Discussion Nvidia Blackwell in Q1-2025

[MrTeal]

Everything I read, said the opposite. It more difficult to do the DC-DC conversion on 48V to ~1V range of ICs, than it is to go from 12V.

For DC Downconverter spec sheets I have seen, the larger the delta between Vin and Vout, the lower the efficiency.

So if you ran 48V DC to a GPU, you would end up with much more waste heat. That would make GPU power problem worse, not better.

Much more is a bit of a stretch. They might be a bit less efficient, but a purpose built 48V-> 1V convertor will be closer in efficiency to a 12V-1V convertor than a wide input range convertor operating at the top end of the range. I'd be surprised if it was even 1 percentage point less efficient, considering you can buy complete 48V-12V solutions that are 98% efficient. That's just the VRM on the GPU though, the flipside is lower losses in cabling and connectors.

The transition problems would be a pain though. Not just needing a new power supply, but there's a huge range of components used in the VRMs themselves tailored to 12V->~1V for both CPU and GPU VRMs with maximum efficiency. It'd probably be more likely if direct 48V conversion takes off in the datacenter and that whole supply chain builds up first.

 

[Heartbreaker]

Much more is a bit of a stretch. They might be a bit less efficient, but a purpose built 48V-> 1V convertor will be closer in efficiency to a 12V-1V convertor than a wide input range convertor operating at the top end of the range. I'd be surprised if it was even 1 percentage point less efficient, considering you can buy complete 48V-12V solutions that are 98% efficient. That's just the VRM on the GPU though, the flipside is lower losses in cabling and connectors.

The transition problems would be a pain though. Not just needing a new power supply, but there's a huge range of components used in the VRMs themselves tailored to 12V->~1V for both CPU and GPU VRMs with maximum efficiency. It'd probably be more likely if direct 48V conversion takes off in the datacenter and that whole supply chain builds up first.

The cables in a PC are inches long though. This isn't data center where they might have miles of DC voltage distribution cable. The impetus to take the load of cables inside a PC isn't pressing.

Most of the Data center 48V I have seen is just Stepping down to 12V first before it hits the MB, and then doing normal 12v-1v conversions on the MB. Which would be utterly pointless inside a PC.

Eventually we might see 48V-1V inside a PC, but that will likely take 10+ years. There simply isn't a pressing need for 48V in PC, and no real benefit even. It may come along years after it becomes so common in Data Center that it become easier and cheaper to switch to 48V...

 

[jpiniero]

https://videocardz.com/press-release/msi-announces-geforce-rtx-50-gaming-laptops-confirms-geforce-rtx-5090-24gb-gddr7-and-dlss4

5090 Laptop supporting 24 GB confirmed. MSI has a bunch of models with the 5070 Ti/5080L as well and also has AMD CPU options too.

 

[Heartbreaker]

https://videocardz.com/press-release/msi-announces-geforce-rtx-50-gaming-laptops-confirms-geforce-rtx-5090-24gb-gddr7-and-dlss4

5090 Laptop supporting 24 GB confirmed. MSI has a bunch of models with the 5070 Ti/5080L as well and also has AMD CPU options too.

Interesting. Maybe 3GB VRAM is arriving soon, because it's extremely unlikely that they are using a cut down GB202 in a laptop. So that's most likely a GB203 256bit bus, with 3GB VRAM chips.

 

[Timmah!]

Less than 24 hours left till CES and still ZERO leaks on 5080/5090 perf???


Perhaps Nvidia could bundle new PSU with 2 grand cards? If that makes them perform better (even if it's just lower power) than what's the problem? Get top models on board of thise - whoever is buying 5090 will be ok with it, as long as it brings real benefits though

I want to see Octane render performance asap. 4090 at 16k cores was almost 2x as fast as 3090 at 10k. This one is gonna have 21k, so the core increase gonna drop from 60 percent to just about 30. Wonder if they can make up for it with architectural changes, or it simply wont be that much more performant. If it wont, i hope it will be reflected in price, that wont go up (despite the rumors).

 

[Win2012R2]

I hope they got them clocked at least 3.3 Ghz, much higher memory bandwidth gives hope that perf gains will be very good at 4k at least

 

[Saylick]

I know price "leaks" are very much last minute, but since it's quite literally hours before CES, I figure what's the harm in sharing this:

https://twitter.com/x/status/1876182767828107402

 

[Timmah!]

