Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[reb0rn]

If they push big gains in RTX mostly they are insane!
I don`t want to see more then 10% die wasted on that crap

 

[xpea]

Kopite claims that GA102 hasn't taped out yet, and it will tape out in March/April .

His track record when it goes to Nvidia info is a little bit better than yours so you better back it up with anything.

trust me or not, I don't care but my track record is 100 times better than these twittos who change opinion every week or so... Just go back in time, I gave exclusive leaks that everyone copied and that were proven to be true...

 

[xpea]

It taped out last year on Samsung's 7nm, which we now know didn't work right, and forced nVidia to switch over to TSMC.

Not true and I'm 100% sure. I have no idea who can believe this sh*t.
Imagine for a second, Nvidia spent 300 millions dollars and 6 months on a tape out and then cancel their plan when everything is done.. yeah... sure... if anyone have a little bit of idea how Nvidia or any major fabless company qualify their manufacturing process and perform conformance analysis on pre silicon before tape-out, they know how this information is stupid...

Edit: I have no time to find a better video but this one gives an small idea on how Nvidia works on silicon failure analysis (keep in mind, it was 4 years ago, now the tools are much more powerful)

 

[Glo.]

trust me or not, I don't care but my track record is 100 times better than these twittos who change opinion every week or so... Just go back in time, I gave exclusive leaks that everyone copied and that were proven to be true...

Your track record is not 100 times better than Kopite's .

 

[xpea]

Your track record is not 100 times better than Kopite's .

nobody cares. Find something I said that became wrong and send me in private. Let's keep this topic clean.

 

[Glo.]

nobody cares. Find something I said that became wrong and send me in private. Let's keep this topic clean.

Its not about your track record, but his(Kopite's) track record. When he leaks something about Nvidia, and I mean leaks, not speculates, it is(at least, always was) 100% accurate. So I won't disregard his words, that Nvidia Gaming GPUs hasn't taped yet, just because you say so.

 

[xpea]

Its not about your track record, but his(Kopite's) track record. When he leaks something about Nvidia, and I mean leaks, not speculates, it is(at least, always was) 100% accurate. So I won't disregard his words, that Nvidia Gaming GPUs hasn't taped yet, just because you say so.

I have my (rock solid) sources and 2 gaming dies have been tapeout at TSMC last year (it was the plan all along, this samsung nonsens is a proof that Kopite source is not first hand)

 

[Dribble]

Only 40% faster than RTX 2080 Ti?

That's laughable...

For this very reason I very much doubt in this rumor, because it means that Nvidia genuinely ******d up.

Why? The 2080Ti die is huge, they will have used a smaller one for the 3080Ti. In addition they are thinking about the 4080Ti. 40% is actually more then they need to get people to buy the new card which is Nvidia's primary aim. Nvidia are therefore better off saving that extra performance so they can make the 4080Ti (probably stuck on the same process) more impressive. No good overdoing the 3080Ti (unless you really need too due to AMD) if it effects your ability to sell the 4080Ti later.

 

[uzzi38]

Why? The 2080Ti die is huge, they will have used a smaller one for the 3080Ti. In addition they are thinking about the 4080Ti. 40% is actually more then they need to get people to buy the new card which is Nvidia's primary aim. Nvidia are therefore better off saving that extra performance so they can make the 4080Ti (probably stuck on the same process) more impressive. No good overdoing the 3080Ti (unless you really need too due to AMD) if it effects your ability to sell the 4080Ti later.

If Nvidia have only aimed for a 40% uplift then they themselves have opened the door for AMD to compete on the very high end through nothing but negligence.

Not beat. But definitely compete. And it only gets worse with AMD iterating more quickly generation on generation than Nvidia for the time being.

I don't think Nvidia are so incompetant, especially now that they inadvertantly have gotten some extra time to correct that mistake. Something I guess they can thank Sammy for.

 

[Stuka87]

Not true and I'm 100% sure. I have no idea who can believe this sh*t.
Imagine for a second, Nvidia spent 300 millions dollars and 6 months on a tape out and then cancel their plan when everything is done.. yeah... sure... if anyone have a little bit of idea how Nvidia or any major fabless company qualify their manufacturing process and perform conformance analysis on pre silicon before tape-out, they know how this information is stupid...

nVidia said they were going to use Samsung 7nm last July. It wasn't until December that nVidia stated they would be using TSMC for 7nm products.

