NVIDIA GeForce "Ampere" GPUs Built on Samsung 8nm Instead of TSMC 7nm?

[btarunr]

NVIDIA's upcoming GeForce "Ampere" family of GPUs will be built almost entirely on Samsung's 8 nanometer silicon fabrication process that's derived from its 10 nm node; rather than TSMC's 7 nm process, according to kopite7kimi, a source with a high hit-rate with NVIDIA rumors in the past. The 8LPP silicon fabrication node by Samsung is an extension of the company's 10LPP (10 nm) node. Both have the same fin pitch, but reductions are made in the areas of gate pitch (down by 6%) resulting in a transistor density of over 61 million/mm². Apparently NVIDIA's entire high-end product stack, including the GA102 silicon that powers at least three high-end consumer SKUs, are expected to be based on Samsung 8LPP.



View at TechPowerUp Main Site

 

[my_name_is_earl]

I hope it's not 10% increase in performance and 300% increase in price. It just works!

 

[TheLostSwede]

I hope it's not 10% increase in performance and 300% increase in price. It just works!

Nah, we're getting 15% for a 250% price increase this time around

 

[Fluffmeister]

Two years later, all we need now is for the competition to turn up.

*yawn*

 

[Ashtr1x]

We already knew that TSMC 7NP and Samsung EUV 8nm nodes were tapped for Nvidia GPUs, this puts more light, GA100 is on 7NP while GA102 are on Samsung node then. No idea how it's going to impact but I do not doubt Nvidia at this point, their Ray Tracing technology is 2 years ahead of the unreleased AMD GPUs and Console HW. Plus their performance in Notebooks (Esp with Slotted PCIe GPU chipsets, MXM) is also unchallenged since a long time, 780M onwards to be precise Kepler, Maxwell, Pascal, Turing, that's 4 generations.

 

[jeremyshaw]

Are you sure about the 61MT/mm^2 figure? That would be pretty close to TSMC's claims for their 7nm HPC variant (66MT/mm^2, iirc; Ampere is at ~65MT/mm^2 and AMD Renoir is estimated at ~63MT/mm^2 as well).

Of course, the real interest is in their electrical characteristics.

Also:

8LPP is not Samsung's most advanced 8nm, that would be 8LPU - Samsung is claiming the latter is their best 8nm for SoCs that required *both* high clocks and high density.

-and-

Are you sure rumormills aren't confusing this with Nvidia's roadmap for their Tegra/Xavier devices? Orin was slated for 8LPP (no direct sources from Nvidia or Samsung, merely "we heard"), and it is an arm A78 + Ampere product (also a few RISC-V data management cores inside).

Given 8LPU was a refinement of 8LPP - specifically for HPC applications - it would make more sense for 8LPU to be used over 8LPP in these rumors. Admittedly, 8LPP was available in 2017, whereas 8LPU took until 2018 to be launch.


Either way, fun to speculate.
EDIT: source of the rumor doesn't even state 8LPP, just Samsung 8nm.

 

[watzupken]

This is in line with previous rumors about Nvidia using Samsung's 10nm for Ampere. True or not, I think we will find out pretty soon.

However if that is really the case, I feel Nvidia may be underestimating competition. Its the story of the tortoise and the hare. While RDNA may not be as efficient as Turing at this point despite being on 7nm, they have been closing that gap moving from Vega 7 (also on 7nm) to RDNA. While there is nothing concrete about RDNA2, I presume there should be further improvements with efficiency (based on what they have been marketing). Interesting to see how the next gen graphic war heating up.

 

[Deleted member 185088]

They can use two nodes for different products, say high end in the better node and worse node for the low end.

 

[cucker tarlson]

I always suspected 3000 is neither ampere nor 7nm tsmc.
we'll see.

 

[prtskg]

This is in line with previous rumors about Nvidia using Samsung's 10nm for Ampere. True or not, I think we will find out pretty soon.

