News RTX 5090 may be surprisingly svelte — twin-slot, twin-fan model on the way, says leaker

[Amdlova]

Low end this time will start at 400 us
Middle 700 900 us
High 1000 2000
Unbelievable 4000

 

[spongiemaster]

Nvidia would need to pull a rabbit out of its hat to get significant performance improvements while sticking with a revised version of the existing node. As noted above, if consumer Blackwell GPUs still use 4NP, the only way to get more performance is with bigger chips and more power.

They've done it before with Maxwell following Kepler on the same node. 980Ti was slightly larger than 780Ti. Much faster while using slightly less power.

 

[SyCoREAPER]

Let's see if they mess up with generation as bad as the 40. Artifical shortages, low VRAM, smaller bus bandwidth, the yet to be determined safe connector**, price.

**F-Me I wasn't as uncertain or bother by Y2K as I am this damn connector. Use to let my computer run while doing am errand, not anymore...

If there's a generation to skip or be a late adopter for, it's this one. New cooling tech in supposedly substantial performances uplifts? I don't want to be a beta tester for that.

 

[YSCCC]

Somehow in my mind a default water cooling solution is coming... a twin slot, water block with a 360mm AIO

 

[Metal Messiah.]

Some update:

RTX 5090 will reportedly sport 28GB GDDR7 and 448-bit bus ! ​

So it will not use all the available 512-bit memory bus of the GB202 die, and only 14 memory modules would be used out of 16. So that should give 28GB VRAM. More than enough for any gaming GPU.

They could reserve the remaining VRAM for later Ti variants, if need be, or used it on a ProViz RTX GPU.

https://www.chiphell.com/forum.php?mod=viewthread&tid=2606354&page=4#pid55038013

448-bit Memory bus akin to GTX 260 era more like ! lol.

ALSO, assuming Nvidia would utilize 28 Gbps GDDR7 memory modules initially, we could see a total bandwidth of up to 1568 GB/s on the 5090. *speculation*

 

[Metal Messiah.]

Off Topic.

Is Geforce GTX G-Assist making a comeback after 7 years ? Latest tweet from Nvidia. Weird.

https://twitter.com/x/status/1795966310532481174

View: https://x.com/NVIDIAGeForce/status/1795966310532481174

 

[TechyIT223]

If these cards again support the controversial 12+4 pin vhpwr Gen 5 connector, then I would be surely skipping them.

 

[Metal Messiah.]

Svelte means slender. It's used to describe people, not things, and it implies a certain elegance.

Couldn't agree more ! The article has a horrible title/heading for a tech site by the way.

The same was with another previous story with the "Golden Pig" name attached to it !

 

[TechyIT223]

I hope they got rid of this piece of crap power connector that is a big engineering failure.

What's the point of miniaturizing a connector when the GPUs are so huge anyway... ?

Gotta agree. I hate to see the same 12VHPWR cable used again in these cards. I don't trust these cables despite the new revision released recently.

I bet we will see more melting power cables and GPUs with Blackwell generation as well 😂

16 pin cable gives me a nightmare

 

[JarredWaltonGPU]

They've done it before with Maxwell following Kepler on the same node. 980Ti was slightly larger than 780Ti. Much faster while using slightly less power.

That was a very different era of GPUs, though. I'm not saying it can't be done at all, but Kepler was clearly a non-optimal GPU configuration and Maxwell improved things a lot (and dumped some stuff that was eating power as well, IIRC). Ada Lovelace is, by all appearances, quite optimized. To get more performance without increasing power use (or even reducing it) would imply that Ada left a lot of room on the table for future improvements. I simply don't believe that's been true of the past several generations of GPUs from Nvidia.

 

[Joseph_138]

I don't think 3nm is ready for prime time yet. I heard it has major scaling issues, so performance may not be as good as we think it will be. The process needs perfecting. RTX 50 may not be such a good bargain. It might be a good idea to sit this one out, for most people.

 

[Metal Messiah.]

Nope. The Blackwell GB202 silicon should be built on the same TSMC 4N foundry node used on the B100/200. TSMC 4NP is just a higher performing version of the 4N node used for the GH100 GPU.

TSMC still considers the 4N as a derivative of the 5 nm EUV node. Though, we expect at least a 30% increase in transistor density.

NVIDIA is not using a TSMC 3nm-class node for Blackwell. Despite sticking to a 4nm-class node, NVIDIA was able to squeeze more transistors into a single die though.

 

[TechyIT223]

Man these TSMC nodes are a bit confusing for me to grasp. Especially the naming part should be more simple to understand