Discussion Nvidia Blackwell in Q1-2025

[Meteor Late]

W r o n g.

Look, I'm not only basing this on how bad Blackwell is. I'm just watching performance jumps relative to transistor count increase.
In the past, we had cards like 1080 ti that didn't increase transistor count all that much and also didn't get a huge bandwidth improvement like Blackwell and still got a massive performance improvement over its predecessor.
What we have nowadays, is a 4090 that, while seemingly impressive at first glance, was what, like 50% faster than 3090 ti and 60-65% higher than the cut down version 3090, all the while having an insane 2.7x the transistor count, while 1080 ti got higher performance delta while having like 1.5x the transistor count.
And you may say, well, that's Nvidia, AMD is much better, 9070 XT is a good jump in PPA in the same node. Well yeah, but 7900 XTX was a failure in meeting the performance target.

 

[jpiniero]

while 1080 ti got higher performance delta while having like 1.5x the transistor count.

1080 Ti was on a better node tho. TSMC marketing says 16 nm is 50% faster than just 20 nm.

 

[Golgatha]

Taking a step back, I was looking at a few older reviews to see why I purchased the 4090 I have now. I was coming from a 3080 Ti, which retailed for $1200 and for $1600 or 33% more money, I was getting nearly double the performance. I also got double the VRAM compared to the 3080 Ti, which has made my current 4090 age really, really well even only 2 years out from the purchase. There are games like Diablo 4 and Cyberpunk 2077, which I'm playing regularly, that when turned up utilize 15-16GB of VRAM.

In comparison, at least 25% more money for roughly 20-25% performance uplift seems like a horrible deal. Add on top of this that 24GB vs 32GB of VRAM most likely isn't going to be an issue for a long time to come, if ever given the raster performance is at best 25% better. I'm likely to run out of GPU performance before lack of VRAM would become an issue over the next few years.

 

[blackangus]

The AIB pricing is outrageous for the most part if the leaks are correct. I be really curious to see how sales go at those prices after the 1st 3 months.

 

[jpiniero]

The AIB pricing is outrageous for the most part if the leaks are correct. I be really curious to see how sales go at those prices after the 1st 3 months.

TBF, the leaked 5080 prices aren't that far off from AIB 4080 Supers.

 

[blckgrffn]

TBF, the leaked 5080 prices aren't that far off from AIB 4080 Supers.

TBF, it's basically a GDDR7 4080 Super.

 

[Thunder 57]

Taking a step back, I was looking at a few older reviews to see why I purchased the 4090 I have now. I was coming from a 3080 Ti, which retailed for $1200 and for $1600 or 33% more money, I was getting nearly double the performance. I also got double the VRAM compared to the 3080 Ti, which has made my current 4090 age really, really well even only 2 years out from the purchase. There are games like Diablo 4 and Cyberpunk 2077, which I'm playing regularly, that when turned up utilize 15-16GB of VRAM.

In comparison, at least 25% more money for roughly 20-25% performance uplift seems like a horrible deal. Add on top of this that 24GB vs 32GB of VRAM most likely isn't going to be an issue for a long time to come, if ever given the raster performance is at best 25% better. I'm likely to run out of GPU performance before lack of VRAM would become an issue over the next few years.

Reee but it's the fastest GPU ever and worthy of at least 4.5 stars. That's about the only argument I am hearing and it falls totally flat for reasons you have outlined. The same 4.6 stars that the objectivly better 4090 got. I'm surprised they haven't published an article (opinion?) yet saying "just buy it" like they did with the 3080.

 

[jpiniero]

TBF, it's basically a GDDR7 4080 Super.

True.

 

[blackangus]

TBF, the leaked 5080 prices aren't that far off from AIB 4080 Supers.

And those are crazy at the moment too.

 

[maddie]

And those are crazy at the moment too.

I say let them keep it. Covid marked a pricing inflection point for companies psychology.

 

[Thunder 57]

I say let them keep it. Covid marked a pricing inflection point for companies psychology.

Well yes. They realized people would pay much higher than what used to be common prices and they would still sell out. That keeps happening so they have no incentive to go back to lower prices.

 

[Meteor Late]

1080 Ti was on a better node tho. TSMC marketing says 16 nm is 50% faster than just 20 nm.

4090 was also a massive node jump over 3090 and 3090 ti, TSMC 7nm was much better in terms of performance than Samsung 8nm and Nvidia jumped to 5nm not 7, and look at the density difference, almost 3 times higher density from Samsung 8nm to TSMC 5nm!

 

[poke01]

Geekerwan review

 

[adroc_thurston]

TSMC 7nm was much better in terms of performance than Samsung 8nm

It wasn't. It was lot less dense, though.

and look at the density difference, almost 3 times higher density from Samsung 8nm to TSMC 5nm!

Yeah TSM nodes are really good at density.

 

[Thunder 57]

Interesting post on TPU.

Bodes really bad for 5070 which has 6144 cores vs 7168 of the 4070 Super and they cannot really clock it any higher due to same semiconductor process.

Maybe that is why Nvidia is being so generous with the 5070 price?

 

[MrTeal]

Interesting post on TPU.



Maybe that is why Nvidia is being so generous with the 5070 price?

What's a 4070 Super?
It's going to be twice as fast* as a 4070, and $50 cheaper. You'd be crazy to not buy one.

