Why doesnt AMD and Nvidia make cards at GloFlow, SMIC or others at 14nm?

[Nena Trinity]

MINERS ARE BUYING ALL YOUR RTX 3080s!!! - YouTube

Okay we are all painfully aware there is a new mining boom, so I was thinking why could they not make RDNA2 & Ampere with 14nm? It would probably have better yields and help fill the markets hunger for graphics chips? Is there ANY reason it would not work?! Remember we had 14nm Vega(56 & 64) and Vega 7nm(VII) so it should work?!!

[CommanderAlex]

But why 14nm when we are already past that point on 7nm/5nm??? 

[Quackers101]

when will they do 28nm??????????????

[jaslion]

Because their cards were designed on the current process node and you can't just scale them back to a larger node? It isn't as simple as just make it bigger that is not how it works at all

[Eigenvektor]

8 minutes ago, Nena Trinity said:

…so I was thinking why could they not make RDNA2 & Ampere with 14nm?

As @jaslion said, you can't just scale up the card. You'd have to redesign and revalidate it, both of which cost time and money. In the end you'd have a card that is way more power hungry and runs a lot hotter than its 7nm equivalent. How many customers would go for that instead of waiting?

 

And, depending on the number of transistors, the 14nm equivalent of a 7nm chip might actually be too large to manufacture. There are size limits to chips and the larger a chip gets, the lower your yields. Lower yields in turn raise prices.

[jaslion]

2 minutes ago, Eigenvektor said:

As @jaslion said, you can't just scale up the card. You'd have to redesign and revalidate it, both of which cost time and money. In the end you'd have a card that is way more power hungry and runs a lot hotter than its 7nm equivalent. How many customers would go for that instead of waiting?

 

And, depending on the number of transistors, the 14nm equivalent of a 7nm chip might actually be too large to manufacture. There are size limits to chips and the larger a chip gets, the lower your yields. Lower yields in turn raise prices.

People tried with whole wafer chips in the past and well it was a big failure with the incredibly high failure rate and how impossibly hard it was too cool and whatnot. They are trying again with riscV stuff tho and it is moderately successful at best by allowing failed dyes and parts of them on the big chip by basically treating it as a whole bunch of chips on their own in one big one.

[Nena Trinity]

14 minutes ago, CommanderAlex said:

But why 14nm when we are already past that point on 7nm/5nm??? 

Yields are still shit, also 7 and 8 can only be made by TSMC and Samsung! It would double production.

14 minutes ago, Quackers101 said:

when will they do 28nm??????????????

Never that is way to big...

12 minutes ago, jaslion said:

Because their cards were designed on the current process node and you can't just scale them back to a larger node? It isn't as simple as just make it bigger that is not how it works at all

14nm can be scaled to 7nm why not the other way? Vega proved it can be done! Radeon VII anyone?!!

[Eigenvektor]

2 minutes ago, jaslion said:

People tried with whole wafer chips in the past and well it was a big failure with the incredibly high failure rate and how impossibly hard it was too cool and whatnot. They are trying again with riscV stuff tho and it is moderately successful at best by allowing failed dyes and parts of them on the big chip by basically treating it as a whole bunch of chips on their own in one big one.

Interesting. But if I understand you correctly, the whole "wafer chip" essentially needs to be designed to be tolerant to production errors. I would also assume that such a big chip would be quite expensive to manufacture, so not exactly the type of design you want in a GPU.

[Vishera]

7 minutes ago, Nena Trinity said:

14nm can be scaled to 7nm why not the other way?

It's a problem because you will have less transistors due to the difference in transistor density.

Less transistors = less performance,and there is also the efficiency and it's effect on both power consumption and heat,it will be hotter and more power hungry on 14nm,

You have seen the giant coolers that NVIDIA use for the 30 series,and the power requirements for 30 series are high already.

[jaslion]

3 minutes ago, Nena Trinity said:

Yields are still shit, also 7 and 8 can only be made by TSMC and Samsung! It would double production.

Never that is way to big...

14nm can be scaled to 7nm why not the other way? Vega proved it can be done! Radeon VII anyone?!!

It wasn't scaled tho it was remade based on existing stuff. This same thing has happened with a bunch of cpu's in the past too. You literally can't just shrink stuff.

Yields aren't bad.

You say 28nm is too big but so is 14nm now. Those gpu's quite literally couldn't exist on 14nm as the chances of them consuming so much power and producing so much heat they can never operate in a regular desktop computer is just certain to happen.

 

Upping production doesn't matter if the product physically cannot exist. That and yields would be awful. The bigger a dye the worse yields get basically.

