Strong indication nvidia is moving from samsung to TSMC foundry in parallel

[rectifryer]

for manufacturing of current RTX cards.

https://www.youtube.com/watch?v=OGe3VriThqs

Cliffs:

-Samsung can't keep up
-Venders getting single digit percentages of orders they placed (from comments, not video)
-TSMC has extra availability since huawei is not a stable partner
-Nvidia has done this before
-8nm cards out now from samsung
-7nm RTX cards will come from TSMC early next year

I'm going to wait to see if a 7nm RTX3090ti comes available LMAO

[bongowongo]

for manufacturing of current RTX cards.

https://www.youtube.com/watch?v=OGe3VriThqs

Cliffs:

-Samsung can't keep up
-Venders getting single digit percentages of orders they placed (from comments, not video)
-TSMC has extra availability since huawei is not a stable partner
-Nvidia has done this before
-8nm cards out now from samsung
-7nm RTX cards will come from TSMC early next year

I'm going to wait to see if a 7nm RTX3090ti comes available LMAO

So you're saying 7nm makes the card a lot better?

[bartosh]

So you're saying 7nm makes the card a lot better?

Depends on how 8nm Samsung compares to 7nm tsmc since neither are truly that size

https://www.hpcwire.com/2020/06/01/1...c-be-replaced/

Same way that intel 10nm is comparable to amd 7nm

But yeah, tsmc 7 is better than Samsung 8

[EliteBrah]

samsung 8nm is 7.6nm in actual

and tsmc 7nm is 7.4nm in actual

cards would have same bios, same boost behavior, ext. does the 2% smaller size equate to better overclocking? not likely but maybe you can squeeze an additional 15Mhz out of it from slightly cooler temps bringing out better boost table.
your better off investing in a hybrid card as opposed to waiting for 7nm

[rectifryer]

So you're saying 7nm makes the card a lot better?

I'm saying it will be the actual launch of the Nvidia rtx 3000 series.

[EliteBrah]

It's not clear they will bring the 3080 and 3090 to 7nm anyways. Evga is more concerned with the other ampere cards as thats where the sustained demand will be. I.E. 3060, 3060 super, 3070, ext

[Greg14]

So you're saying 7nm makes the card a lot better?

I think actually being able to get one of these cards will also make the card a lot better

[R3L3NTL3SS]

Lmao @ people thinking that somehow TSMC cards would be any better. They wouldn't be. IF - IF Nvidia uses TSMC, it will be alongside with Samsung and likely be for the lower end SKU's I bet.

But by all means - I hope everyone waits for this new TSMC mega card they've all dreamed up in their head. Then maybe I can finally get my hands on a 3080 sooner rather than later.

[FAPhaggot]

So a quick note about nodes

Up until 28nm, transistors used to look like McDonald's french fries.



Naming them was pretty easy: how wide's the shorter side of your french fry?

Then once they ran into problems shrinking those, they developed a new transistor technology that looks like Chick Fil A fries.



and chit got more complicated.

On Intel's 14nm node for example, the size of each of those waffle squares is actually 42nm each. But if you look where the thick waffles cross over the thin ones, the thin ones are surrounded on 3 sides. So Intel's reasoning is like, "our french fries are 42nm apart, but the french fries are now surrounded on 3 sides, 42nm /3 = 14nm equivalent for a french fry that only touches one side."

Other manufacturers took more complex routes for naming their nodes, generally with the goal of fudging it smaller to look good for marketing.

-------------------------------------

So I couldn't tell you the exact device sizes for Samsung 8nm and TSMC 7nm, but they're from the same generation, they're in the same ballpark. The differences in size and speed between them aren't a huge issue.

What the killers are for Samsung are

-power consumption/heat (they have leaky french fries that sap a lot of current)

-and yield.