However if that is really the case, I feel Nvidia may be underestimating competition. Its the story of the tortoise and the hare. While RDNA may not be as efficient as Turing at this point despite being on 7nm, they have been closing that gap moving from Vega 7 (also on 7nm) to RDNA. While there is nothing concrete about RDNA2, I presume there should be further improvements with efficiency (based on what they have been marketing). Interesting to see how the next gen graphic war heating up.

If the gap between AMD & Nvidia close, hopefully Intel will be more competitive & we'll have better priced gpu.

 

[BoboOOZ]

This is in line with previous rumors about Nvidia using Samsung's 10nm for Ampere. True or not, I think we will find out pretty soon.

However if that is really the case, I feel Nvidia may be underestimating competition. Its the story of the tortoise and the hare. While RDNA may not be as efficient as Turing at this point despite being on 7nm, they have been closing that gap moving from Vega 7 (also on 7nm) to RDNA. While there is nothing concrete about RDNA2, I presume there should be further improvements with efficiency (based on what they have been marketing). Interesting to see how the next gen graphic war heating up.

From what I can gather, NVidia did underestimate in the past the potential of RDNA2, at least until the information with the specs of the 2 consoles came out. Because of that, they tried to haggle with TSMC over prices, so they used Samsung as an alternative to pressure TSMC. It didn't work out very well and they were forced to reconsider when they saw the boost clocks of the PS5.
They might very well have masks ready for 7nm and 8nm, because what matters a lot is not only the performance/density of a node, for Nvidia yields and capacity are paramount, because they sell a lot of cards...

IMO, the efficiency of RDNA is not so bad, it's just that AMD decided to go with a cheap small die and overclock it as much as possible. in order to increase profit margins.

 

[watzupken]

IMO, the efficiency of RDNA is not so bad, it's just that AMD decided to go with a cheap small die and overclock it as much as possible. in order to increase profit margins.

I don't feel its bad, but considering that this is starting to catch up with Turing which is 12nm (essentially a refined 16nm), they are still at least 1 step behind. Hopefully RDNA2 will close this gap.

Smaller chip is likely because
1. 7nm is not cheap. This is the cutting edge fab slightly more than a year back
2. AMD is introducing a brand new product, RDNA.

They may be using RDNA as a test on 7nm (pipe clearing) before they introduce higher end parts on 7nm.

 

[BoboOOZ]

I don't feel its bad, but considering that this is starting to catch up with Turing which is 12nm (essentially a refined 16nm), they are still at least 1 step behind. Hopefully RDNA2 will close this gap.

Smaller chip is likely because
1. 7nm is not cheap. This is the cutting edge fab slightly more than a year back
2. AMD is introducing a brand new product, RDNA.

They may be using RDNA as a test on 7nm (pipe clearing) before they introduce higher end parts on 7nm.

Well, I guess my wording may be misleading. The 251 sqmm die of the Navi 10 is cheap due to its size and to the fact that being used in 4 sku's with different CU counts and clocks allows using most of the dies, improving yields. It's not a bad strategy, they need to make money, too, if they want to expand and be really able to compete with Nvidia.
As for the cutting edge, AMD/ATI strategy historically was always to go to the most advanced node first, although many times it has penalized them because the process wasn't mature enough and they had to pay the price of additional work to tame the process.

 

[ratirt]

From what I can gather, NVidia did underestimate in the past the potential of RDNA2, at least until the information with the specs of the 2 consoles came out. Because of that, they tried to haggle with TSMC over prices, so they used Samsung as an alternative to pressure TSMC. It didn't work out very well and they were forced to reconsider when they saw the boost clocks of the PS5.
They might very well have masks ready for 7nm and 8nm, because what matters a lot is not only the performance/density of a node, for Nvidia yields and capacity are paramount, because they sell a lot of cards...

IMO, the efficiency of RDNA is not so bad, it's just that AMD decided to go with a cheap small die and overclock it as much as possible. in order to increase profit margins.

