Rumor | NVIDIA GeForce RTX 3090 Ampere GPU to sport 4,992 CUDA cores and 12 GB of 18 Gbps VRAM

[P4-630]

Apparently based on the GA102, the GeForce RTX 3090 may feature 4,992 CUDA cores and 12 GB of GDDR6 VRAM. Running at 18 Gbps, the RTX 3090 would be one of NVIDIA's flagship Ampere consumer GPUs.

Treat this with some scepticism for the time being. While most of the details contained herein are from @kopite7kimi, a reliable leaker of information about upcoming NVIDIA hardware, we have been unable to verify them.

In line with previous rumours about the Ampere architecture, @kopite7kimi claims that NVIDIA is developing a GPU codenamed the GA102. Referred to as the RTX 3090 for simplicities sake, other information about the GA102 suggests that it will find its way into multiple GPUs.

At any rate, @kopite7kimi claims that the RTX 3090 would be the "full-fat" GA102. However, they do state that the card will feature 4,992 CUDA cores, a way of previous estimates. Apparently, the card will feature 39 Texture Processing Clusters (TPCs) too, which would be three more clusters than the full TU102 configuration has.

Referred to as "THE ULTIMATE GEFORCE" by @Ragdoll_Kitties, the GA102 is said to include 12 GB of GDDR6 VRAM running at 18 Gbps. Apparently due to arrive by "August or later", this ties in with previous rumours about consumer Ampere GPUs slipping to Q3 or Q4 2020.

However, Ampere is also said to be based on a 10 nm process. Hence, it is no longer necessarily clear what performance levels to expect from the upcoming architecture. If true, then AMD could have a big advantage over NVIDIA headed into the next-generation mainstream desktop GPUs.

NVIDIA GeForce RTX 3090 Ampere GPU to sport 4,992 CUDA cores and 12 GB of 18 Gbps VRAM

Apparently based on the GA102, the GeForce RTX 3090 may feature 4,992 CUDA cores and 12 GB of GDDR6 VRAM. Running at 18 Gbps, the RTX 3090 would be one of NVIDIA's flagship Ampere consumer GPUs.

www.notebookcheck.net

 

[dgianstefani]

Nope. Nvidia won't stick with only 12gb. I also have strong doubts about Samsung 10nm when RDNA2 is on 7nm+.

 

[P4-630]

I also have strong doubts about Samsung 10nm when RDNA2 is on 7nm+.

Rumor: Next gen Nvidia RTX chip will be based on Samsung 10nm node

All new cards will be RTX. SLi will be available to GA102 ONLY. https://nl.hardware.info/nieuws/71288/gerucht-alle-ampere-kaarten-krijgen-rtx-en-worden-gemaakt-op-samsungs-10nm-node

www.techpowerup.com

 

[notb]

I also have strong doubts about Samsung 10nm when RDNA2 is on 7nm+.

Why bother? Especially for a high-end desktop part?
GPUs are efficient anyway. They won't look bad next to Radeons.

More frugal 7nm would be more useful in mobile chips.

 

[ppn]

10nm could be the same density if not better than the 7nm DUV 41 mTr/mm2 that AMD plans for the RDNA2.
14nm 1050Ti made by samsung is 25 mTr, second gen 10nm must be at least 35-45 mTr.

Nope. Nvidia won't stick with only 12gb. I also have strong doubts about Samsung 10nm when RDNA2 is on 7nm+.

Do you honestly believe that. Just watch them, they will stick with 12GB for many years. no problem

 

[dgianstefani]

For their top tier GPU? No. 16GB, 2080 replacement might get 12GB. 2060 replacement I expect 8GB.

 

[ppn]

We are stuck with 8 and 12 until the 50-series, 24GB is Quadro territory. When RX 5900 has 12GB why bother.
full fat RTX 3080 3584 Cuda can work with 8GB just like a 20% fatser 2080 Super.

 

[Vya Domus]

10nm could be the same density if not better than the 7nm DUV 41 mTr/mm2 that AMD plans for the RDNA2.

It could have better density but that wont matter much, it's all about performance, clocks and such. Historically, Samsung's nodes haven't been about that at all.

 

[Flanker]

hmm 4992 CUDA cores as opposed to 4352 cuda cores, 15% increase.
Oh btw, I though x90 has always been for dual GPU cards?

 

[Vayra86]

hmm 4992 CUDA cores as opposed to 4352 cuda cores, 15% increase.
Oh btw, I though x90 has always been for dual GPU cards?

Maybe its the new moniker for an SLI capable card, as its the only one and also top end.

Anyway... this is all a little too much dust & smoke for me to go on as of yet. Needs more substance, though it doesn't look impossible.

 

[Darmok N Jalad]

Total memory is going to depend on bus width. 12GB would suggest a 384bit bus, consistent with previous ultra cards (Titan). If so, the next step is 24GB, not 16GB. 16GB would require a 416 bus—far more complex. It’s all speculation, sure, but it’s based on history and what nVidia currently offers. It would make sense for them to chase faster RAM vs adding more bus width.

 

[ShurikN]

Total memory is going to depend on bus width. 12GB would suggest a 384bit bus, consistent with previous ultra cards (Titan). If so, the next step is 24GB, not 16GB. 16GB would require a 416 bus—far more complex. It’s all speculation, sure, but it’s based on history and what nVidia currently offers. It would make sense for them to chase faster RAM vs adding more bus width.

16GB would require a 256 or 512-bit bus, depending on the ammount of memory chips. 416 is 13GB. 32bit per chip.

 

[Aquinus]

That die is going to be absolutely massive. If this is true, I predict yields will be garbage.

 

[Darmok N Jalad]

16GB would require a 256 or 512-bit bus, depending on the ammount of memory chips. 416 is 13GB. 32bit per chip.

Yeah, my bit-math is off. Me =

Still, I can’t see them dropping such a big chip down to 256bit, or going all the way to 512bit.

 

[ShurikN]

If this is true, I predict yields will be garbage.

And pricing atrocious

Still, I can’t see them dropping such a big chip down to 256bit, or going all the way to 512bit.

Well 12GB worked for them so far. 384bit with 18Gbps chips is a lot of bandwidth.

 

[Assimilator]

Quite plausible that NVIDIA is repurposing the -90 suffix from their dead dual-GPU card line to mean "supports SLI". As such, it is possible that it will sport 16GB VRAM (as a high-end card, particularly one geared towards dual-GPU, is likely to need the largest framebuffer it can get).

I don't buy the Samsung 10nm rumour - NVIDIA has used TSMC for all its high-end GPUs practically forever, I doubt they would be willing to take a chance with a relative minnow like Samsung for such a large and complex die. For the smaller GPUs it's quite likely, but that will be further down the line.

 

[Valantar]

Hm, an actually believable rumor. Interesting. Memory bandwidth is equally interesting though as noted by many above, and IMO it's highly unlikely that Nvidia will go for a 512-bit bus (large, expensive, power hungry), so 12GB sounds plausible to me. Then again we don't actually need an increase here - while there are plenty of games hitting 8GB or more VRAM usage today, most of them don't actually need all they're using - clearly seen by cards with <8GB not falling off a cliff in terms of performance at the same settings. And with PCIe 4.0 and NVMe SSDs becoming the new norm for game development, streaming in assets when they are actually needed will be less of a liability. Not to mention that both GPU vendors and several API makers are talking about implementing better methods for determining which assets are actually used and which aren't, lowering VRAM needs. The main question becomes if a 384-bit bus is capable of feeding a bigger chip, but with the increase in memory speed roughly matching the increase in cores that sounds likely. Looking forward to something more concrete and official though.