Samsung won't be up front about what their yield rates are, but it's rumored to be like 30%. IE, if you try to manufacture three 3090s, two will be broken. With some of those, you might be able to lazer cut out say some defective CUDA cores and call it a 3080, but some of those won't be salvageable. Single-digit yields would be absolutely catastrophic if true, like somewhere between "how did this node actually get released", and "how is your semiconductor fab arm still in business."

On the other hand, TSMC openly brags that their 7nm yields are above 90%. Some people are wondering if they've had to actually damage perfectly good Ryzen 9 processors to make Ryzen 7s and chit, cause 19 out of 20 Ryzen 9s they try to make are going to come out flawless. Like, this is the polar opposite of Samsung chit.

So if NVidia partners up with Chad TSMC, they will basically get every graphics card they order. I don't know how much room TSMC is offering or how much they plan to buy, but there's not going to be this "whoops, all your fukkin dies broke" like they're getting from Samsung. It's going to be a very reliable supply. And the heat issues that have people calling the 3090 a radiator and chit, those will be improved too, just by virtue of switching processes.

As far as how long redesigning everything for TSMC 7 nano would take... the digital layout for the RAM memory and chit is all automated so that can be done in like a week, the analog layout for the clocks and power network is a lot more complicated and would have to all be redone by hand, and all the packaging and chit would have to be re-verified for the smaller parts (ie, do our graphics card pins still send out data accurately with these new parts?). Ballpark estimate, I figure doing the latter two would take about 6 months, but NVidia will probably go slave-driver mode and shave a month or two off that by forcing their engineers to work like fukkin 80 hour weeks. It'd suck dick to work there right now, srs.

Lmao @ people thinking that somehow TSMC cards would be any better. They wouldn't be. IF - IF Nvidia uses TSMC, it will be alongside with Samsung and likely be for the lower end SKU's I bet.

But by all means - I hope everyone waits for this new TSMC mega card they've all dreamed up in their head. Then maybe I can finally get my hands on a 3080 sooner rather than later.

They won't have any different performance (maybe a 5% gain like 2070/2070 Super tops), but you'll actually be able to fukking buy them.

[rocketfish]

samsung's foundry has bad yields.. reported to be 30%.. terrible numbers. (i know it doesn't work that way cause SOMETIMES you can cut it down and make them a 3080 or 3070 basically)

tsmc had beef with nvidia over pricing a while back, but obviously you get what you pay for and I think nvidia learned their lesson because Nvidia RTX 3000 series isn't selling out because everyone and there momma wants one, but because stores are only receiving 3-5 cards a week... some stores zero.

when nvidia wanted tsmc, tsmc told them no thanks we are full, but I think nvidia is giving in to their pricing now.

This does not mean rtx 3090/3080/3070 will be on TSMC.. but their next cards, like a refreshed version next year in 2021

[bartosh]

So a quick note about nodes

Up until 28nm, transistors used to look like McDonald's french fries.



Naming them was pretty easy: how wide's the shorter side of your french fry?

Then once they ran into problems shrinking those, they developed a new transistor technology that looks like Chick Fil A fries.



and chit got more complicated.

On Intel's 14nm node for example, the size of each of those waffle squares is actually 42nm each. But if you look where the thick waffles cross over the thin ones, the thin ones are surrounded on 3 sides. So Intel's reasoning is like, "our french fries are 42nm apart, but the french fries are now surrounded on 3 sides, 42nm /3 = 14nm equivalent for a french fry that only touches one side."

Other manufacturers took more complex routes for naming their nodes, generally with the goal of fudging it smaller to look good for marketing.

-------------------------------------

So I couldn't tell you the exact device sizes for Samsung 8nm and TSMC 7nm, but they're from the same generation, they're in the same ballpark. The differences in size and speed between them aren't a huge issue.

What the killers are for Samsung are

-power consumption/heat (they have leaky french fries that sap a lot of current)

-and yield.

