NVIDIA Refreshing GeForce RTX 30 Series on TSMC’s 7 nm Process?

[Tsing]

Image: NVIDIA



We may not have to wait until NVIDIA’s next architecture (i.e., “Hopper”) for GeForce RTX graphics cards built on a 7 nm process. According to a DIGITIMES article pertaining to the incredible demand for TSMC’s services, green team could be releasing refreshed GeForce RTX 30 Series GPUs that leverage the semiconductor manufacturing leader’s 7 nm node.



“…rumors say Nvidia’s previously launched annual masterpiece RTX 30 Series uses Samsung Electronics’ heavily discounted 8nm process, but will switch to TSMC’s 7nm process in 2021,” according to a translation provided by RetiredEngineer (chiakokhua)...

Continue reading...

 

[Grimlakin]

2021... the year Nvidia announces newer SMALLER GPU's and everyone cheers!

 

[Armenius]

No, it's not believable that the 3070, 3080 and 3090 would be changed to 7nm so soon. It isn't unprecedented for a process change mid-generation, but given the R&D and testing costs as they are right now I find it hard to believe. Quite a few 9-series cards were migrated to 55nm from 65nm, and some 200-series started life on 65nm and later moved to 55nm. The economics are quite different now than they were back then, however. What is believable is that the 3050 and 3060 would be on 7nm.

 

[Dan_D]

I'm having a hard time believing this myself.

 

[LazyGamer]

Given how much technology Nvidia tends to iterate each generation, a mostly-direct port of Ampere to 7nm could make sense.

I'm more skeptical of there being any real availability for them though. TSMCs 7nm node is seeing high demand from many industries, and while Nvidia is certainly a heavyweight and longtime customer, they seemed to be working much closer with Samsung this generation.

What is believable is that the 3050 and 3060 would be on 7nm.

If there's 7nm capacity to be had at TSMC, this is pretty believable. Especially if these are RTX parts and not hybrids like the 16x0's were. These parts see significant volume in power-limited scenarios and are valued as much for their gaming performance as they are for their utility as accelerators for other applications.

 

[kcthebrewer]

I was under the understanding that process nodes were incompatible and would require a complete(ish) redesign to be moved over.
I think this is more likely for mobile 30 series.

 

[LazyGamer]

I was under the understanding that process nodes were incompatible and would require a complete(ish) redesign to be moved over.
I think this is more likely for mobile 30 series.

It'll be a full-on engineering project to get it done, but that's still small potatoes next to a completely new architecture.

 

[Stoly]

A couple of things to consider.

Moving to a smaller process is no small feat, even more when its a different manufacturer.

Nvidia contracts manufacturing in advance, probably its covered for a year at least. Not easy to cancel

On the first note, nvidia has already designed A100 for TSMC 7nm, so moving A102 to TSMC may not be that hard.

On the 2nd note, while it may not be possible to move the RTX 3090/3080/3070 to TSMC for manufacturing contract reasons. They could have RTX3060 or even Ti versions of current cards.

It begs the question as why didn't nvidia go 7nm in the first place, even on Samsung.

 

[Brian_B]

It begs the question as why didn't nvidia go 7nm in the first place,

Well, at least for TSMC, I have a feeling it was just available capacity. Apple and AMD have a very large chunk of their capacity booked up.

On Samsung's front, 8nm vs 7nm at Samsung, that could very well be differences in the nodes - not all 7nm are the same, some are optimized for low power, some are optimized for high performance, some are optimized for other things, I'm sure, and nVidia decided the 8nm was the best bet for Ampere at the time.

 

[Stoly]

Well, at least for TSMC, I have a feeling it was just available capacity. Apple and AMD have a very large chunk of their capacity booked up.

On Samsung's front, 8nm vs 7nm at Samsung, that could very well be differences in the nodes - not all 7nm are the same, some are optimized for low power, some are optimized for high performance, some are optimized for other things, I'm sure, and nVidia decided the 8nm was the best bet for Ampere at the time.

AFAIK the 8nm process was "tailored" for nvidia. It wasn't designed for such large chips, but it was easier to adapt than 7nm .

Kind of reminds me of Turing 12nm which could not be made on 10nm.
BTW Turing never made the transition to 10nm despite rumors before the super models were announced.

 

[UFO]

It begs the question as why didn't nvidia go 7nm in the first place, even on Samsung.

It's political, and not a result of any "bad decisions" on Nvidia's part.
US lawmakers were pressuring TSMC to make chips in America as recent as last Janruary,
and allthough rumers were that TSMC foundry"Fed up" with Nvidia's chip designers,
the Samsung order was to make certain Nvidia had a product to deliver;
and a last minute 7nm order obviously isn't possible

 

[Auer]