[jaslion]

3 minutes ago, Eigenvektor said:

Interesting. But if I understand you correctly, the whole "wafer chip" essentially needs to be designed to be tolerant to production errors. I would also assume that such a big chip would be quite expensive to manufacture, so not exactly the type of design you want in a GPU.

Oh yeah totally. Was more to illustrate that sure big chips have been and are again attempted to be made but it's not doable on a wide scale/at all.

[Eigenvektor]

3 minutes ago, Nena Trinity said:

14nm can be scaled to 7nm why not the other way? Vega proved it can be done! Radeon VII anyone?!!

I don't think anyone is saying that it is technically impossible. But it doesn't make much sense. A 14 nm chip would be bigger, require more power and hence more cooling. So a 14nm variant would be worse than its 7nm counterpart in every way.

[Sakkura]

The Vega 10 GPU used in the RX Vega cards had 12.5 billion transistors.

 

The Navi 21 GPU used in the RX 6000 cards has 26.8 billion transistors.

 

Manufacturing that at 14nm would not be feasible. Way too expensive, way too power-hungry, and even the yields wouldn't be good since the chip would be enormous. Over a thousand mm² in theory.

[Eigenvektor]

8 minutes ago, jaslion said:

Oh yeah totally. Was more to illustrate that sure big chips have been and are again attempted to be made but it's not doable on a wide scale/at all.

Gotcha, so there's no real "size limit" from a technical perspective but as you said above, due to bad yields and manufacturing costs it doesn't make economical sense to exceed a certain size.

[porina]

We have a real world example of taking something designed for a smaller node and making it on a bigger one: Rocket Lake. Ok, it isn't exactly out yet, but it essentially takes the 10nm core design from Ice Lake mobile (mid 2019) and re-lay it to be made on their 14nm process. We're looking at roughly a year and a half between Ice Lake and Rocket Lake. Semiconductor design unfortunately isn't something that can be done quickly. If they were designed from the start, in parallel, to be made on multiple processes it might be an option. But to decide later on, it'll be near end of life by the time it comes to market. Rocket Lake is expected to be out for less than a year before it is succeeded.

 

There's another big problem, literally: die size. The higher end GPUs are already large packages. To make them on older process would make them even bigger... it is not going to end well.

[xg32]

doesn't matter to amd and nvidia, they still sell cards, backporting is almost never worth it.

[Vishera]

21 minutes ago, porina said:

We have a real world example of taking something designed for a smaller node and making it on a bigger one: Rocket Lake. Ok, it isn't exactly out yet, but it essentially takes the 10nm core design from Ice Lake mobile (mid 2019) and re-lay it to be made on their 14nm process. We're looking at roughly a year and a half between Ice Lake and Rocket Lake. Semiconductor design unfortunately isn't something that can be done quickly. If they were designed from the start, in parallel, to be made on multiple processes it might be an option. But to decide later on, it'll be near end of life by the time it comes to market. Rocket Lake is expected to be out for less than a year before it is succeeded.

 

There's another big problem, literally: die size. The higher end GPUs are already large packages. To make them on older process would make them even bigger... it is not going to end well.

Also Rocket Lake doesn't have 10 cores because of the back porting.

[Mark Kaine]

my wild guess is it would work but it would be hot as hell and have even more issues than current "next gen GPUs" 

 

Also, in all seriousness, would likely get into trouble with environmental regulations for being reckless af... 

 

 

Also neither nv or amd care who they sell to, I imagine they even prefer "miners" as they buy in bulk, which vendor doesn't like that... 

[Nena Trinity]

6 hours ago, Sakkura said:

The Vega 10 GPU used in the RX Vega cards had 12.5 billion transistors.

 

The Navi 21 GPU used in the RX 6000 cards has 26.8 billion transistors.

 

Manufacturing that at 14nm would not be feasible. Way too expensive, way too power-hungry, and even the yields wouldn't be good since the chip would be enormous. Over a thousand mm² in theory.

 

6 hours ago, Eigenvektor said:

Gotcha, so there's no real "size limit" from a technical perspective but as you said above, due to bad yields and manufacturing costs it doesn't make economical sense to exceed a certain size.

 

6 hours ago, porina said:

We have a real world example of taking something designed for a smaller node and making it on a bigger one: Rocket Lake. Ok, it isn't exactly out yet, but it essentially takes the 10nm core design from Ice Lake mobile (mid 2019) and re-lay it to be made on their 14nm process. We're looking at roughly a year and a half between Ice Lake and Rocket Lake. Semiconductor design unfortunately isn't something that can be done quickly. If they were designed from the start, in parallel, to be made on multiple processes it might be an option. But to decide later on, it'll be near end of life by the time it comes to market. Rocket Lake is expected to be out for less than a year before it is succeeded.