 

[DXDiag]

Not true and I'm 100% sure. I have no idea who can believe this sh*t.
Imagine for a second, Nvidia spent 300 millions dollars and 6 months on a tape out and then cancel their plan when everything is done.. yeah... sure... if anyone have a little bit of idea how Nvidia or any major fabless company qualify their manufacturing process and perform conformance analysis on pre silicon before tape-out, they know how this information is stupid...

Yes I don't believe this nonsense either, NVIDIA collaborated heavily with TSMC on 12nm process, they wouldn't simply dumb them on 7nm and go for a different foundry. Not when TSMC is the leading 7nm manufacturer.

 

[Hitman928]

nVidia said they were going to use Samsung 7nm last July. It wasn't until December that nVidia stated they would be using TSMC for 7nm products.

Here's a link from an announcement at the time.

Big win for Samsung as NVIDIA picks its 7nm process for next-gen GPU

Samsung has made no secret of the fact that it wants to offset the decline in its memory business with ...

www.sammobile.com

The move has done the trick. NVIDIA Korea chief Yoo Eung-joon confirmed during a press conference today that production of the company’s next-gen GPU architecture will be carried out by Samsung on its 7nm EUV process. “It is meaningful that Samsung Electronics’ 7-nanometer process would be used in manufacturing our next-generation GPU,” Yoo said.
Yoo didn’t say just how much of the foundry production will be handled by Samsung, only suggesting that it would be “substantial.” This isn’t the first time that NVIDIA is relying on Samsung’s manufacturing. It has done that for the GTX 1050 and GTX 1050 Ti cards so this isn’t an unprecedented development. Ampere-based NVIDIA graphic cards are set to hit the market next year.
NVIDIA’s decision might also be based on supply considerations. TSMC’s 7nm node has already seen very high demand from the likes of Apple and AMD. Opting for Samsung’s 7nm process may enable NVIDIA to have more supply so that it can effectively meet demand.

However, another report said that Nvidia was staying with TSMC. Although supposedly they are saying Nvidia made an announcement saying they are sticking with TSMC but I couldn't find that announcement anywhere, only the one that they would use Samsung 7 nm for a substantial amount of their 7 nm designs.

Taiwan Semiconductor Manufacturing Company victim of fake news | Taiwan News | 2019-07-08 11:15:00

NVIDIA deal to produce next-generation GPUs still on.NVIDIA has refuted a claim by a Korean publication that NVIDIA had chosen Samsung over TSMC to produce its 7-nm GPUs. | 2019-07-08 11:15:00

www.taiwannews.com.tw

Both could be true in that the original plan was to use both Samsung and TSMC and maybe Samsung was to handle the big dies due to wafer availability.

 

[Stuka87]

Yes I don't believe this nonsense either, NVIDIA collaborated heavily with TSMC on 12nm process, they wouldn't simply dumb them on 7nm and go for a different foundry. Not when TSMC is the leading 7nm manufacturer.

Has nothing to do with "dumping TSMC". The fact of the matter is that nVidia is a Tier 3 customer. Meaning they can only get the capacity that tier 1 and 2 companies don't use. Tier 1 has companies like Apple, and Tier 2 has companies like AMD (Which has far more chips made than nVidia due to AMD also having CPU and APU production on 7nm). When Apple moved to purchasing 5nm capacity, AMD in turned bought up all of the capacity that Apple let go for 7nm (Most likely for Consoles).

This is thought of as being the reason nVidia had intended to go with Samsung, because Samsung had a lot more capacity available.

 

[Stuka87]

However, another report said that Nvidia was staying with TSMC. Although supposedly they are saying Nvidia made an announcement saying they are sticking with TSMC but I couldn't find that announcement anywhere, only the one that they would use Samsung 7 nm for a substantial amount of their 7 nm designs.

Both could be true in that the original plan was to use both Samsung and TSMC and maybe Samsung was to handle the big dies due to wafer availability.

Yeah, its kind of muddy. Ultimately I think we will end up seeing their next chips on TSMC, with a possible exception being chips for Nintendo, which may end up on Samsung. If Samsung is having issues, its most likely with larger chips (or has been historically anyway). So the smaller SoC that the Switch uses would be a better fit for them.

 

[NostaSeronx]

Samsung 8LPU, anytime after Q1 2020. (GA) (Essentially, Turing w/ huge power efficiency gain and HPC ULVT > 2GHz?)
TSMC N5, anytime after Q4 2020. (GH) (Re-uses the above CUDA version, but is a from-scratch post-CUDA architecture)

Just throwing this here.

 

[Glo.]