Samsung won't be up front about what their yield rates are, but it's rumored to be like 30%. IE, if you try to manufacture three 3090s, two will be broken. With some of those, you might be able to lazer cut out say some defective CUDA cores and call it a 3080, but some of those won't be salvageable. Single-digit yields would be absolutely catastrophic if true, like somewhere between "how did this node actually get released", and "how is your semiconductor fab arm still in business."

On the other hand, TSMC openly brags that their 7nm yields are above 90%. Some people are wondering if they've had to actually damage perfectly good Ryzen 9 processors to make Ryzen 7s and chit, cause 19 out of 20 Ryzen 9s they try to make are going to come out flawless. Like, this is the polar opposite of Samsung chit.

So if NVidia partners up with Chad TSMC, they will basically get every graphics card they order. I don't know how much room TSMC is offering or how much they plan to buy, but there's not going to be this "whoops, all your fukkin dies broke" like they're getting from Samsung. It's going to be a very reliable supply. And the heat issues that have people calling the 3090 a radiator and chit, those will be improved too, just by virtue of switching processes.

As far as how long redesigning everything for TSMC 7 nano would take... the digital layout for the RAM memory and chit is all automated so that can be done in like a week, the analog layout for the clocks and power network is a lot more complicated and would have to all be redone by hand, and all the packaging and chit would have to be re-verified for the smaller parts (ie, do our graphics card pins still send out data accurately with these new parts?). Ballpark estimate, I figure doing the latter two would take about 6 months, but NVidia will probably go slave-driver mode and shave a month or two off that by forcing their engineers to work like fukkin 80 hour weeks. It'd suck dick to work there right now, srs.



They won't have any different performance (maybe a 5% gain like 2070/2070 Super tops), but you'll actually be able to fukking buy them.

Quoting because this is great info. Read this

[Destor]

Yields would tie directly to cost right? Seems surprising these cards could be offered at their prices with such low yields

[R3L3NTL3SS]

Yields would tie directly to cost right? Seems surprising these cards could be offered at their prices with such low yields

Most likely the huge majority of them are able to be saved and used in 3080, 3070's.
That's probably why the price difference between the 3080 and 3090 is so massive. If out of 100 chips, only 30 are good enough for a 3090, that definitely explains the massive price gap.

[R3L3NTL3SS]

On a possibly related subject, videocardz.com is now saying that the rumored 3080 20gb card, as well as all the other rumored 3000 variants, have been cancelled.
Not sure if you can truly call it cancelled since, as far as I know, it was never anything more than rumor, but it certainly is interesting that this comes out around the same time as rumors of Nvidia looking to TSMC for chips.

https://videocardz.com/newz/nvidia-a...-rtx-3070-16gb

[ClivesTriceps]

If it increases supply, then cool.

Doesn't AMD use TSMC as well? Imagine being so desperate to sell cards you need to share the same house as your competitor lol

[FAPhaggot]

Yields would tie directly to cost right? Seems surprising these cards could be offered at their prices with such low yields

Yep.

Semiconductor manufacturing is weird in that the cost is alllll overhead. Building a modern fab plant costs 10 billion dollars. Each of the lithography machines cutting the wafers is like $100M each. But the raw materials for the wafers themselves? Just a few cents each. The saying in the business is that for an old process where the machinery's already paid for itself, "the cost of the chip is just the sand and the gold."

So figuring out the costs is complicated. You're right that Samsung having 70% of failed wafers = 70% of potential money that your tooling's not paying back. But the price they're asking for NVidia to manufacture each of those wafers mostly depends on how much money they owe on the machinery and the building.

Samsung 8nm is an extension of Samsung 10nm, so they already own the site and the tooling, they're just refining them. On the other side, TSMC built a whole new plant for 7 nano and bought a chit ton of new tooling from Germany, so their expenses are a lot higher. The lithography machines they ordered were actually flown in on a cargo 747 and were so big they had to be taken out through the plane's nose; wonder what the UPS Store bill for that one was.

