Discussion Nvidia Blackwell in Q1-2025

[Grooveriding]

Feel this really sums it up. Doesn’t feel worthy of a new generation moniker.

 

[SolidQ]

1 year ago, since 16:58. Magic ball worked well

MSI 5090

 

[Heartbreaker]

1 year ago, since 16:58. Magic ball worked well

You mean this stopped clock, among the multitudes of stopped clocks, and their multitudes of windbag claims stumbled into getting something vaguely correct. About as correct as your daily horoscope.

Whatever it is, I have already block tubed that channel as well, so it's just a black empty box for me.

Videocardz reports all the leaks/rumors, that I can read in less than a minute, so no need to waste half an hour listening to windbags...

About the only NVidia Leaker, I take seriously is Koptie7Kimi, who has very high accuracy, and is either an actual insider, or knows an insider very well, and I just read their leaks on Videocardz.

 

[jpiniero]

MSI GeForce RTX 5090 Suprim Liquid SOC Review

Thanks to fully its integrated watercooling solution, the MSI GeForce RTX 5090 Suprim Liquid SOC can achieve excellent temperatures and lower noise levels than the Founders Edition. It's also highly overclocked, delivering additional performance out of the box.

www.techpowerup.com

5090 Liquid only gets another 5%.

 

[Ranulf]

1 year ago, since 16:58. Magic ball worked well

MSI 5090

There are just so many memes one can use here. I've gone thru like 3-4 in just 2 minutes.

761 watts? Doc Brown says you're gonna need a bolt of lightning to power that thing. That or an APU starter to power up that card to hit the 747watts in Cyberpunk.

 

[maddie]

Videocardz reports all the leaks/rumors, that I can read in less than a minute, so no need to waste half an hour listening to windbags...

Well said. Time, the only resource fixed in quantity.

 

[MrTeal]

MSI GeForce RTX 5090 Suprim Liquid SOC Review

Thanks to fully its integrated watercooling solution, the MSI GeForce RTX 5090 Suprim Liquid SOC can achieve excellent temperatures and lower noise levels than the Founders Edition. It's also highly overclocked, delivering additional performance out of the box.

www.techpowerup.com

5090 Liquid only gets another 5%.

There's only another 25W to hit the hard power cap, so that's not unexpected. You'd definitely need to volt mod it to run more power into it.

 

[Heartbreaker]

There's only another 25W to hit the hard power cap, so that's not unexpected. You'd definitely need to volt mod it to run more power into it.

Voltage modding is generally bad idea, but on a 600W card it's asking for disaster.

 

[MrTeal]

Voltage modding is generally bad idea, but on a 600W card it's asking for disaster.

Why do you say that?

 

[Timmah!]

1 year ago, since 16:58. Magic ball worked well

Nothing magical about this. We could have done this easily, since the number of cores was already known at that point, and we could presume the clocks are not going to go another gigahertz up, as they went when moving from Ampere to Lovelace.
I mean 3090 to 4090 was 60 percent more cores and higher clocks. 4090 to 5090 is 30 percent more clocks and similar (actually lower) clocks. I mean we all here could have guessed here the perf increase is gonna end up as he predicted, under those circumtances. It would be actually surprising, if the perf increase was significantly more, as that would require some magical architectural improvements.

 

[Meteor Late]

Nothing magical about this. We could have done this easily, since the number of cores was already known at that point, and we could presume the clocks are not going to go another gigahertz up, as they went when moving from Ampere to Lovelace.
I mean 3090 to 4090 was 60 percent more cores and higher clocks. 4090 to 5090 is 30 percent more clocks and similar (actually lower) clocks. I mean we all here could have guessed here the perf increase is gonna end up as he predicted, under those circumtances. It would be actually surprising, if the perf increase was significantly more, as that would require some magical architectural improvements.

