News Nvidia Will Use TSMC for Ampere 8nm Successor in 2021 (Rumor)

[Admin]

Recent reports suggest that Nvidia is moving to TSMC's 7nm process for its Ampere GPUs.

Nvidia Will Use TSMC for Ampere 8nm Successor in 2021 (Rumor) : Read more

 

[hotaru.hino]

Something of note since I saw in the other comments that people think NVIDIA and TSMC have a rocky relationship due to NVIDIA chewing TSMC out.

My cursory glance of the interwebs on the history of this points to two prominent instances where NVIDIA wanted to jump on a new process node, but had trouble with it. The first was the GeForce FX series, going on the then-new 130nm node and TSMC were having issues with that node ([1] [2] [3]). The second was the GeForce 400 series at 40nm ([1] [2]). While you could argue that if TSMC had issues at 40nm, then AMD would've brought it up as well since the Radeon HD 5000 series was based on the same process, but apparently AMD only had to pay per chip, whereas NVIDIA had to pay per wafer.

If anything, I think NVIDIA learned its lesson to not jump on the latest process node (hence why Turing was 12nm even though 7nm was around) or at least take a more cautious approach. This time around, since everyone and their mother wants TSMC 7nm, it was probably better from a production capacity standpoint to go with someone who wasn't as burdened. Though given the shortages we're seeing with Ampere...

Either way, while sure, NVIDIA does have its hiccups with its architectures not being up to snuff, them yelling at their manufacturer isn't entirely unwarranted.

 

[JarredWaltonGPU]

Something of note since I saw in the other comments that people think NVIDIA and TSMC have a rocky relationship due to NVIDIA chewing TSMC out.

My cursory glance of the interwebs on the history of this points to two prominent instances where NVIDIA wanted to jump on a new process node, but had trouble with it. The first was the GeForce FX series, going on the then-new 130nm node and TSMC were having issues with that node ([1] [2] [3]). The second was the GeForce 400 series at 40nm ([1] [2]). While you could argue that if TSMC had issues at 40nm, then AMD would've brought it up as well since the Radeon HD 5000 series was based on the same process, but apparently AMD only had to pay per chip, whereas NVIDIA had to pay per wafer.

If anything, I think NVIDIA learned its lesson to not jump on the latest process node (hence why Turing was 12nm even though 7nm was around) or at least take a more cautious approach. This time around, since everyone and their mother wants TSMC 7nm, it was probably better from a production capacity standpoint to go with someone who wasn't as burdened. Though given the shortages we're seeing with Ampere...

Either way, while sure, NVIDIA does have its hiccups with its architectures not being up to snuff, them yelling at their manufacturer isn't entirely unwarranted.

The thing is, Nvidia went with TSMC N7 for Ampere A100, so it's mostly a question of price and capacity I think. Clearly, the A100 is the "price doesn't even matter" market segment. But AMD is doing Zen 2, Zen 3, Navi 10, Navi 12, and Navi 14 all on N7 -- and most likely Navi 21/22 as well. Apple has been doing A12 since 2018 and A13 (N7P) in 2019. Obviously a much smaller chip than Nvidia's GPUs, but even so, no one else is complaining about TSMC's 7nm nodes. Which is the problem, apparently: they're so good that TSMC can charge more, and demand has surpassed (wafer) supply.

 

[Kamen Rider Blade]

The question is, will Jensen Huang let his ego get in the way of good business decisions?

Will Jensen Huang accept the higher prices per waffer and go crawling back to TSMC and beg them to manufacture their GPU's?

Will Samsung try to sweeten the per Waffer price even more to keep nVIDIA as a partner?

Time will tell.

 

[JarredWaltonGPU]

The question is, will Jensen Huang let his ego get in the way of good business decisions?

Will Jensen Huang accept the higher prices per wafer and go crawling back to TSMC and beg them to manufacture their GPU's?

Will Samsung try to sweeten the per Wafer price even more to keep Nvidia as a partner?

Time will tell.

Considering Nvidia has already partnered with TSMC, I strongly doubt that it was pure ego getting in the way. If Nvidia went to TSMC last year and said, basically, "We need about 200,000 wafers of GA102 by September, 400,000 wafers of GA104 by October ... oh, and we're still in need of more wafers for GA100," the answer might simply have been that TSMC couldn't do it on the timeframe Nvidia wanted.

 

[hotaru.hino]

The thing is, Nvidia went with TSMC N7 for Ampere A100, so it's mostly a question of price and capacity I think. Clearly, the A100 is the "price doesn't even matter" market segment. But AMD is doing Zen 2, Zen 3, Navi 10, Navi 12, and Navi 14 all on N7 -- and most likely Navi 21/22 as well. Apple has been doing A12 since 2018 and A13 (N7P) in 2019. Obviously a much smaller chip than Nvidia's GPUs, but even so, no one else is complaining about TSMC's 7nm nodes. Which is the problem, apparently: they're so good that TSMC can charge more, and demand has surpassed (wafer) supply.

I would like to add since now that I look at it, that NVIDIA probably doesn't want to go on the latest and greatest thing the moment it comes out. 7nm now is fine, but 7nm a couple of years ago had teething problems and a lot of fabs that were looking into going 7nm simply dropped it because it was costing too much.

EDIT: Actually I take back what I said about Turing, since it was likely just Volta with RT cores tacked onto it, Volta was a 12nm design to begin with, and it was likely designed before 7nm had its first "production grade" part manufactured.

