News Leak claims RTX 5090 has 600W TGP, RTX 5080 hits 400W — up to 21,760 cores, 32GB VRAM, 512-bit bus

[Admin]

Renowned hardware leaker kopite7kimi posted potential specs for the RTX 5090 and 5080, seemingly confirming many of the previous rumors. The RTX 5090 will supposedly have up to 21,760 CUDA cores, a 512-bit memory interface, and a 600W TGP; the RTX 5080 will basically be half those values, but with a 400W TGP.

Leak claims RTX 5090 has 600W TGP, RTX 5080 hits 400W — up to 21,760 cores, 32GB VRAM, 512-bit bus : Read more

 

[Notton]

I think it's the first option. 2x GB203 glued together, but clocked lower.
How else would a dual 400W turn into 600W?

Speculation
Did they choose 600W because they didn't want to add another 16pin 12VHPWR connector?
or does the CoWoS-L interconnect bottleneck and not benefit from more power into the chip?
or they are leaving headroom for an 800W 5090Ti?

I expect the 5090 to run cool, so long as your case can keep up with the total heat it dumps. That is a huge die and would be easy to cool with a vapor chamber.

 

[Giroro]

If it has 32GB of memory on a 512-bit bus then it's going to be an AI/Workstation card, and it will be priced accordingly.

 

[Gururu]

The line between enthusiast and professional is getting thinner.

 

[Elusive Ruse]

I sure hope the 5080 is not coming with the same 16GB of VRAM even if it’s GDDR7.

 

[Flayed]

They probably want to limit the vram on the 5080 so that the AI boffs have to buy the 5090

 

[newtechldtech]

I dont think the 5090 has two chips.. there is a missing 5080 ti to fill inbetween in the future .

 

[UnforcedERROR]

Keeping 16GB on the X080 and below parts is crazy, really. Why only give the TOTL that upgrade?

 

[thestryker]

The RTX 4090 has 16384 cores on a cut down AD102 (18432 maximum) and is 602mm^2 on TSMC N5. N4P (assuming they're using the same process for enterprise and client) isn't enough smaller to make up for the core count increase within the same die size. So to me that would mean the 5090 is either harvested GB100 or indeed a dual die solution. With the word about yield issues on early Blackwell (sounded like largely interconnect related) it's possible nvidia was able to pivot and get more usage of the die over for client.

**assuming this leak is accurate

 

[TeamRed2024]

24GB VRAM should be enough for anyone... well... for me anyway. I have seen anything even come close to maxxing it out so the fact the 5090 will have 32 isn't a dealmaker either way.

 

[valthuer]

Percentage wise, can somebody please explain to me the projected performance increase of 5090 over 4090?

 

[TeamRed2024]

Percentage wise, can somebody please explain to me the projected performance increase of 5090 over 4090?

If I were a betting man I would say not even close to the gap the 4090 has over everything else.

 

[valthuer]

24GB VRAM should be enough for anyone... well... for me anyway. I have seen anything even come close to maxxing it out so the fact the 5090 will have 32 isn't a dealmaker either way.

If you love your card, don’t try Star Wars: Outlaws. At 4K Outlaw RT settings, it can sometimes consume over 21 GBs of VRAM, even with DLSS at performance.

Sweet Jesus! 🤣

If I were a betting man I would say not even close to the gap the 4090 has over everything else.

That’s exactly what I’m afraid of: that it won’t be worth the hefty price.

 

[umeng2002_2]

The trend with all leaks is that the leaks about performance are about 15% too high and the prices about 15% too low.

 

[Flayed]

Percentage wise, can somebody please explain to me the projected performance increase of 5090 over 4090?

25% more cores plus faster memory, so more than 25% I guess.

 

[TeamRed2024]

25% more cores plus faster memory, so more than 25% I guess.

If it's the same $1599 MSRP as the 4090 the solution is simple... sell the 4090 for $1000 and then pay the difference.

 

[vanadiel007]

As long as it's not 1.21 gigawatts we should be ok.

 

[Mama Changa]

The RTX 4090 has 16384 cores on a cut down AD102 (18432 maximum) and is 602mm^2 on TSMC N5. N4P (assuming they're using the same process for enterprise and client) isn't enough smaller to make up for the core count increase within the same die size. So to me that would mean the 5090 is either harvested GB100 or indeed a dual die solution. With the word about yield issues on early Blackwell (sounded like largely interconnect related) it's possible nvidia was able to pivot and get more usage of the die over for client.

**assuming this leak is accurate

On the existing low NA lithography that N5 and N4 are made on the die size limit is over 800mm^2. Combine that with N4P's 11% higher density and Nvidia can easily do a single die, no need for dual dies at all. It's the next gen high-NA EUV that has a die size limit of just over 400mm^2 that will cause problems and why after Blackwell, Nvidia will have to go chiplet

 

[Conor Stewart]

I think it's the first option. 2x GB203 glued together, but clocked lower.
How else would a dual 400W turn into 600W?

Speculation
Did they choose 600W because they didn't want to add another 16pin 12VHPWR connector?
or does the CoWoS-L interconnect bottleneck and not benefit from more power into the chip?
or they are leaving headroom for an 800W 5090Ti?

I expect the 5090 to run cool, so long as your case can keep up with the total heat it dumps. That is a huge die and would be easy to cool with a vapor chamber.

Generally power and performance do not scale linearly, once you pass a threshold adding more power doesn't do much for performance, just look at the Ryzen 9000 series reviews, specifically stock vs PBO enabled, around double the power for maybe 10 % more performance.

It could be a case of with the 5080 they are pushing it as far as possible in terms of power because they have the headroom, given they have 675 W available (600 W from power connector plus 75 W from PCIe slot) so pushing to 400 W isn't too hard but it may not actually gain much performance versus if they capped it at say 300 W. Then with the 5090 they may just not be pushing it so hard and cap the whole thing at 600 W but are still able to get most of the total theoretical performance out of the chip.

For applications where you want high performance but low power you generally want to go for more cores but lower clocks and lower power limits to run those cores more efficiently since running the cores close to their maximum performance isn't very efficient.

 

[thestryker]

On the existing low NA lithography that N5 and N4 are made on the die size limit is over 800mm^2. Combine that with N4P's 11% higher density and Nvidia can easily do a single die, no need for dual dies at all.

I don't think you caught the point being made: if it's not die harvested and it's a single die that means they'd be making a client GPU around the same size as their flagship enterprise model. This makes zero financial sense at all given the disproportionate impact on yield and GPUs per wafer.

 

[Makaveli]

24GB VRAM should be enough for anyone... well... for me anyway. I have seen anything even come close to maxxing it out so the fact the 5090 will have 32 isn't a dealmaker either way.

all depends on what you do on the machine.

If you only play games then yes.

Anyone that runs LLM's / AI need as much VRAM as possible.