I'm not sure what you are saying. How can Nvidia underestimate in the past RDNA2 when it didnt come out yet? How can Nvidia underestimate consoles which are not out yet? Besides, how can you compare consoles to a PC? The frequency is not everything you know. If NV is judging anything is due to RDNA not RDNA2 or upcoming consoles. So far we know these are coming and that is basically all. Especially if you consider the RDNA2 cards.
RDNA is not bad but it is way of a stretch to say anything about RDNA2 at this point. Consoles is one thing but the discrete PC cards is other.

 

[BoboOOZ]

I'm not sure what you are saying. How can Nvidia underestimate in the past RDNA2 when it didnt come out yet? How can Nvidia underestimate consoles which are not out yet? Besides, how can you compare consoles to a PC? The frequency is not everything you know. If NV is judging anything is due to RDNA not RDNA2 or upcoming consoles. So far we know these are coming and that is basically all. Especially if you consider the RDNA2 cards.

All processors and hardware architectures are designed and created several years in advance of the actual launch. When the strategy for each product is chosen, the choices are made based on projections of the market situation and of the competition, it's what I called here "estimation".

If you think a large company like Nvidia bases all their decisions only on proucts that have been launched ayear ago, you are utterly misled. Nvidia's marketing is very competent and extremely proactive, note how they managed to put AMD in trouble with the launch of the 5600XT by launching the EVGA 2060 KO, or how they launched the Super sku's when AMD launched the 5700XT. That's clever marketing and it's always based on having knowledge in advance about what the competition is doing. All significant companies have this, it's just that Nvidia's marketing seems way more competent than AMD's, so when they make a mistake it is much more remarkable.

RDNA is not bad but it is way of a stretch to say anything about RDNA2 at this point. Consoles is one thing but the discrete PC cards is other.

They're essentially the same thing, and an experienced engineer will be most probably capable of estimating the outcome of the translation from APU to a discrete graphic card in terms of TDP, frequencies and efficiency. Now I'm not that engineer, but Nvidia has loads of them.

 

[Ibotibo01]

I thought Nvidia will bring out 10 NM but A100 is 7NM and Ampere architecture. I thought it should be 10NM because manufacturing on silicon has limited density, if TSMC and other companies use silicon , they will use silicon up to 1.4NM. Universities are working on Carbon Nanotubes but it will use 10 years or 15 years later. AMD is bring out new hardware very fast. They must use(because of manufacturing) 1.4NM, 2NM, 3NM, 3NM+, 5NM, 5NM+. AMD will release 7NM+ in this year and AMD passed 10NM. AMD releases new GPUs-CPUs in 1-2 years. If they releases products this speed and passes manufacturing process, AMD will compete to Nvidia or Intel for 10 years(it wouldn't compete for 10+ years). Nvidia and Intel are slow due to Carbon Nanotubes. All in all, it is my guess.

I believe that Nvidia will release 7NM Ampere in this year but maybe they can release it on 8NM.
If we look up transistor densities,

#1	TSMC’s 5nm EUV (RDNA3, ZEN4)	171.3
#2	TSMC’s 7nm+ EUV (RDNA2?, ZEN3)	115.8
#4	Intel’s 10nm	100.8*
#5	TSMC’s 7nm (Mobile)	96.5
#6	Samsung’s 7nm EUV	95.3
#7	TSMC’s 7nm (HPC) (Ampere?) (RDNA1)	66.7
#8	Samsung’s 8nm (Ampere?)	61.2
#9	TSMC’s 10nm	60.3
#10	Samsung’s 10nm	51.8
#11	Intel’s 14nm	43.5
#12	Nvidia's Turing	34-38
#13	GlobalFoundries 12nm	36.7
#14	TSMC’s 12nm	33.8
#15	Samsung / GlobalFoundries 14nm (GCN5.0/GP107)	32.5
#16	TSMC’s 16nm (Pascal)	28.2

NVIDIA’s 12nm FFN process is a customized version of TSMC’s 12nm process, which is better than the average 12nm FFC process.