I have no idea how the math works out on "30% yield on a process that's already had one run to pay for itself (with some of those 70% defects able to be recovered as 3080s)" vs "90%+ yield on a brand new super-expensive process". And the exact $$$ per wafer figures are probably company confidential anyway. But you can be sure that Samsung ended up being cheaper.

NVidia's usually worked with TSMC in the past, so it's kind of weird to see them shift gears to Samsung. IIRC, it happened earlier this year like this spring or summer, even when TSMC had confirmed absolute Chadly yields by Christmas 2019. https://www.reddit.com/r/AMD_Stock/c...firmed_at_009/

So my speculation, backed by absolutely nothing, is that NVidia felt price pressure to compete with AMD Big Navi, and figured they could get an edge by producing the RTX 3000s on cheaper Samsung wafers whereas AMD would spend more to use TSMC. If they stuck with TSMC like usual, it'd be a push. If anyone in the room had two firing brain cells, they would have brought up the yield differential, so I assume that "yields chitting the bed" was decided to be a tolerable risk.

[Destor]

Yep.

Semiconductor manufacturing is weird in that the cost is alllll overhead. Building a modern fab plant costs 10 billion dollars. Each of the lithography machines cutting the wafers is like $100M each. But the raw materials for the wafers themselves? Just a few cents each. The saying in the business is that for an old process where the machinery's already paid for itself, "the cost of the chip is just the sand and the gold."

So figuring out the costs is complicated. You're right that Samsung having 70% of failed wafers = 70% of potential money that your tooling's not paying back. But the price they're asking for NVidia to manufacture each of those wafers mostly depends on how much money they owe on the machinery and the building.

Samsung 8nm is an extension of Samsung 10nm, so they already own the site and the tooling, they're just refining them. On the other side, TSMC built a whole new plant for 7 nano and bought a chit ton of new tooling from Germany, so their expenses are a lot higher. The lithography machines they ordered were actually flown in on a cargo 747 and were so big they had to be taken out through the plane's nose; wonder what the UPS Store bill for that one was.

I have no idea how the math works out on "30% yield on a process that's already had one run to pay for itself (with some of those 70% defects able to be recovered as 3080s)" vs "90%+ yield on a brand new super-expensive process". And the exact $$$ per wafer figures are probably company confidential anyway. But you can be sure that Samsung ended up being cheaper.

NVidia's usually worked with TSMC in the past, so it's kind of weird to see them shift gears to Samsung. IIRC, it happened earlier this year like this spring or summer, even when TSMC had confirmed absolute Chadly yields by Christmas 2019. https://www.reddit.com/r/AMD_Stock/c...firmed_at_009/

So my speculation, backed by absolutely nothing, is that NVidia felt price pressure to compete with AMD Big Navi, and figured they could get an edge by producing the RTX 3000s on cheaper Samsung wafers whereas AMD would spend more to use TSMC. If they stuck with TSMC like usual, it'd be a push. If anyone in the room had two firing brain cells, they would have brought up the yield differential, so I assume that "yields chitting the bed" was decided to be a tolerable risk.

Interesting insight brah

[FAPhaggot]

Interesting insight brah

Glad you guys liked it; just stuff I've picked up through work and school. VLSI knowledge also gets the HBBs soaking, nomsayin.

[nkiritsis13]

Nvidia has to do something with the sh*t show this launch has been, no matter what kind of daft "demand issue" spin they try to pull. There is only so long they can pretend that they aren't producing enough cards to meet the demand even if you subtract scalpers mass buying, and a switch to TSMC is pretty much an admission that this truly was a supply problem all along.

On a possibly related subject, videocardz.com is now saying that the rumored 3080 20gb card, as well as all the other rumored 3000 variants, have been cancelled.
Not sure if you can truly call it cancelled since, as far as I know, it was never anything more than rumor, but it certainly is interesting that this comes out around the same time as rumors of Nvidia looking to TSMC for chips.

https://videocardz.com/newz/nvidia-a...-rtx-3070-16gb

Hopefully this is just indicative of them shifting from 8nm to 7nm, hence why the 8nm cards are crossed out there.