Hopefully its the same case for 5090 as well, in comparison to 4090, even if the the perf difference between 5080 and 90 gonna be greater.

 

[gdansk]

Hopefully its the same case for 5090 as well, in comparison to 4090, even if the the perf difference between 5080 and 90 gonna be greater.

Die too big. $2000 minimum.

 

[Saylick]

Hopefully its the same case for 5090 as well, in comparison to 4090, even if the the perf difference between 5080 and 90 gonna be greater.

I doubt it, especially since GB202 is literally double GB203.

 

[Timmah!]

Die too big. $2000 minimum.

Still smaller than Turing, Which was like 1100 - 6 years ago. Not 20.
And its on the same-ish node as Lovelace, thats why its bigger in the first place. One would expect with more time the processes gonna mature and become less expensive. Or allow for bigger chips at the same price.

 

[Saylick]

Still smaller than Turing, Which was like 1100 - 6 years ago. Not 20.
And its on the same-ish node as Lovelace, thats why its bigger in the first place. One would expect with more time the processes gonna mature and become less expensive. Or allow for bigger chips at the same price.

That's the crux. The cost per transistor has not decreased since 16nm like it historically did. Factor inflation into the equation and Jenson's love for fat margins and it's really easy to see $2000 MSRP. Then you have people who have no impulse control over their gaming purchases and the street price for the 5090 easily becomes $2500.

 

[adroc_thurston]

That's the crux. The cost per transistor has not decreased since 16nm like it historically did.

It actually increased which is lmao.

 

[Josh128]

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

 

[Heartbreaker]

$1999 minimum for 5090. 4090's have been selling close to that. With significantly larger die, and 50% more VRAM (likely more expensive VRAM) the increase is warranted.

Putting myself in the place of someone pricing it, I'd start it at $2199, you can always lower it later if that's too high, but I suspect they would still sell out at that price.

 

[Grooveriding]

When does the announcement happen?

 

[jpiniero]

When does the announcement happen?

The keynote starts in 2 hours.

 

[gdansk]

Putting myself in the place of someone pricing it, I'd start it at $2199, you can always lower it later if that's too high, but I suspect they would still sell out at that price

The point is the 5090 is a value and you're ruining it. It has to look good compared to the 5080. That's the entire point of the spoiler part being $1200. It works worse if the price is higher on the halo part. "The more you buy..." isn't just a line that's the actual strategy.

 

[Heartbreaker]

It actually increased which is lmao.

Transistor prices haven't increased. There is the occasional short term blip up in transistor prices but the long term trend is still down. Just down at a MUCH slower rate, than the good old days.

TU102 is in the ballpark of 20 Billion transistors, GB102 is likely around 100 Billion, 5X as many transistors.

I'll agree that its a significantly more expensive die, but it's not 5X as expensive.

 

[Josh128]

The point is the 5090 is a value and you're ruining it. It has to look good compared to the 5080. That's the entire point of the spoiler part being $1200. It works worse if the price is higher on the halo part. "The more you buy..." isn't just a line that's the actual strategy.

60% more perf for only 50% more price is a frickin' steal!!

 

[gdansk]

60% more perf for only 50% more price is a frickin' steal!!

View attachment 114369

You're getting 50% more for only 25% more. Look at that value. It is hard to argue with (provided the 5080 is priced poorly).

They really have it worked out.

 

[Josh128]

You're getting 50% more for only 25% more. Look at that value. It is hard to argue with (provided the 5080 is priced poorly).

They really have it worked out.

Remind me again the prices to which you are referring??

 

[gdansk]

Remind me again the prices to which you are referring??

$2000 and $1600 (last seen in 2024) and $1200. They're good prices. Don't go to Lisa's, you won't get these prices.

 

[Saylick]

Since we’re 30 min from the presentation, why don’t we drop in our final predictions for pricing here. Closest one wins, and may Lord Jenson have mercy on our wallets!

 

[gdansk]

Well the 5080 is $200 cheaper than I expected. It's fair then excluding parametric yield. Half the die, half the price.