Nvidia is reportedly partnering with TSMC for 7nm Ampere production next year

DefaultD

Posts: 8   +13
Did TSMC have capacity for them in Q3/Q4, or had it already been promised to others?
I suspect TSMC's capacity on the different fabs will go something like this: Apple is on TSMC's 5 nm for their iPhone 12s (I read in a previous article, Apple snatched up all initial capacity on 5 nm). AMD will be switching to 5 nm next year with Zen 4 (and their next graphics cards?), which will then leave the TSMC 7nm fabs open for other customers (Nvidia). That's what I would guess based on previous articles and rollout dates I've read.

Nvidia will benefit from the mature manufacturing process while their designs are already well-optimized Remember, their 12 nm GPUs were still beating AMD's 7 nm GPUs last generation...we'll see how the new generations stack up with the AMD launch later this month (looking to see performance of RDNA2 on 7nm+).
 

Bullwinkle M

Posts: 380   +280
Wondering when they'll switch to 5nm, now that'll be an interesting year.
Well.....
That's what I was hoping, but TSMC has 2 nodes online @ 7nm (one EUV and one non-EUV)

If NVidia finds themselves at the back of the line @ TSMC, it may end up being the older 7nm node which is no better than Intels 10nm process (if you are counting transistors per given area)

EUV EUV EUV EUV!
 

daimag

Posts: 15   +2
TSMC 7nm is expensive. Never mind MSRP, these cards will be traded way above current SKU. Gamers need more stimulus package from government.
 

Cooe14

Posts: 77   +107
I heard they left TMSC because of low yields...and that Samsung was the reason that they could have so much better pricing. So I am quite confused about this news of them going back to TSMC.
Yields had absolutely NOTHING to do with it. The problem was the TSMC was refusing to give Nvidia the pricing Nvidia wanted at the capacity levels Nvidia wanted. So Nvidia tried to use their clout & threatened to leave for Samsung unless TSMC dropped their prices & complied to their terms. Samsung for its efforts was willing to give Nvidia a CRAZY good deal on its 8nm wafers to bring them over. Instead, TSMC called their bluff & gave all that 2020 7nm capacity to AMD & Apple instead, leaving Nvidia with no choice but to go with Samsung & their MUCH worse node, but with MUCH cheaper pricing.

(And even with the absurdly good deal Samsung is giving Nvidia on 8nm wafers, the RTX 3000 series is still BARELY making any $ atm, with the RTX 3080 having a supposed BOM ["Bill of Materials"] cost of ≈$600. The dies are all just stupid big again, and the crappy node requires seriously beefed out PCB & VRM designs. Plus GDDR6X isn't particularly cheap. Nvidia & Jensen in particular ended up just playing themselves... and HARD!).
 
  • Like
Reactions: Bamda
Nvidia was likely trying to leverage moving to samsung to get a better deal. Doesn't look like it's worked given the current status of the 30xx series and their quick move back.
It doesnt work like that. If ampere is made on samsung it means the GPU design was designed for samsung's 8nm years ago. You cant just ship the design off to tsmc and churn out gpus. It wont work. It was not designed for tsmc's process technology. It was made for samsung's.
 
  • Like
Reactions: Endymio
So you think NVidia secretly likes to throw away a lot of very expensive chips due to poor performance? And that, instead of scaling, they are happy to resell their cards for hundreds of dollars more than MSRP?
 

Evernessince

Posts: 5,461   +6,134
It doesnt work like that. If ampere is made on samsung it means the GPU design was designed for samsung's 8nm years ago. You cant just ship the design off to tsmc and churn out gpus. It wont work. It was not designed for tsmc's process technology. It was made for samsung's.
Pedantic and needless comment. We are all aware this change doesn't happen overnight.

The time-frame isn't soon, it's 2021.

Both AMD and Nvidia have released the same GPU design on a smaller node within one year in the past, it can certainly be done. I would not be surprised if Nvidia's Ampere was design compatible with both samsung 8nm and TSMC 7nm. Ultimately there's no way to tell unless you ask Nvidia.
 

Endymio

Posts: 1,084   +901
Ultimately there's no way to tell unless you ask Nvidia.
Nonsense. We know more than enough about Samsung's 8LPP process to know it cannot possibly be design-compatible with any of TSMC's 7nm offerings -- density, gate pitch, metal pitch, patterning rules: all are different. Even TSMC's own N7+ process isn't design-compatible with its N7/N7P, and it's far closer to these than is Samsung.
 
  • Like
Reactions: Bamda and daimag

Evernessince

Posts: 5,461   +6,134
Nonsense. We know more than enough about Samsung's 8LPP process to know it cannot possibly be design-compatible with any of TSMC's 7nm offerings -- density, gate pitch, metal pitch, patterning rules: all are different. Even TSMC's own N7+ process isn't design-compatible with its N7/N7P, and it's far closer to these than is Samsung.
1) You make an assumption again, as is your style, without proof.
2) It doesn't matter either way

Pointless.
 

Evernessince

Posts: 5,461   +6,134
Assumption? No.

Breakdown of the CPP, M2P, and densities for Samsung's 8nm and TSMC's 7nm processes in the links below. There are patterning rule differences as well.

https://semiwiki.com/semiconductor-manufacturers/tsmc/285856
https://semiwiki.com/semiconductor-manufacturers/samsung/samsung-10nm-8nm-and-7nm
Where does it state interoperability? It doesn't

You are assuming that those are the only factors that determine design compat. You are certainly no engineer in the field.

Stop talking about a field you are not qualified to speak in depth in.

And again, irrelevant.
 

Endymio

Posts: 1,084   +901
You are assuming that those are the only factors that determine design compat.
They're certainly not the only factors by far. But even one such factor will prevent design-compatibility ... and those charts detail several.

You are certainly no engineer in the field.
I'm no auto engineer either, but I know a Mercedes V-8 engine block won't mate to a 4-cylinder Honda head. Claiming that Samsung's 8LPP node is design-compatible with TSMC's 7nm offerings is equally absurd.