At the very least they could maybe meet somewhere in the middle because 10 GB just doesn't seem future proof enough for me, especially when even AMD is stuffing their cards with 16 GB.

Glad you guys liked it; just stuff I've picked up through work and school. VLSI knowledge also gets the HBBs soaking, nomsayin.

On spread, but great write-up and the fries comparison is on point.

If it increases supply, then cool.

Doesn't AMD use TSMC as well? Imagine being so desperate to sell cards you need to share the same house as your competitor lol

It's not actually uncommon at all, e.g. the RTX cards were TSMC 12nm while the AMD Radeon RX 5000 series used TSMC 7nm.

The reason being that there really aren't a lot of options when it comes to semiconductor manufacturers.
*And I guess we see what happens when Nvidia went with Samsung.

[R3L3NTL3SS]

At the very least they could maybe meet somewhere in the middle because 10 GB just doesn't seem future proof enough for me, especially when even AMD is stuffing their cards with 16 GB.

IMO AMD throwing 16gb on their cards is just a marketing/selling point because they know chip for chip they can't compete with Nvidia.
10gb is really plenty, even for 4K gaming. The only time I think I've read anyone exceeding is with lots of texture mods, etc.
At the end of the day, you can slap all the extra memory you want on a card. You hit the point of diminishing returns extremely quickly.
The rumored 20gb 3080 would probably have been maybe 5-8% more performance over the 10gb model. And they were going to charge $200-$300 more for it. Not worth it at all IMO.

[FAPhaggot]

IMO AMD throwing 16gb on their cards is just a marketing/selling point because they know chip for chip they can't compete with Nvidia.
10gb is really plenty, even for 4K gaming. The only time I think I've read anyone exceeding is with lots of texture mods, etc.
At the end of the day, you can slap all the extra memory you want on a card. You hit the point of diminishing returns extremely quickly.
The rumored 20gb 3080 would probably have been maybe 5-8% more performance over the 10gb model. And they were going to charge $200-$300 more for it. Not worth it at all IMO.

There's more that graphics cards are used for than just desktop gaming. For deep learning, memory is critical. For VR, memory is critical. Calling it empty marketing is taking too narrow of a look at it.

Good thing that I've invested into TSMC

U gonna be rich, brah.

[R3L3NTL3SS]

There's more that graphics cards are used for than just desktop gaming. For deep learning, memory is critical. For VR, memory is critical. Calling it empty marketing is taking too narrow of a look at it.

True, but 99% of people here are buying for gaming.

[FAPhaggot]

True, but 99% of people here are buying for gaming.

I don't even know if I'd say that much. IL uses his video cards for machine learning to create deep fakes. King SWRV uses his for VR to whack off to virtual girls. The market's broadening.

[MiscInformed]

Yep.

Semiconductor manufacturing is weird in that the cost is alllll overhead. Building a modern fab plant costs 10 billion dollars. Each of the lithography machines cutting the wafers is like $100M each. But the raw materials for the wafers themselves? Just a few cents each. The saying in the business is that for an old process where the machinery's already paid for itself, "the cost of the chip is just the sand and the gold."

So figuring out the costs is complicated. You're right that Samsung having 70% of failed wafers = 70% of potential money that your tooling's not paying back. But the price they're asking for NVidia to manufacture each of those wafers mostly depends on how much money they owe on the machinery and the building.

Samsung 8nm is an extension of Samsung 10nm, so they already own the site and the tooling, they're just refining them. On the other side, TSMC built a whole new plant for 7 nano and bought a chit ton of new tooling from Germany, so their expenses are a lot higher. The lithography machines they ordered were actually flown in on a cargo 747 and were so big they had to be taken out through the plane's nose; wonder what the UPS Store bill for that one was.