References:

7nm vs 10nm vs 14nm: Fabrication Process - Tech Centurion

Everything about 7nm, 10nm and 14nm Fabrication Process of TSMC, Intel, and Samsung. Transistor Density & Ranking of different technology nodes.

www.techcenturion.com

TSMC 12nm grabbed NVIDIA Turing from Samsung 10nm

After analyzing the new generation of Turing architecture of NVIDIA, the graphics chip maker NVIDIA recently pointed out that NVIDIA originally chose Samsung‘s 10nm process technology in process technology, but ultimately chose TSMC‘s 12 Nano FFN technology, which is the same process as the...

www.ecdatasheet.com

Intel’s Manufacturing Roadmap from 2019 to 2029: Back Porting, 7nm, 5nm, 3nm, 2nm, and 1.4 nm

www.anandtech.com

 

[HABO]

From what I can gather, NVidia did underestimate in the past the potential of RDNA2, at least until the information with the specs of the 2 consoles came out. Because of that, they tried to haggle with TSMC over prices, so they used Samsung as an alternative to pressure TSMC. It didn't work out very well and they were forced to reconsider when they saw the boost clocks of the PS5.
They might very well have masks ready for 7nm and 8nm, because what matters a lot is not only the performance/density of a node, for Nvidia yields and capacity are paramount, because they sell a lot of cards...

IMO, the efficiency of RDNA is not so bad, it's just that AMD decided to go with a cheap small die and overclock it as much as possible. in order to increase profit margins.

Definitelly, they are using 1,2 volts in their cards, Nvidia 1-1,050. Bad effiviency was just decision how to have good performance with small dies. With 1V and big die it will be another story.

 

[renz496]

From what I can gather, NVidia did underestimate in the past the potential of RDNA2, at least until the information with the specs of the 2 consoles came out. Because of that, they tried to haggle with TSMC over prices, so they used Samsung as an alternative to pressure TSMC. It didn't work out very well and they were forced to reconsider when they saw the boost clocks of the PS5.
They might very well have masks ready for 7nm and 8nm, because what matters a lot is not only the performance/density of a node, for Nvidia yields and capacity are paramount, because they sell a lot of cards...

IMO, the efficiency of RDNA is not so bad, it's just that AMD decided to go with a cheap small die and overclock it as much as possible. in order to increase profit margins.

looking at how nvidia operate for years i don't think nvidia ever underestimate what AMD capable of. also we heard a lot about this story where nvidia try to use samsung to negotiate better price but personally i think it does not really make sense. historically TSMC has always been nvidia primary partner because there is no other foundry out there have the capabilities to do what nvidia want with their silicon. samsung at best only make nvidia GP107 in the past. they don't even get to make mid range nvidia GPU. so what kind of advantage samsung have so nvidia can negotiate better price with TSMC? and in TSMC case if nvidia want to play it the hard way there is always others that will want to fill up the capacity like AMD. and early on we also heard about how nvidia will not going to get any 7nm capacity from TSMC because "it is too late" and all the capacity has been bought by AMD. and yet here we are with A100 being manufactured by TSMC 7nm. in fact nvidia most likely has been building them since late last year or early this year hence A100 being delivered to several nvidia customer much earlier before the official unveil in may.

 

[BoboOOZ]

looking at how nvidia operate for years i don't think nvidia ever underestimate what AMD capable of. also we heard a lot about this story where nvidia try to use samsung to negotiate better price but personally i think it does not really make sense. historically TSMC has always been nvidia primary partner because there is no other foundry out there have the capabilities to do what nvidia want with their silicon. samsung at best only make nvidia GP107 in the past. they don't even get to make mid range nvidia GPU. so what kind of advantage samsung have so nvidia can negotiate better price with TSMC?

