Nvidia Post-Volta (Ampere?) Rumor and Speculation Thread

And it seems "enough" for nVidia. 20x0 rtx are made by samsung too, right ? The process seems fine... In real gfx products it's not like tsmc's are destroying samsung's...

 

Turings (GTX16/RTX20) are manufactured on TSMC 12nm (aka improved 16nm)
Edit: the only(?) NVIDIA GPU manufactured so far at Samsung is GP107

 

And GP108?

 

Ah you're right. I don't know why I had in mind that nVidia was doing Turings with a custom Samsung 12nm solution. I'm getting old : /

 

True, GP108 is Samsung too

 

The CEO put the argument to rest today when he announced that NVIDIA will be using 7nm from both TSMC and Samsung for their next gen GPUs, with the majority of orders from TSMC.

https://wccftech.com/nvidia-tsmc-7nm-next-generation-gpu-ampere/

 

From that link

''This makes 2020 all the more exciting since we will have full-fledged GPU lineups from all three vendors, Intel, NVIDIA and AMD.''

Interesting indeed, any more info on RTX3000 series, like how many TF they will be, RT etc?

 

Huh, consumer GPUs from TSMC and professional line (Volta successor) from Samsung? I know a few mobile companies will tape out the same SOC on both foundries, but afaik it's not common and seems expensive.

Either way, sad to see Samsung get shorter shrift from most companies. This just seems to confirm TSMC offers the better process right now, but with all of 2 contract cutting edge foundries left everyone prioritizing one over the other is the slow path to a monopoly. At least AMD has announced they'll not be jumping on TSMCs 5nm right quick, leaving the possibility of switching to Samsung circa 2021/2 open.

 

Could Samsung be doing the cards which use its Flashbolt HBM2E?
Is this what the (refresh) "V100S" is using?

 

You'll never build HPC GPUs at Samsung, TSMC is a too strong partner in this space and executing flawless. Samsung would be a too high risk, even if their process wouldn't be that bad. Low-Midrange should be again from Samsung, like we had it with pascal. It might even be, that Samsungs process is so bad, that Nvidia had to switch to TSMC with consumer products. How much delay would you get if you do that? I think that i read about ~6 months sometime ago, if you switch pretty late in the design process (still before tapeout of course).

Yes, TSMC already has more or less a monopoly, but the companies have no choice. You need the best process to compete and TSMC is just superior. It seemed as if Samsung might have some advantage with EUV, but it turned out no. Samsungs roadmap is also very weak compared to TSMC, so TSMC will be 1-2 years ahead in the future. Only their 3GAA timeframe seems aggressive, but they had so many delays with their process, that i wouldn't take their roadmap seriously.

 

Well unfortunately, nV decided to do EXACTLY THAT, and it didn't work even a little bit.

 

Some other reasons for staying with TSMC's foundry is their 5nm advantages.

https://semiengineering.com/5nm-vs-3nm/

 

Which GPU are you referring to?

 

But that is not what was being said when Nvidia announced their Samsung deal. Didn't nvidia come right out and say (8 months ago) that they were shopping elsewhere because the price TSMC was charging for 7nm...? And that Samsung EUV process was much cheaper, etc..


And, nobody has mentioned the fact that Dr Lisa Su signed a deal with TSMC over 3 years ago, for exclusive rights for 7nm TSMC in 2018. Which didn't allow Nvidia to adopt and have volume orders last year. Thus, forcing Nvidia to seek Samsung and belay their new architecture. Which is still being shopped around, as nobody has heard of Ampere/Hopper being taped out.

 

Wait, What???

 

They didn't announce any deals.

No.

Lisa Su didn't sign any "exclusive" deals with TSMC either.

The only thing which is "delaying" NV's next gen architecture right now is the apparent inability of all of NV's competitors to beat whatever products NV has currently on 16/12nm. Some of said competitors aren't able to beat NV's products even with the help of said 7nm production process.

 

Does 7nm process help? Jen-Hsun Huang said it doesn't, because it doesn't bring any advantages.

 

It does, when you're not power limited at transistor counts which can be achieved with less advanced production process for similar or lower cost.

 

There is a time factor to his statement, he means "currently", as in vs 12nm during the launch of Turing.

 

Not true. No such agreement exists (and its punishable by law BTW)

Nope. Ampere taped out on March 2019.