This is thought of as being the reason nVidia had intended to go with Samsung, because Samsung had a lot more capacity available.

Well, considering that almost nobody would manufactuore anything on Samsung's process, that is anywhere near the performance, die size and volume that Nvidia wanted to acghieve...

Unluckily for Nvidia, Samsung's process is a turd.

 

[DXDiag]

Has nothing to do with "dumping TSMC". The fact of the matter is that nVidia is a Tier 3 customer. Meaning they can only get the capacity that tier 1 and 2 companies don't use. Tier 1 has companies like Apple, and Tier 2 has companies like AMD (Which has far more chips made than nVidia due to AMD also having CPU and APU production on 7nm). When Apple moved to purchasing 5nm capacity, AMD in turned bought up all of the capacity that Apple let go for 7nm (Most likely for Consoles).

Source for NVIDIA being Tier 3 customer? NVIDIA has been with TSMC far longer than AMD ever did, and contrary to your statement, NVIDIA does produce large shipments, selling units (dGPUs, iGPUs, Data Center GPUs .. etc) that rival AMD's CPUs and GPUs combined (before Ryzen). So I don't believe for a second that TSMC would demote NVIDIA to tier 2 at all.

And NVIDIA is not a naive company that wings it's process node placements on the fly, they pre-book it several years in advance, (especially with a close partner like TSMC). AMD will not get to buy all of Apple's share unless NVIDIA allows it by declining to buy any share for themselves, which is unlikely to ever happen. The fact that we are seeing leaks for big Ampere chips confirms that as well, NVIDIA is prioritizing their TSMC shares to produce their large Data Center GPUs, after which they will produce their consumer chips.

So I strongly believe that story to be not true at all.

 

[Glo.]

Source for NVIDIA being Tier 3 customer? NVIDIA has been with TSMC far longer than AMD ever did, and contrary to your statement, NVIDIA does produce large shipments, selling units (dGPUs, iGPUs, Data Center GPUs .. etc) that rival AMD's CPUs and GPUs combined (before Ryzen). So I don't believe for a second that TSMC would demote NVIDIA to tier 2 at all.

And NVIDIA is not a naive company that wings it's process node placements on the fly, they pre-book it several years in advance, (especially with a close partner like TSMC). AMD will not get to buy all of Apple's share unless NVIDIA allows it by declining to buy any share for themselves, which is unlikely to ever happen. The fact that we are seeing leaks for big Ampere chips confirms that as well, NVIDIA is prioritizing their TSMC shares to produce their large Data Center GPUs, after which they will produce their consumer chips.

So I strongly believe that story to be not true at all.

Fab capacity usually happens with auction. The faster and higher bid for larger capacity - the more chance you will be picked as a the client for manufacturing of your products.

Which is why Apple can so easily get TSMC's node so early. They have heaps of cash to put into TSMC. AMD is a long client of TSMC, and they manufacture critical chips for logistical reasons(CPUs). TSMC will put priority for AMD, because of this reason.

Because AMD manufactures CPUs and GPUs Nvidia, AMD can afford buyout on heaps of capacity in advance, which is impossible for Nvidia, to book large amounts of wafers, like AMD can.

From sheer manufacturing volume, when you take into account what AMD actually makes with TSMC, AMD dwarfes Nvidia manufacturing. Nvidia is GPU only company.

CPUs, GPUs, APUs, Console SoC's. Thats what AMD will make on 7 nm TSMC's processes. Its Nvidia who will get whatever wafers are left in TSMC fabs, from capacity point of view.

Which was the whole reason why they went with Samsung. They believed Samsung will have performance, yield and capacity to fulfill Nvidia.

 

[Stuka87]

Source for NVIDIA being Tier 3 customer? NVIDIA has been with TSMC far longer than AMD ever did, and contrary to your statement, NVIDIA does produce large shipments, selling units (dGPUs, iGPUs, Data Center GPUs .. etc) that rival AMD's CPUs and GPUs combined (before Ryzen). So I don't believe for a second that TSMC would demote NVIDIA to tier 2 at all.

And NVIDIA is not a naive company that wings it's process node placements on the fly, they pre-book it several years in advance, (especially with a close partner like TSMC). AMD will not get to buy all of Apple's share unless NVIDIA allows it by declining to buy any share for themselves, which is unlikely to ever happen. The fact that we are seeing leaks for big Ampere chips confirms that as well, NVIDIA is prioritizing their TSMC shares to produce their large Data Center GPUs, after which they will produce their consumer chips.

So I strongly believe that story to be not true at all.