Not necessarily, because each core itself can always be more powerful, like in Turing.
That being said, it's also probably fair to consider how small 1080 ti die was all things considered, so it made sense that 2080 ti was able to have more powerful cores.
It looks like the days of great architectural boosts are over, since we are in a situation where both die size jumps are not possible anymore due to how big dies already are nowadays, and process node jumps will be smaller and smaller.
I'm comparing the number of transistors between 3090 and 4090 and I cannot believe what I'm seeing, helps put things into perspective, yeah of course 4090 was impressive, the difference in transistor count is out of this world, and the difference in performance between Samsung 8nm and TSMC 5nm is also HUGE.

 

[Thunder 57]

Anyone else think Tom's was a bit... generous giving this thing 4.5/5 stars?

 

[Elfear]

Anyone else think Tom's was a bit... generous giving this thing 4/5/5 stars?

I thought TPU's review(s) were rather glowing too.

 

[Thunder 57]

I thought TPU's review(s) were rather glowing too.

Haven't seen that one but that does not surprise me. Tom's says some things just don't work, the drivers are rather early, the performance increase isn't as much as past generations and the power usage went up a lot. But 4.5/5! Thankfully they got some crap about it in the comments section. Jarred's response was basically "well the review isn't quite finished yet". Then why give it a rating?

 

[adroc_thurston]

It looks like the days of great architectural boosts are over

W r o n g.

 

[Saylick]

Haven't seen that one but that does not surprise me. Tom's says some things just don't work, the drivers are rather early, the performance increase isn't as much as past generations and the power usage went up a lot. But 4.5/5! Thankfully they got some crap about it in the comments section. Jarred's response was basically "well the review isn't quite finished yet". Then why give it a rating?

Gotta please their handlers at Nvidia if they want to keep receiving Day 1 review samples moving forward.

 

[amenx]

W r o n g.

Curious.. what do you think the perf uplift would have been if blackwell was on 3nm?

 

[CP5670]

MSI GeForce RTX 5090 Suprim Liquid SOC Review

Thanks to fully its integrated watercooling solution, the MSI GeForce RTX 5090 Suprim Liquid SOC can achieve excellent temperatures and lower noise levels than the Founders Edition. It's also highly overclocked, delivering additional performance out of the box.

www.techpowerup.com

5090 Liquid only gets another 5%.

I would prefer this card for the low noise and the heat going straight out of the case, not for the performance. But at $2500 it's hard to justify. This is a much bigger markup than the $200 that the 4090 version had.

 

[igor_kavinski]

W r o n g.

Folks, this is most likely a hint towards UDNA performance

 

[adroc_thurston]

Curious.. what do you think the perf uplift would have been if blackwell was on 3nm?

Node doesn't change the SM change being non-existent, upon of which they also won no fmax.
The uarch is the problem, not the node.

Folks, this is most likely a hint towards UDNA performance

not just RDNA5.
NV really oughta pick the slack, cuz sm100 was their Next Big Thing.

 

[SolidQ]

for RTX 5090?

 

[Timmah!]

Equaling the number of FP and INT cores was the meh idea. If they used that die-space to increase the number of FP cores instead, they could have at least better performance in stuff like blender. Which would suit me

 

[igor_kavinski]

Equaling the number of FP and INT cores was the meh idea. If they used that die-space to increase the number of FP cores instead, they could have at least better performance in stuff like blender. Which would suit me

Look on the bright side. You save money, power bill, getting rid of all that heat. And maybe now you could pick up a third 4090 once they start thrashing the used market.

 

[Timmah!]

Look on the bright side. You save money, power bill, getting rid of all that heat. And maybe now you could pick up a third 4090 once they start thrashing the used market.

Yeah, and i ducktape it outside of the case :-D
But yeah, if the new product is as not good enough as we hoped, not feeling as urged to get it and save the money in the process is always a silver-line
I saw the tweet by OTOYs CEO yesterday, hyping 5090 as the first GPU that MIGHT cross 2000 octanebench points threshold (with OC) - 4090 gets cca 1250 for comparison - so that would be actually decent speed-up. I mean, given blender and vray are only around 30 percent faster.

 

[jpiniero]

Node doesn't change the SM change being non-existent, upon of which they also won no fmax.
The uarch is the problem, not the node.

It would have helped if they had used N4P I think.