 

[igor_kavinski]

Maybe that is why Nvidia is being so generous with the 5070 price?

Math is wrong. That's 6144 x 4 = 24K cores when DLSS4 is in full swing. Must I do everything around here?

 

[igor_kavinski]

What's a 4070 Super?
It's going to be twice as fast* as a 4070, and $50 cheaper. You'd be crazy to not buy one.

Nuh uh. Don't twist Honorable Mr. Jensen's words. It's 4090 performance for $549. Because he said so.

 

[moonbogg]

I saw the Asus pricing at $2800. People keep paying, so it does make sense to charge what people are willing to pay. I saw some commentors with 4090's now saying that $2800 was too high for them, but that just means they have finally hit their personal limits while others will continue to pay up. So, next gen the top Asus card should be around $3500 maybe. Not saying I like it, but it does make sense and I hope to see this trend continue. It makes me giddy and fills my heart with joy to see all the painfully stupid fools who keep burning money finally getting what they deserve: being priced out of their flagship product. As long as they could afford it, then didn't care. They paid the prices and talked their trash to those who could not afford it and in doing so, they destroyed the GPU market for everyone. Have fun with that $2800 you idiots, next time it'll be $2800 + 30%.

 

[Timmah!]

TBF, the leaked 5080 prices aren't that far off from AIB 4080 Supers.

In Europe, they are.
I can see bunch of 4080S for 1200 tops at local retailer. Which is inline with the 999 + VAT pricetag. 5080s, from the prices that were leaked, are seemingly in the 1500~1800 range. Now those prices might be placeholders, but i would not hold a breath thats the case. Each company will have then no doubt one MSRP model...how available, there is another question.

https://preview.redd.it/real-prices-for-rtx-5090-and-rtx-5080-in-spain-21-vat-v0-ouypg13zzede1.png?width=620&format=png&auto=webp&s=5ff69d2c96b43df279dc01b2e80b7e5b489fc27b

 

[Elfear]

I saw the Asus pricing at $2800. People keep paying, so it does make sense to charge what people are willing to pay. I saw some commentors with 4090's now saying that $2800 was too high for them, but that just means they have finally hit their personal limits while others will continue to pay up. So, next gen the top Asus card should be around $3500 maybe. Not saying I like it, but it does make sense and I hope to see this trend continue. It makes me giddy and fills my heart with joy to see all the painfully stupid fools who keep burning money finally getting what they deserve: being priced out of their flagship product. As long as they could afford it, then didn't care. They paid the prices and talked their trash to those who could not afford it and in doing so, they destroyed the GPU market for everyone. Have fun with that $2800 you idiots, next time it'll be $2800 + 30%.

I saw that too and about gagged. $2000 is already sky high but adding another 40% to that for a fancy cooler is ludicrous. 🙄

 

[Thunder 57]

What's a 4070 Super?
It's going to be twice as fast* as a 4070, and $50 cheaper. You'd be crazy to not buy one.

And the vanilla 4070 has 5888 CUDA cores vs 6144. Not inspiring a lot of confidence and definitely not twice as fast. But Jensen himself did say it would be at 4090 levels so we'll see /s. I'd guess 15% faster based on nothing but my gut. Either it would be DoA for me because of VRAM. The 5070 Ti looks to be the only other interesting card thus far.

 

[Mopetar]

Definitely NOT a step forward for RT.

It gets above that magical 60 FPS line for some titles, so even if it's not an amazing increase in terms of raw percentage, I think it works psychologically for consumers that are aiming at a particular experience.

I still don't understand what's the point of the 512-bit bus here. Is this card really much worse with a 384-bit bus? maybe one could downclock the memory to see how much worse the performance is?

NVidia made this GPU to sell cards to the professional and data center markets and the bigger bus is needed there. Gamers just get it for free even if it doesn't seem to be doing much. I find it rather surprising that the significant bandwidth increase isn't doing more.

I'd be more grateful if nVidia gave up some of that margin they're enjoying. In the second quarter of fiscal 2025, Nvidia's gross margin was 75%, and that's up from 50% in 2012.

That's pretty much driven by the AI boom more than anything. NVidia would make a lot more money selling 5090's to other markets that will pay $$$$$ for a GPU like that.

So, where's the $1500 24GB 384bit part that replaces the 4090 at 450W outright? Seems doable and there is such a gulf from 5080 to the 5090 using the 4080S as a stand in.

AMD has nothing to compete in that space right now so Nvidia won't bin their dies to create a product to fill it. Why sell their die for $1,500 when they can get $2,000 for it. Don't want to spend that much? Too bad, buy a 5080 instead then. It's not like you have alternative options.

Otherwise every 1,000 of those $1,500 is half a million dollars in revenue they aren't getting. That's clearly pocket change to Nvidia, but you don't get to be a multi-billion dollar company like they are without sweating those details. They're not too different from Apple in that regard.

 

[igor_kavinski]

I find it rather surprising that the significant bandwidth increase isn't doing more.

I don't want to defend Nvidia but it's possible that this is a forward looking architecture that will shine brightly once work graphs and neural rendering become commonplace. We'll know more once CnC publish their 5090 microbenchmarking article.

 

[Josh128]

Very curious that 5090 performs almost 9% faster on average in 4K Max RT when paired with a 13900K than with a 9800X3D in Tom's Hardware review. Thats quite a large margin in the most relevant settings to a 5090. Has it been tested in any other review?