Loweer price, and , as a consequence, better price/performance ratio.

and in TSMC case if nvidia want to play it the hard way there is always others that will want to fill up the capacity like AMD. and early on we also heard about how nvidia will not going to get any 7nm capacity from TSMC because "it is too late" and all the capacity has been bought by AMD. and yet here we are with A100 being manufactured by TSMC 7nm. in fact nvidia most likely has been building them since late last year or early this year hence A100 being delivered to several nvidia customer much earlier before the official unveil in may.

Nvidia had booked some volume at TSMC, but not a lot. You maybe don't realize, but NVidia is a much larger company and, hence, a much larger client than AMD, so they buy larger volumes. Anyway, Nvidia tried to pressure TSMC but lost, so 2 months ago they booked all the capacity that they could:

Nvidia, AMD book out spare TSMC 7nm production lines for next-gen GPUs

Nvidia and AMD have picked up any spare capacity on TSMC’s 7nm foundry as the companies ready their next generational GPUs. The report comes from Taiwanese rag Digitimes that says the new orders will help to ensure TSMC posts relatively strong results compared with its competition.

www.notebookcheck.net

 

[ratirt]

All processors and hardware architectures are designed and created several years in advance of the actual launch. When the strategy for each product is chosen, the choices are made based on projections of the market situation and of the competition, it's what I called here "estimation".

If you think a large company like Nvidia bases all their decisions only on proucts that have been launched ayear ago, you are utterly misled. Nvidia's marketing is very competent and extremely proactive, note how they managed to put AMD in trouble with the launch of the 5600XT by launching the EVGA 2060 KO, or how they launched the Super sku's when AMD launched the 5700XT. That's clever marketing and it's always based on having knowledge in advance about what the competition is doing. All significant companies have this, it's just that Nvidia's marketing seems way more competent than AMD's, so when they make a mistake it is much more remarkable.

NV doesn't know what AMD will release. How they can know that? they may predict something but this does not include NV knowing what AMD's RDNA2 will be capable of.
NV didnt put AMD in trouble by releasing 2060 KO but it was NV response to the 5600XT's performance after it had been released.
The Super skus on the other hand where released to counter AMD's products and then AMD has tweaked the BIOS for those by giving it to AIB's
All of this doesn't mean NV had known about AMD's products and vice versa.
Do you think AMD knows what NV new graphics are gonna be like? I doubt it.

They're essentially the same thing, and an experienced engineer will be most probably capable of estimating the outcome of the translation from APU to a discrete graphic card in terms of TDP, frequencies and efficiency. Now I'm not that engineer, but Nvidia has loads of them.

Consoles are not the same thing as PC's even though they share kinda same counterparts. These are dedicated for gaming while PC's are not just for gaming and that is the difference. Consoles don't need that much horse power. Consoles are dedicated for games and what NV can get out of those is that they will have the capability of ray tracing but it won't tell them much where the PC's discrete cards RDNA2 will end up. NV will prepare for various scenarios but it isn't like they know exactly what RDNA2 will bring in terms of performance. Unless they listen to AMD's keynotes about it but as you already know, we need to wait for the reviews and actual benchmarks to judge the performance.
Prediction may be correct but also may be way off.

 

[BoboOOZ]

NV doesn't know what AMD will release. How they can know that? they may predict something but this does not include NV knowing what AMD's RDNA2 will be capable of.
NV didnt put AMD in trouble by releasing 2060 KO but it was NV response to the 5600XT's performance after it had been released.
The Super skus on the other hand where released to counter AMD's products and then AMD has tweaked the BIOS for those by giving it to AIB's
All of this doesn't mean NV had known about AMD's products and vice versa.
Do you think AMD knows what NV new graphics are gonna be like? I doubt it.

