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.

 

[Hitman928]

Source? I cannot find anywhere where they say it has been shipping for 2+ months.

They said they sent some to a lab to do COVID-19 research but we have no idea how many were sent. It could have been a pre-production sample type volume. I haven't watched the videos yet but I've skimmed a couple of the articles and haven't seen a date yet of when GA100 will be shipping to partners or when partners will have solutions ready. Maybe someone else has seen those details.

 

[DXDiag]

Source? I cannot find anywhere where they say it has been shipping for 2+ months.

NVIDIA today announced that the first GPU based on the NVIDIA Ampere architecture, the NVIDIA A100, is in full production and shipping to customers worldwide.

NVIDIA's New Ampere Data Center GPU in Full Production

NVIDIA today announced that the first GPU based on the NVIDIA Ampere architecture, the NVIDIA A100, is in full production and shipping to customers worldwide. The A100 draws on design breakthroughs in the NVIDIA Ampere architecture — offering the company's largest leap in performance to date...

www.techpowerup.com

Other early adopters include national laboratories and some of the world's leading higher education and research institutions, each using A100 to power their next-generation supercomputers. They include:

• Indiana University, in the U.S., whose Big Red 200 supercomputer is based on HPE's Cray Shasta system, will support scientific and medical research, and advanced research in AI, machine learning and data analytics.
• Jülich Supercomputing Centre, in Germany, whose JUWELS booster system being built by Atos is designed for extreme computing power and AI tasks.
• Karlsruhe Institute of Technology, in Germany, which is building its HoreKa supercomputer with Lenovo, will be able to carry out significantly larger multi-scale simulations in the field of materials sciences, earth system sciences, engineering for energy and mobility research, and particle and astroparticle physics.
• Max Planck Computing and Data Facility, in Germany, with its next-generation supercomputer Raven built by Lenovo, provides high-level support for the development, optimization, analysis and visualization of high-performance-computing applications to Max Planck Institutes.
• The U.S. Department of Energy's National Energy Research Scientific Computing Center, located at Lawrence Berkeley National Laboratory, which is building its next-generation supercomputer Perlmutter based on HPE's Cray Shasta system to support extreme-scale science and develop new energy sources, improve energy efficiency and discover new materials.

 

[sontin]

And nVidia has upgraded their own Saturn-V with 560 DXG systems (total of 4,480 A100 GPUs) to 4,8 Exaflops.

 

[maddie]

Making claims and ignoring calls for some evidence. Over and over again.

 

[Hitman928]

Nvidia reports earnings 1 week from today. If they shipped Ampere for revenue in Q1 it should come to light then. GA100 is quite a beast. I'm extremely impressed by how big they were able to make this die on 7 nm. I wonder what their yields are for that thing, even with some clusters disabled in each die. They definitely went all out to assert their HPC dominance in the current market.

 

[Ajay]

Source? I cannot find anywhere where they say it has been shipping for 2+ months.

From AT Front Page: Installation at Argonne National Labs. Not sure of the date, but they are already there.

 

[Ajay]

Apparently, Gaming Ampere will be based off A102 cards (already taped out) and not A100 ones. So potentially all off Samsung (84SM | 5376 Cuda Cores | 384 bus width)

Huh... GA102 taped out? Where is that info from?

IDK, maybe he's tuning in info using special antennae. In all likelihood, the consumer code name isn't even Ampere (like Votla/Pascal). I was hoping NV would dribble out a bit of info for us poor enthusiasts, but, alas, they are leaving us dangling . Guess they really are waiting for AMD RDNA2 GPUs to drop.

 

[Hitman928]

I think it's coming from this:

https://twitter.com/x/status/1237687606869188608

 

[Mopetar]

Nvidia reports earnings 1 week from today. If they shipped Ampere for revenue in Q1 it should come to light then. GA100 is quite a beast. I'm extremely impressed by how big they were able to make this die on 7 nm. I wonder what their yields are for that thing, even with some clusters disabled in each die. They definitely went all out to assert their HPC dominance in the current market.

The yields are probably pretty good, even for such a massive die, considering 7nm has been out for a while now. I forget exactly where I read it, but some months ago AMD had made some statements regarding their own yields (% of good dies) for Zen 2 that led to an estimated defect density no more than .05 defects per square centimeter for the 7nm process.

If you plug that into a wafer calculator with a 30 cm. x 30 cm. die size (it's not exact, but it's close enough without having exact measurements) it comes out to ~66% good dies. If they're disabling a few clusters so they can salvage otherwise defective dies and ensure a particular performance level for all chips then it isn't hard to believe that they're probably getting at least 90% of chips making the cut.

Now while all that sounds impressive it is important to keep in mind that they can only fit about 50 of these onto a wafer because they're so large, but for what they'll be able to charge for one of these I'm not sure they're too worried.

 

[uzzi38]

I wonder if his configuration for GA102 is true and what kind of performance uplift they'd get compared to the 2080ti if it were.

He stated afterwards that his configurations for all of the gaming cards are wrong. He's one of the people that have been talking about the 8LPP retape since January.

