot. It started with fake cores, now fake frames...

It might help with certain AI stuff as the whole gpu could now do just int32 all at the same time? else yeah it seems kind if pointless change. I assume a core can do int+int while another core does fp+fp now. vs the same to cores before doing a mixed workload. not sure if that changes much expect scheduling were which OP runs?

 

[Win2012R2]

yeah it seems kind if pointless change

Would not be pointless for games if those cores were true FP32+INT as in at the same time for all of them.

Maybe they just lacked transistor budget to do it and it's new foundation for the 60 series...

I assume a core can do int+int while another core does fp+fp now.

Not how I read it - each "core" (that consists of FP32/INT) can do either FP32 or INT, but that's it.

 

[Win2012R2]

5090 listed in Nvidia shop as "Coming Soon", what's the time sale is supposed to start?

Looks like at 2pm UK time

 

[SolidQ]

290? Nice margin for NV

Spitballing Nvidia's RTX 5090 GPU die manufacturing costs — die could cost as little as $290 to make

Not too cheap.

www.tomshardware.com

 

[Heartbreaker]

290? Nice margin for NV

Spitballing Nvidia's RTX 5090 GPU die manufacturing costs — die could cost as little as $290 to make

Not too cheap.

www.tomshardware.com

IMO, if that was true, AMD would make big die GPU to compete.

Both AMD and NVidia are behaving like the dies are ridiculously expensive at this process level.

 

[Win2012R2]

IMO, if that was true

Why would it not be?

4N had 2 years of tweaking, plus frequency is not pushed, so that means binning by it is a lot easier, a fair few bits on chip are disabled too, totally plausible - and yet has no effect on AMD's decision not to compete on high end.

 

[NTMBK]

Would not be pointless for games if those cores were true FP32+INT as in at the same time for all of them.

Maybe they just lacked transistor budget to do it and it's new foundation for the 60 series...



Not how I read it - each "core" (that consists of FP32/INT) can do either FP32 or INT, but that's it.

Nvidia GPUs work in a block of 32 "threads", called a warp. (It's more like a SIMD lane than a real thread or core, Nvidia terminology is garbage.) So all 32 "cores" in a warp are executing the same instruction at once, or are masked off and not doing anything.

 

[Win2012R2]

So all 32 "cores" in a warp are executing the same instruction at once, or are masked off and not doing anything.

Ok, so explain this then - why Turin supported FP32 and INT executing at the same time which gave speed up since 1/3 of instructions are INT? That did lead to real observed speed up in games.

 

[GTracing]

IMO, if that was true, AMD would make big die GPU to compete.

Both AMD and NVidia are behaving like the dies are ridiculously expensive at this process level.

You can do the math yourself. Nvidia's wafer cost might be a little higher or lower than the $16000 quoted by the article.

There are other costs. After it's produced it needs binning, packaging, and shipping. Then after it's sold to an AIB partner it needs 32GB of GDDR7, the PCB with all the electronic components, and the heatsink. The partner has their own overhead with customer support, marketing, warranty, sales, and shipping. Then the store takes their own margin.

But even after all those other costs, they're not that expensive. Nvidia has huge margins.

 

[Heartbreaker]

You can do the math yourself. Nvidia's wafer cost might be a little higher or lower than the $16000 quoted by the article.

There are other costs. After it's produced it needs binning, packaging, and shipping. Then after it's sold to an AIB partner it needs 32GB of GDDR7, the PCB with all the electronic components, and the heatsink. The partner has their own overhead with customer support, marketing, warranty, sales, and shipping. Then the store takes their own margin.

But even after all those other costs, they're not that expensive. Nvidia has huge margins.

Show me where TSMC advertises this $16000 number...

 

[GTracing]

Show me where TSMC advertises this $16000 number...

They don't. Wafer prices aren't public and different customers pay different amounts. But all the leaks put N4P in that range.

 

[Win2012R2]

Can't be ****ed to F5 hoping "coming soon" will turn in to product, Nvidia you can keep it...

 

[CakeMonster]

Yep this looks like a shitshow, following a few threads of potential buyers, and there were cards listed by scalpers a couple of minutes after the drop, while no one could load a site or even see the buy button on the actual shops.

 

[jpiniero]

There's an OOS 5080 with an "official" price of $99,999 at Newegg.

 

[Heartbreaker]

Microcenter in Denver CO:

View attachment 115913

I'm almost hoping they are poor students being paid well by scalpers because parking lot camping in Colorado in late January, just for a card marginally better (majority of stock are 5080s) than what was available for 2 years makes no real sense.

 

[pj-]

Will be interesting to see if 5090 remains impossible to get for a long time. Can't believe all microcenters in the US got 233 total

 

[Win2012R2]

3 trillion dollar corpo can't get enough stock for a product with 80%+ margin to at least keep people happy buying at overpriced "MSRP".

 

[Ranulf]

Seeing reports that Bestbuy sold out in 5mins starting at 9am EST.

 

[Win2012R2]

Well, this is a very good time for AMD to launch their new card - as long as it's in stock

 

[pj-]

Seeing reports that Bestbuy sold out in 5mins starting at 9am EST.

I looked at 9:00:xx and they were all immediately sold out

 

[Golgatha]

I looked at 9:00:xx and they were all immediately sold out

Saw in Discord Micro Center was allotted 5x 5090 and and 70x 5080 GPUs for sale. What a joke, but I guess that's one way to push folks to buy a 5080. As for me, I'm solidly in the wait for the 6000 series 2-3 years from now and enjoy gaming on my 4090. Maybe I'll just Sharpie 5080 Ti on my 4090 to remind myself what a horrible non-uplift the 5000 series is.

 

[CakeMonster]

Consolation prize, new driver is out (572.16).

• DLSS Transformer Model Upgrade - Enables the latest transformer AI model for DLSS Super Resolution, Ray Reconstruction, and DLAA for all GeForce RTX users, when the aforementioned features are ON in-game
• DLAA & Ultra Performance Modes - Sets the internal rendering resolution for DLSS Super Resolution, enabling DLAA or Ultra Performance mode in games lacking native support, when Super Resolution is ON in-game

 

[NTMBK]

Ok, so explain this then - why Turin supported FP32 and INT executing at the same time which gave speed up since 1/3 of instructions are INT? That did lead to real observed speed up in games.

It dual issues a vector FP instruction at the same time as a vector INT instruction.

 

[MrTeal]

Seeing reports that Bestbuy sold out in 5mins starting at 9am EST.

BB Canada was broken all morning. I'm not aware of anyone who actually bought one manually. Apparently Stellar took the majority of the stock.

 

[adroc_thurston]

AMD would make big die GPU to compete.

AMD doesn't make big dies. They make big packages.