I think AMD know less about Nvidia than Nvidia knows about AMD? How? Opens source information analysis plus other means, known as orporate intelligence.
AMD is well-known for being very "leaky" :
https://adoredtv.com/overvolted-9-amds-leaky-ship/

Consoles are not the same thing as PC's even though they share kinda same counterparts. These are dedicated for gaming while PC's are not just for gaming and that is the difference. Consoles don't need that much horse power. Consoles are dedicated for games and what NV can get out of those is that they will have the capability of ray tracing but it won't tell them much where the PC's discrete cards RDNA2 will end up. NV will prepare for various scenarios but it isn't like they know exactly what RDNA2 will bring in terms of performance. Unless they listen to AMD's keynotes about it but as you already know, we need to wait for the reviews and actual benchmarks to judge the performance.
Prediction may be correct but also may be way off.

You're beating a dead strawman here, I agree with you consoles are different animals than PC, but I wasn't talking about that.
What I said is that a good engineer can extract a lot of information about how a hardware architecture will perform as a discrete graphics card, knowing what clock frequencies it can hit as an APU in a TDP constrained console. What matters here is not how many FPS will have Big Navi in Red Dean Redemption 2. What Nvidia needs to know is only how many TFlops and what memory bandwidth will it have, approximately. I hope it's clearer now.

 

[ratirt]

You're beating a dead strawman here, I agree with you consoles are different animals than PC, but I wasn't talking about that.
What I said is that a good engineer can extract a lot of information about how a hardware architecture will perform as a discrete graphics card, knowing what clock frequencies it can hit as an APU in a TDP constrained console. What matters here is not how many FPS will have Big Navi in Red Dean Redemption 2. What Nvidia needs to know is only how many TFlops and what memory bandwidth will it have, approximately. I hope it's clearer now.

Good engineers focus on the product their company wants to release, to mitigate any inconsistency or problems they may or may not face during design or other, not to try extracting any possible performance measures from a competitor's product that has not been released knowing fraction of what it will be capable of.

I think AMD know less about Nvidia than Nvidia knows about AMD? How? Opens source information analysis plus other means, known as orporate intelligence.
AMD is well-known for being very "leaky" :
https://adoredtv.com/overvolted-9-amds-leaky-ship/

It is leaks and as a prediction it may be or may not be true. I'm sure NV realizes that and instead, NV is trying to make its new card as good as possible and adjust to the market when AMD and NV release the cards and start competing. BTW there are leaks about NV as well and I'm sure AMD is not focusing on these leaks instead on their product release.

 

[Wilson]

I hope it's not 10% increase in performance and 300% increase in price. It just works!

It's about 30-35% increase in performance peak (Raw, excluding RT). It's up to them how to cut their chip and what to sell those pieces for

 

[BoboOOZ]

Good engineers focus on the product their company wants to release, to mitigate any inconsistency or problems they may or may not face during design or other, not to try extracting any possible performance measures from a competitor's product that has not been released knowing fraction of what it will be capable of.

It is leaks and as a prediction it may be or may not be true. I'm sure NV realizes that and instead, NV is trying to make its new card as good as possible and adjust to the market when AMD and NV release the cards and start competing. BTW there are leaks about NV as well and I'm sure AMD is not focusing on these leaks instead on their product release.

This will be my last post on this subject, because I think we discussed it enough.
But basically I think you idealize what is competition in a free market. Companies do not try to make the best product that they can make. Companies try to make the most money they can make. They do that usually by doing the cheapest product possible but which is still better than the competition, or at least competitive with the price of the competition.
Or, as they say, when you're chased by the bear, you don't have to be faster than the bear, you just have to be faster than your friend .

 

[yeeeeman]

This has been discussed a million times now. Higher end GPUs will be made on 7nm and lower end on 8nm.
As for product stacking, they usually do their best within the limits they have and then, closer to launch they see what the competitor is cooking and try to outmatch it. In this case, nvidia thought a bog standard GA104 will do just fine with rdna2, but they found out that AMD has much better performance than anticipated. So they moved 3080 series from ga104 to ga102 which is made on 7nm and that was meant for titan card.