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.

 

[IntelUser2000]

The 2.5x density gain from GV100 to GA100 is consistent with CPPxMMP comparisons between TSMC's 16nm and their 7nm process. Samsung's "8nm" still represents a significant density gain of 2x over TSMC's 16nm generation.

Yes, GV100 is actually TSMC 12nm but either density gains haven't materialized or GV100 isn't using whatever density gains "12nm" offers over 16nm.

The difference in density between Samsung 8nm and TSMC 7nm is about 25%.

 

[Bouowmx]

GA102 being 627 mm^2 is in line with 1.5x TU102 transistors (1.5 * 18600 M = 27900 M) at 45.2 M/mm^2, which is what the Qualcomm Centriq CPU achieved at Samsung 10 nm.

1.5x transistors has been the progression of the top GeForce GPU since Maxwell.

 

[IntelUser2000]

@Bouowmx That's actually a good observation.

 

[Ajay]

Yes, GA100 is actually TSMC 12nm but either density gains haven't materialized or GA100 isn't using whatever density gains "12nm" offers over 16nm.

I think you didn't write what you meant.

 

[Ajay]

Jenhsen was specifically asked about 7 nm product orders, not about ALL of their next gen products.

And he said SPECIFICALLY that 7 nm products will be handled in majority by TSMC, and in small portion - Samsung.

Nobody has asked Jenhsen about 8 nm from SS.

So in this sense: GA100 - 7 nm TSMC - correct, which we already have seen. Anything below GA100 - 8 nm Samsung.

Hmm, wonder what GPUs will be produced on Samsung 7EUV.

 

[IntelUser2000]

I think you didn't write what you meant.

Thanks. GV100 I mean.

 

[alcoholbob]

The 2.5x density gain from GV100 to GA100 is consistent with CPPxMMP comparisons between TSMC's 16nm and their 7nm process. Samsung's "8nm" still represents a significant density gain of 2x over TSMC's 16nm generation.

Yes, GA100 is actually TSMC 12nm but either density gains haven't materialized or GA100 isn't using whatever density gains "12nm" offers over 16nm.

The difference in density between Samsung 8nm and TSMC 7nm is about 25%.

Well then 505mm2 makes sense for big Navi then. Once you account for density differences GA102 and Navi 20 will basically the same number of transistors.

 

[Glo.]

Well then 505mm2 makes sense for big Navi then. Once you account for density differences GA102 and Navi 20 will basically the same number of transistors.

Assuming AMD hasn't increased the transistor density.

 

[Tup3x]

Assuming that it's going to be Samsung's 8nm process, I wonder how much cheaper it is than TSMC's 7nm process. If Samsung can provide steady supply and the process doesn't hold back the cards too much... I guess it's not going to be a problem.

 

[BenSkywalker]

TSMC revenue taking a big hit, if you don't follow the market at all I can see why that didn't sound bad-

Wei added that TSMC hasn't seen "significant order reduction" from customers, but it would still pay attention to supply chain dislocation and weaker global demand for applications, including consumer electronics and automotive.
For the first quarter, the firm reported revenue increased 42% year on year to NT$310.6 billion (10.3 billion USD) due to high performance computing-related (HPC) and 5G smartphone growth. It predicts "flattish" revenue moving into the second quarter of between $10.1 billion and $10.4 billion.

TSMC cuts growth outlook for 2020

Chipmaker admits it cannot outrun the coronavirus.

www.zdnet.com

They went from up 42% for Q1 to expecting full year to be up low teens. I know it doesn't work exactly like this due to some level of seasonality but +42%, 0%, 0%, 0% would be up 10.5%, low teens ish, for the year. It was a major hit with spare capacity snagged up by the GPU teams.

Perhaps a better way to word it would've been a major uptick in spare capacity.

For the record, if people were claiming RDNA2 was going to be Samsung I'd be pointing out the same things there.

 

[Glo.]

Orders are secured YEARS in advance.

There is no other way for this.

 

[BenSkywalker]

Huawei orders, their second largest customer, had existing orders cleared off the books in a matter of months, article already linked.

Yes, some orders are made years in advance, but I've already posted articles stating the increase in spare capacity, reduction in projected revenue, and increased short lead time orders for nVidia. My apologies I could only link reputable news sources and not anonymous tweets.

 

[Glo.]

You forget about one thing.

Next gen. gaming cards already taped out.

If there is a point in time when "sources" can KNOW full well about what Process those products are - its after the tape out.

There is no other way around it. Nvidia GPUs are made on 8 nm Samsung process.

 

[BenSkywalker]

Missed this point previously.

The two can't be related to one another - you do realise that, right? Nvidia picking up very significant amount of additional N7 wafers would be an effect of not having enough in the first place and planning on dual-sourcing more than normal or alternatively responding to a rapidly growing market that's shown immense growth as of late with additional volume. Like for example, the datacentre market.