 

There's another big problem, literally: die size. The higher end GPUs are already large packages. To make them on older process would make them even bigger... it is not going to end well.

 

6 hours ago, Vishera said:

Also Rocket Lake doesn't have 10 cores because of the back porting.

 

3 hours ago, Mark Kaine said:

my wild guess is it would work but it would be hot as hell and have even more issues than current "next gen GPUs" 

 

Also, in all seriousness, would likely get into trouble with environmental regulations for being reckless af... 

 

 

Also neither nv or amd care who they sell to, I imagine they even prefer "miners" as they buy in bulk, which vendor doesn't like that... 

But the idea is to make it more power hungry so the miners wouldnt wanna touch it! Also we would maybe at best reach a RX 6600 or 6700 but considering EVERYTHING IS SOLD OUT, why not do it? AMD and Nvidia will be able to sell em anyway people are desperate... UwU

[Mark Kaine]

9 minutes ago, Nena Trinity said:

But the idea is to make it more power hungry so the miners wouldnt wanna touch it!

I mean from that perspective it makes sense, but let's go crazy a bit for a moment, so they do this... need ridiculous over sized coolers, likely "1500w" PSUs, and then as I said will likely still run into reliability issues because of how hot these things run, in then undersized little PC boxes... 

 

What I'm saying is it doesnt really make sense from the manufacturer side, and even consumer side might not actually like it.

 

13 minutes ago, Nena Trinity said:

EVERYTHING IS SOLD OUT,

it's not though...? 

 

I just got a 3070 and could choose between 5, 6... cards at least. 

 

(and in hindsight I stupidly forgot to check on Ali express for the card I really wanted... an i game 3070) ¯\_(ツ)_/¯

[Nena Trinity]

2 minutes ago, Mark Kaine said:

I mean from that perspective it makes sense, but let's go crazy a bit for a moment, so they do this... need ridiculous over sized coolers, likely "1500w" PSUs, and then as I said will likely still run into reliability issues because of how hot these things run, in then undersized little PC boxes... 

 

What I'm saying is it doesnt really make sense from the manufacturer side, and even consumer side might not actually like it.

 

it's not though...? 

 

I just got a 3070 and could choose between 5, 6... cards at least. 

 

(and in hindsight I stupidly forgot to check on Ali express for the card I really wanted... an i game 3070) ¯\_(ツ)_/¯

Here the only GPUs in stock were the GT 710, GT 1030 and RX 5500 XT some other stores have them but jacked the price up to insanity which at that point what is the point? Nope they only have 1650s now... UwU

[The Blackhat]

This is all very obviously just a troll post.

[Vishera]

3 minutes ago, Mark Kaine said:

What I'm saying is it doesnt really make sense from the manufacturer side, and even consumer side might not actually like it.

Correct,bigger chips have lower yields and are more expensive,since a larger die = less GPUs/CPUs per wafer.

[Nena Trinity]

2 minutes ago, Vishera said:

Correct,bigger chips have lower yields and are more expensive,since a larger die = less GPUs/CPUs per wafer.

The architecture also contribute to better performance & efficiency right? Even if the top model chip doesnt equal a 3080 and we are stuch with a 3060Ti or 3070 its not a total loss...

[Mark Kaine]

3 minutes ago, Nena Trinity said:

Here the only GPUs in stock were the GT 710, GT 1030 and RX 5500 XT some other stores have them but jacked the price up to insanity which at that point what is the point? Nope they only have 1650s now... UwU

... to be honest, I just checked, they're still available, but not as many and prices went up quite a bit compared to a week ago... but considering the new year just started that makes sense and I expect stock to get a bit better again soon... 

 

Also where's here, eBay usually has them, ali express has them, and if you think you can ever buy one at msrp I must tell you that you're mistaken, you need to at least calculate 200$/€ more... which honestly makes sense cause msrp was never meant serious and aftermarket cards are actually better than FE and simply more pricey too. 

 

 

6 minutes ago, The Blackhat said:

This is all very obviously just a troll post.

It's also kinda obvious you don't really know what trolling even is (unless you're trolling lol) 

 

6 minutes ago, Vishera said:

Correct,bigger chips have lower yields and are more expensive,since a larger die = less GPUs/CPUs per wafer.

oh I was right with something, mark the calendar, folks!

 

(no seriously I thought more about failure rate, design cost, heat, pitchforks, etc, but that also makes sense!)