I have no idea how the math works out on "30% yield on a process that's already had one run to pay for itself (with some of those 70% defects able to be recovered as 3080s)" vs "90%+ yield on a brand new super-expensive process". And the exact $$$ per wafer figures are probably company confidential anyway. But you can be sure that Samsung ended up being cheaper.

NVidia's usually worked with TSMC in the past, so it's kind of weird to see them shift gears to Samsung. IIRC, it happened earlier this year like this spring or summer, even when TSMC had confirmed absolute Chadly yields by Christmas 2019. https://www.reddit.com/r/AMD_Stock/c...firmed_at_009/

So my speculation, backed by absolutely nothing, is that NVidia felt price pressure to compete with AMD Big Navi, and figured they could get an edge by producing the RTX 3000s on cheaper Samsung wafers whereas AMD would spend more to use TSMC. If they stuck with TSMC like usual, it'd be a push. If anyone in the room had two firing brain cells, they would have brought up the yield differential, so I assume that "yields chitting the bed" was decided to be a tolerable risk.

When I walked away from the industry years ago, we were using 0.25um libraries to design with and 0.18um libraries were just coming online. It's amazing to see that 7nm design rules are in production. That means (roughly and simplistically speaking), transistor count for a given die size has quadrupled every 4 years or so. It seems Moore's Law is still holding true.

[FAPhaggot]

When I walked away from the industry years ago, we were using 0.25um libraries to design with and 0.18um libraries were just coming online. It's amazing to see that 7nm design rules are in production. That means (roughly and simplistically speaking), transistor count for a given die size has quadrupled every 4 years or so. It seems Moore's Law is still holding true.

Believe it or not, 180nm never left brah. It's one of the go-to nodes for pure analog processes to this day. When I interned designing car electronics, all the stuff we did was on 180nm. I interviewed a couple months ago for a job making radiation sensors for spacecraft and they were doing the same thing. There's actually a shortage of fab space around the world right now for manufacturing 180nm chips.

For the non-EE crew: speed, density and power are everything in digital electronics, but what you need for analog is accuracy.

Say I have a computer simulation of a switch that ideally outputs 0V when I turn it off and 1V when I turn it on. But when I build the actual switch, a dozen different real-life imperfections cause it to be at 0V when it's off but only swing up to 0.95V when it's on. For a computer where I only need to read a high or a low, that's absolutely fine. I can still see that 0.95V is supposed to be a high. But say I use this switch to make an audio amplifier, now the top 5% loudest part of my music is going to come out distorted and sounding like chit. That's sort of the core challenge with analog; those little errors really start mattering.

180nm is way obsolete for digital (it's so old it was used in the Playstation 2 CPU), but it found a second life in analog, because every last imperfection got mapped out over the next decade-plus and it was computer modeled accurately down almost to the atom. So when you rig up a switch in your simulation, you can see that it will only output 0.95V when you build it, and then you can correct accordingly.

In automotive, our allowed failed parts per million on a safety-critical part was zero, even with a bottom 0.001% made part, even on a 200,000 mile car driving through the Arctic Circle. We had to model the chip's performance in that chitty Arctic car accurately and make sure it still behaved there, or people are going to die and we get sued. Predictability was everything; much more than size or speed.

Basically with 180nm, all the analog brahs decided that we've got this process, it's small enough to be dense and kind of fast, but big enough to behave more ideally (phuck modeling those tiny waffle fries), big enough to handle a bit of high power without exploding (small devices are really delicate), and it gives uber reliable results between the simulation and the fabricated part. Not to mention, with technology that old, the cost per wafer goes cheap as chit. So they're going to site on that late-90s node, and they're never going to change off it until they either need more speed or you offer a replacement that's just as accurate.

[StainlesSon]

Do you think amd coming out on 28 oct will be on par with the new gtxs?