I don't normally quote wccftech, but they have a translated version of the article: https://wccftech.com/amd-7nm-wafer-...-7nm-capacity-at-tsmc-currently-fully-booked/

TSMC's 7nm production capacity continues to rise. The industry expects monthly capacity to reach 110,000 wafers in 1H'2020. The top 5 customers by order proportion are: Apple, HiSilicon, Qualcomm, AMD, and Mediatek. Except for Mediatek, order share is split at roughly 20% each, depending on seasonality. Mediatek's share is around 13%.

In 2H, Apple is moving over to 5nm. AMD in turn purchased Apple's share of the 7nm capacity, which equates to 30k wafers. AMD was able to do this because they know they have the demand for that. With Ryzen continuing to gain ground, and both the PS5 and XBox Series-X coming out, they will easily consume all of that capacity.

nVidia makes a lot of profit off of what they make, but in terms of volume of chips produced, they are much smaller than the top five listed above.

 

[extide]

NVIDIA has been with TSMC far longer than AMD ever did,

Not necessarily. Don't forget that AMD bought ATi who has been building GPU's at TSMC since the 90's, just like Nvidia. The fact that AMD buys more wafers is also going to give them a bit more weight to throw around. Not a ton, but some.

 

[lobz]

Source for NVIDIA being Tier 3 customer? NVIDIA has been with TSMC far longer than AMD ever did, and contrary to your statement, NVIDIA does produce large shipments, selling

Source for your statement in bold?

 

[lobz]

In other news:
"(...)in light of the spread of the coronavirus, it is (NVIDIA) deferring plans to deliver a webcast keynote(...)"

Can someone please explain to me, what in the frozen hell does a biological virus have to do with a webcast?

 

[jpiniero]

In other news:
"(...)in light of the spread of the coronavirus, it is (NVIDIA) deferring plans to deliver a webcast keynote(...)"

Can someone please explain to me, what in the frozen hell does a biological virus have to do with a webcast?

What are they going to do, broadcast it from Jensen's bedroom?

 

[xpea]

nVidia said they were going to use Samsung 7nm last July. It wasn't until December that nVidia stated they would be using TSMC for 7nm products.

The plan never changed and it's the same from some generations already. TSMC gets the big dies and the major chunk of the business. Samsung the small dies and less business. That's it. Nothing more, nothing less. Like Nvidia did with Pascal, like they are doing with Turing.
The Samsung talk was a PR move to tell the world they won a high profile customer on their most advanced node. Of course they are making it as big as possible. It's marketing 101 and these brainless tweettos distort the reality for sensationalism in a clickbait news...

 

[tajoh111]

Fab capacity usually happens with auction. The faster and higher bid for larger capacity - the more chance you will be picked as a the client for manufacturing of your products.

Which is why Apple can so easily get TSMC's node so early. They have heaps of cash to put into TSMC. AMD is a long client of TSMC, and they manufacture critical chips for logistical reasons(CPUs). TSMC will put priority for AMD, because of this reason.

Because AMD manufactures CPUs and GPUs Nvidia, AMD can afford buyout on heaps of capacity in advance, which is impossible for Nvidia, to book large amounts of wafers, like AMD can.

From sheer manufacturing volume, when you take into account what AMD actually makes with TSMC, AMD dwarfes Nvidia manufacturing. Nvidia is GPU only company.

CPUs, GPUs, APUs, Console SoC's. Thats what AMD will make on 7 nm TSMC's processes. Its Nvidia who will get whatever wafers are left in TSMC fabs, from capacity point of view.

Which was the whole reason why they went with Samsung. They believed Samsung will have performance, yield and capacity to fulfill Nvidia.

Are you sure of that?

Look at costs of revenue which is typically manufacturing costs and you will see AMD and Nvidia are comparable.

Add in that AMD relies more heavily on Global foundaries compared to Nvidia on Samsung and Nvidia could be a bigger spender at TSMC.

One of the possible reasons Nvidia went with 12nm vs 7nm finfet was in September of 2018, 7nm finfet was horrific and the available supply of wafers was only sufficient for low volume products like Vega 20.

Considering AMD is still having supply issues with 7nm, TSMC does not have enough for 7nm even now. As a result, Nvidia likely calculated that the benefit of an early launch with good supply outweighed the process advantage of 7nm given their volume of their launches. If Nvidia launched turing on 7nm, the limited supply availability would hurt them, particularly when the availability of 7nm was bad already at the end of 2019. Just imagine how bad it would have been in september 2018.