 

[InvalidError]

The question is, will Jensen Huang let his ego get in the way of good business decisions?

As pointed out by others, GA100 is on TSMC 7nm, so Nvidia clearly has no problems going to TSMC when cost is no object for ultra-high-margin parts. Also as already mentioned by others, I too believe Nvidia's decision to go with Samsung 8nm is mainly because of TSMC's 7nm backlog - TSMC being superior is pointless if you cannot book anywhere near enough wafers on the process you want to use.

 

[MasterMadBones]

Regardless of what Nvidia's relationship with TSMC is, there are also some noises that Nvidia is already pumping huge amounts of money into Samsung to assist them in developing their nodes. If this is true, it's probably part of a long term plan to be able to avoid potential issues with high demand from TSMC and the inflated pricing that would cause.

hence why Turing was 12nm even though 7nm was around

Mostly because 7nm was a very, very young process at the time and not suitable for large dies at all. Even AMD had availability issues with Vega 20 GPUs more than 4 months later. It was not a matter of taking a "calculated risk", it would simply have been a terrible decision.

 

[NightHawkRMX]

Considering Samsung can't seem to make more than about a single GPU die per week, I would hope they would come back to a company that doesn't have AS MANY production issues.

 

[zodiacfml]

Looks like Samsung has pretty bad yield issues and Nvidia is not happy and so are the customers. The hype for the new cards is really good but users can't get their hands on the cards as quickly, wasting precious hype until AMD announces/releases their new cards.
Congrats to TSMC for doing so well, so well that I'm predicting a future where Intel has fewer fabs until they let go of it like AMD did

 

[wr3zzz]

This really begs the question why UMC and GlobalFoundry chose not to invest in 7nm.

 

[A_paid_reader]

Something of note since I saw in the other comments that people think NVIDIA and TSMC have a rocky relationship due to NVIDIA chewing TSMC out.

My cursory glance of the interwebs on the history of this points to two prominent instances where NVIDIA wanted to jump on a new process node, but had trouble with it. The first was the GeForce FX series, going on the then-new 130nm node and TSMC were having issues with that node ([1] [2] [3]). The second was the GeForce 400 series at 40nm ([1] [2]). While you could argue that if TSMC had issues at 40nm, then AMD would've brought it up as well since the Radeon HD 5000 series was based on the same process, but apparently AMD only had to pay per chip, whereas NVIDIA had to pay per wafer.

If anything, I think NVIDIA learned its lesson to not jump on the latest process node (hence why Turing was 12nm even though 7nm was around) or at least take a more cautious approach. This time around, since everyone and their mother wants TSMC 7nm, it was probably better from a production capacity standpoint to go with someone who wasn't as burdened. Though given the shortages we're seeing with Ampere...

Either way, while sure, NVIDIA does have its hiccups with its architectures not being up to snuff, them yelling at their manufacturer isn't entirely unwarranted.

I think what happened was more like a posturing situation.

Since AMD tease big navi for like years in advance and the fan were hyping it up. To the point that some say big navi is going to be nvidia killer. Obviously nvidia can't just let rumor like that go. So, nvidia probably decide that an important factor for ampere is to have their card launch first. Instead of leaving a gap where AMD can take performance crown.

But when the time come pretty much everyone is launching their stuff. Taking away TSMC time slot. So much so that TSMC couldn't make the ampere launch before big navi happens without some serious money. So, nvidia is left to either
A. let big navi launch first
B. use alternative foundry
and they pick option B with samsung. And here we are.

 

[Olle P]

I think the real question isn't if Nvidia will have consumer GPUs made by TSMC, but when.
If they'll just wait in line for production capacity the cards might be available on the market some time H2'21.
If the GA100 sell considerably worse than expected that might free up resources for consumer GPUs sooner...

 

[JarredWaltonGPU]

I think the real question isn't if Nvidia will have consumer GPUs made by TSMC, but when.
If they'll just wait in line for production capacity the cards might be available on the market some time H2'21.
If the GA100 sell considerably worse than expected that might free up resources for consumer GPUs sooner...

All indications I've seen are that the supercomputer and deep learning research markets are buying DGX A100 basically as fast as Nvidia can produce them. Needless to say, profit margins on such systems are much higher than on consumer cards. Granted, it's difficult to get numbers for how many are actually being sold in this sort of market -- anyone want to get involved with companies offering to see DGX A100 and request a quote? LOL

 

[hotaru.hino]

This really begs the question why UMC and GlobalFoundry chose not to invest in 7nm.

It got too expensive to continue the R&D.

I think what happened was more like a posturing situation.

Since AMD tease big navi for like years in advance and the fan were hyping it up. To the point that some say big navi is going to be nvidia killer. Obviously nvidia can't just let rumor like that go. So, nvidia probably decide that an important factor for ampere is to have their card launch first. Instead of leaving a gap where AMD can take performance crown.

But when the time come pretty much everyone is launching their stuff. Taking away TSMC time slot. So much so that TSMC couldn't make the ampere launch before big navi happens without some serious money. So, nvidia is left to either
A. let big navi launch first
B. use alternative foundry
and they pick option B with samsung. And here we are.

I'd say that's part of the equation, but at the same time, TSMC is pretty much everyone's foundry at this point. They basically manufacture every non-Intel/Samsung processor that's not an embedded system SoC.