Pricing. People complained to no end about how expensive say the 2080Ti was, and yet they seem to be almost all gone at this point. We know that TSMC has been capacity limited for a while now. Prices were set based on what nVidia thought they could sell at that price point. It appears they could've actually gone a bit higher with their top tier, and they went too high with TU106 prices.

Given the competitive land scape Turing faced, maximizing margins made sense, and in any realistic sense selling out of your product the minute before its new and improved model hits the floor is the holy grail of product distribution.

This generation AMD has been saying they have a nVidia killer coming, so getting as many wafers as possible will allow them to price more aggressively and deal with a very different market than they have been facing.

 

[Hitman928]

TSMC revenue taking a big hit, if you don't follow the market at all I can see why that didn't sound bad-

TSMC cuts growth outlook for 2020

Chipmaker admits it cannot outrun the coronavirus.

www.zdnet.com

They went from up 42% for Q1 to expecting full year to be up low teens. I know it doesn't work exactly like this due to some level of seasonality but +42%, 0%, 0%, 0% would be up 10.5%, low teens ish, for the year. It was a major hit with spare capacity snagged up by the GPU teams.

Perhaps a better way to word it would've been a major uptick in spare capacity.

For the record, if people were claiming RDNA2 was going to be Samsung I'd be pointing out the same things there.

Before the pandemic TSMC was predicting ~20% revenue growth for the year. Now they are predicting high single digit to low teens in revenue growth for the year. Yes that's a decent hit in projected growth, but a decrease in projected growth is very different than taking a big hit to revenue.

 

[BenSkywalker]

There is no other way around it. Nvidia GPUs are made on 8 nm Samsung process

I would expect the lowest tier parts to show up on Samsung 8nm based on the available information, although given the state of the supply chain that looks to be a ways off(2 quarters at least I'd say).

You want to believe the guy that was right once, have at it. I'll take the dozens/hundreds of reputable sources that are right every day.

 

[Mopetar]

Orders are secured YEARS in advance.

There is no other way for this.

That's true to a point and part of the reason why there's any discussion of NVidia using Samsung at all, but if a company goes out of business or has to withdraw an order ( or for that matter any number of other reasons. ) a company like TSMC might find themselves with some additional capacity.

It seems like NVidia purchased some wafers from them more recently precisely because they were available. Assuming this is all true we can draw other conclusions from this as well.

 

[Glo.]

I would expect the lowest tier parts to show up on Samsung 8nm based on the available information, although given the state of the supply chain that looks to be a ways off(2 quarters at least I'd say).

You want to believe the guy that was right once, have at it. I'll take the dozens/hundreds of reputable sources that are right every day.

I suggest you check Kopite's track record when it goes for Nvidia GPUs.

He was not correct once. He is correct EVERY SINGLE TIME, when he posts info from his source, and is not his opinion, or speculation.

This is the very reason why even Videocardz quotes him. When he posts something about Nvidia info - there is very high chance it will turn out to be true.

 

[BenSkywalker]

No consumer Ampere cards, no consumer cards this year and they'll be Samsung 8nm.

We know for a fact he was wrong about the first point already, guess we will see on the other two.

 

[Glo.]

No consumer Ampere cards, no consumer cards this year and they'll be Samsung 8nm.

We know for a fact he was wrong about the first point already, guess we will see on the other two.

KEK.

Show any twitt in which he was saying that there won't be any consumer cards, or consumer Ampere this year?

Yep, I just checked, there was no twitt from him on this topic.

 

[BenSkywalker]

Quoting you-

Kopite is best informed when it goes for Nvidia, and he said plainly that there is no such thing as 8 nm Nvidia GPUs, and only Ampere GPU chip is GA100, and gaming cards, which are differenet architecture, have not been taped out yet.

 

[Glo.]

Quoting you-

So you are mistaking him, with me.

Yes, I made a mistake there. I admit. It was not Kopite who was writing it.

 

[BenSkywalker]

I'd have to then ask what he actually did say? Can you link him claiming that there won't be TSMC Ampere consumer cards?

I provided links to my sources.

 

[Dribble]

So you are mistaking him, with me.

Yes, I made a mistake there. I admit. It was not Kopite who was writing it.

You were very sure at the start of the year
"The GPU he is talking about is GA100.

There is no other GPU from Nvidia this year. Period."

You went on and on about him for pages, and on and on and on about what you were 100% sure gonna happen...not surprising people are starting to question you....

 

[jpiniero]

I would expect the lowest tier parts to show up on Samsung 8nm based on the available information, although given the state of the supply chain that looks to be a ways off(2 quarters at least I'd say).

Do you believe it's not on SS8? It does make sense they would do that as I would think it would have been easier to scale up the low end parts rather than port completely.