 

[Hitman928]

The yields are probably pretty good, even for such a massive die, considering 7nm has been out for a while now. I forget exactly where I read it, but some months ago AMD had made some statements regarding their own yields (% of good dies) for Zen 2 that led to an estimated defect density no more than .05 defects per square centimeter for the 7nm process.

If you plug that into a wafer calculator with a 30 cm. x 30 cm. die size (it's not exact, but it's close enough without having exact measurements) it comes out to ~66% good dies. If they're disabling a few clusters so they can salvage otherwise defective dies and ensure a particular performance level for all chips then it isn't hard to believe that they're probably getting at least 90% of chips making the cut.

Now while all that sounds impressive it is important to keep in mind that they can only fit about 50 of these onto a wafer because they're so large, but for what they'll be able to charge for one of these I'm not sure they're too worried.

I think you mean 30 mm x 30 mm , but besides that I'd be very surprised if they were getting 90% yield even with the disabled clusters. As an equivalent, you'd need a 15 mm x 15 mm sized fully enabled chip (~ 1/4 the size) to hit that high of a yield.

Edit: These are all rough numbers of course.

 

[Konan]

IDK, maybe he's tuning in info using special antennae. In all likelihood, the consumer code name isn't even Ampere (like Votla/Pascal). I was hoping NV would dribble out a bit of info for us poor enthusiasts, but, alas, they are leaving us dangling . Guess they really are waiting for AMD RDNA2 GPUs to drop.

I think it's coming from this:

https://twitter.com/x/status/1237687606869188608

That is correct. Partly why I said "apparently". As Uzzi said a few posts back this guy has a record of over a year at being pretty accurate and is a great person to follow.
Food for thought - he is a bit at odds with the Moore's Law videos so will be interesting to see what comes to fruition.

 

[Ajay]

That is correct. Partly why I said "apparently". As Uzzi said a few posts back this guy has a record of over a year at being pretty accurate and is a great person to follow.
Food for thought - he is a bit at odds with the Moore's Law videos so will be interesting to see what comes to fruition.

Okay, my apologies, that guy seems to have better info than most. Has he made any updated posts since NV's announcement about N7 products. I just cannot figure out what's really going on here as I've seen some conflicting rumors.

 

[Konan]

GA100 comes in late(hard launch, physical release) Q3 2020, and this is the first Nvidia next gen GPU.

Can you point me to post in which I specifically said that A100 is not in production?

I think this is what some of the guys are referring to.
GA100 shown today is in full production and has shipped out according to press releases. As above Glo. you said later in the year so some confusion; maybe differences between Fiscal and Calendar quarters? Either way I think that is what they are getting at.

 

[Konan]

Okay, my apologies, that guy seems to have better info than most. Has he made any updated posts since NV's announcement about N7 products. I just cannot figure out what's really going on here as I've seen some conflicting rumors.

Nothing to apologize for Ajay. All good buddy.

He has said -
GA104 has been taped out. GA103 may be cancelled.
Moore's law video is fake - mainly as no USBc
Regarding the 18Gbps GDDR6 speeds he said could be GDDR6x maybe

Also said "Maybe, RTX 3090 will have 4992 CUDA Cores(GA102 - 3TPC = 39TPC). (not full fat)"
^ this was in relation to kittycorgi saying - "GA102 12GB G6 at 18Gbps" who also isn't sure if it is 8nm Samsung or GDDR6x.
I have reservations about those speeds though - For what it's worth, none of the memory vendors are talking about 18 Gbps GDDR6, or at least not publicly

So basically keep an eye on these two -

https://twitter.com/KkatCorgi

https://twitter.com/kopite7kimi

It looks like Moore's Law is Dead (Ampere Youtube video guy) has pulled bits of what they have said and then layed on his own "source" to come up with those videos.

Other interesting people to follow -

https://twitter.com/_rogame

https://twitter.com/chiakokhua

https://twitter.com/peresuslog

 

[Ajay]

GA104 has been taped out. GA103 may be cancelled.

Taped out at Samsung, or TSMC? This is the thing getting my knickers in a twist.

 

[Glo.]

I quoted the post. Here is another post claiming A100 will only be available in Q3+: https://forums.anandtech.com/thread...arch-speculation-thread.2572510/post-40155409

A100 is in production since months and has been shipped to customers in the last weeks.

You can tell simply that you are a liar or manipulator. It will be easier.

Nobody ever claimed that GA100 was not in production, you made that up, or twisted words for your own purposes. Its trolling and against forum rules.


Insulting other users is not allowed.

AT Mod Usandthem

 

[Glo.]

I think this is what some of the guys are referring to.
GA100 shown today is in full production and has shipped out according to press releases. As above Glo. you said later in the year so some confusion; maybe differences between Fiscal and Calendar quarters? Either way I think that is what they are getting at.

Nobody was getting at anything. If you would pay attention to Sontin's posts, he usually twists people's words to fit his own agenda, and manipulation. He blatantly suggested that "somebody" here suggested that GA100 was not in production. Which NOBODY EVER did. When confronted, he refuted with his logic for this.

So there are two possibilities. He is a liar, or a troll. One is bad for him, second is against forum rules.


Insulting users is not allowed (such as calling
them liars or trolls) .

AT Mod Usandthem

 

[Konan]

Nobody was getting at anything. If you would pay attention to Sontin's posts, he usually twists people's words to fit his own agenda, and manipulation. He blatantly suggested that "somebody" here suggested that GA100 was not in production. Which NOBODY EVER did. When confronted, he refuted with his logic for this.

So there are two possibilities. He is a liar, or a troll. One is bad for him, second is against forum rules.

That is fair enough. Either way it looks like you said it will come later but it is here now

 

[Konan]

Taped out at Samsung, or TSMC? This is the thing getting my knickers in a twist.

Well that is really still up in the air where the delineation is between the two in the stack.
If we look at this:

https://twitter.com/x/status/1131031428878094337

He is saying the Ampere gaming series is all on Samsung. Personally, not so sure about that due to Jensen saying TSMC get bulk of orders publicly. So those high volume parts = not on Samsung "interpreted

You can take that list above and compare to this:

https://twitter.com/x/status/1237748855426310144

The SM's here add up to Kopite's numbers by X'ing them together. This person is also saying they are all on Samsung too. So that leaves what these two say Vs what Nvidia has said about TSMC getting the 7nm work "most orders". Surely, doesn't mean most orders are for GA100....

 

[Glo.]

About Kopite. Kopite has been accused on this very thread, by some people, as completely unreliable for Nvidia information, whereas he is actually the most accurate source for Nvidia information and rumors.

If he is claiming something about Nvidia - he never misses. If he is saying that Moore's Law is Dead video is wrong or has fake info, or is totally fake - it has very large chance that in fact Kopite is right and MLID is wrong.

That is fair enough. Either way it looks like you said it will come later but it is here now

Im very sceptical about it being hard launch. And I have very good reasons for this.

Taped out at Samsung, or TSMC? This is the thing getting my knickers in a twist.

Well that is really still up in the air where the delineation is between the two in the stack.
If we look at this:

https://twitter.com/x/status/1131031428878094337

He is saying the Ampere gaming series is all on Samsung. Personally, not so sure about that due to Jensen saying TSMC get bulk of orders publicly. So those high volume parts = not on Samsung "interpreted

There are supply chain reasons for what is Kopite saying. I cannot share with you the details, I promised to not post this anywhere, but the fact that the lineup consists of "only" 107, 106, 104 and 102 dies suggests that in fact it is made on 8 nm SS process.

However the parts could've been taken back into oven for few technical reasons(retooled and retuned), to be more competitive. If they are indeed on 8 nm process, that makes sense, and leaves large possibility, that in fact at the end of the year we will get new Nvidia gaming lineup.

And I mean at least one to two dies, released from higher end of the stack. If they would be on 7 nm we would not see them before very late Q4 of 2020.

 

[Konan]

And I mean at least one to two dies, released from higher end of the stack. If they would be on 7 nm we would not see them before very late Q4 of 2020.

But KittyCorgi said possibly August or later. Other people have eluded to paper launch with CyberPunk - Sept 17th (considering how close Nvidia is to those guys and the special edition cards and mentions) and to get them out asap. Then again others (forbes for one) said Nvidia is waiting to see what BigNavi is... for Price and performance.

 

[Glo.]

But KittyCorgi said possibly August or later. Other people have eluded to paper launch with CyberPunk - Sept 17th (considering how close Nvidia is to those guys and the special edition cards and mentions) and to get them out asap. Then again others (forbes for one) said Nvidia is waiting to see what BigNavi is... for Price and performance.

KIttyCorgi is the least reliable of the "leakers"

Don't take his info too seriously. I never did take his info seriously, I even mentioned this in my post using his die configs, to show the problem with SS's 8 nm process.

 

[Stuka87]

From AT Front Page: Installation at Argonne National Labs. Not sure of the date, but they are already there.

Yeah, they are shipping now. But Q1 ENDED 1.5 months ago.

 

[Mopetar]

I think you mean 30 mm x 30 mm , but besides that I'd be very surprised if they were getting 90% yield even with the disabled clusters.

Yeah millimeters. a 30 cm^2 chip wouldn't even fit on the wafer, never mind the reticle limit

Even if the defect density is higher than estimated, it's still reasonably to assume that they could hit 90% or more good chips. When you have a die this large, the portion of the area that goes towards the clusters or other parts of the chip that could be disabled increases as a percentage of the chip. Unless the circuitry in the critical parts that lack redundancy is more susceptible to defects for whatever reason, it's a lot less likely for a defect to be in that part of the chip.

If you hit a whale with a harpoon it's pretty unlikely to kill it. Even a few aren't likely to bring it down. Tag a cat with one and it's a different story.

At that size they're likely to have issues with running into the power limits for what the card can draw so they might want to disable the least efficient clusters anyway if it allows the others to run at a higher base frequency along with being able to sell chips with a few defective parts.

They get so few chips per wafer that it would be odd for them not to maximize the yield. It's not as though AMD is running around making monster chips like this that are competing head to head with NVidia cards that they have to worry about keeping the performance crown.