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.

 

[Ajay]

Derp, misread. My bad.

Eh, I wasn't very clear in my post.

 

[Glo.]

About 9-10 months is the norm for GPUs as a whole. If there was a tape out right now, then you'd get a chip about November-ish.

Tape out should happen around April, for gaming, next gen GPUs, which makes them Q1/H1 2021 available.

Perfect for GTC 2021 release

 

[uzzi38]

Tape out should happen around April, for gaming, next gen GPUs, which makes them Q1/H1 2021 available.

Perfect for GTC 2021 release

Yeah, sounds fairly likely. For now I'll be very general and say sometime H1 2021 though, you never know with dates like this

 

[insertcarehere]

Happens to everybody . I remember one guy not knowing who the ex-CanardPC chief was Fun times .

Since he is such a good source according to you, I am also gonna assume his radio silence on big Navi must indicate that we're not getting those until end of this year at the earliest.

 

[uzzi38]

Since he is such a good source according to you, I am gonna assume his radio silence on big Navi must indicate that we're not getting those until end of this year at the earliest.

...or that he doesn't ever post AMD based rumours.

Nice try though, but it takes two minutes of checking his Twitter profile to see that.

 

[Glo.]

Since he is such a good source according to you, I am also gonna assume his radio silence on big Navi must indicate that we're not getting those until end of this year at the earliest.

He never posts AMD rumors, or rather - inside information, my friend. He is solely Nvidia focused, If I remember correctly.

 

[insertcarehere]

...or that he doesn't ever post AMD based rumours.

Nice try though, but it takes two minutes of checking his Twitter profile to see that.

He never posts AMD rumors, or rather - inside information, my friend. He is solely Nvidia focused, If I remember correctly.

https://twitter.com/x/status/1130651430312824835

Idk what this is if not an AMD rumor, but OK 👌🏻

 

[Glo.]

https://twitter.com/x/status/1130651430312824835

Idk what this is if not an AMD rumor, but OK 👌🏻

Its his complete speculation. Try another time.

 

[uzzi38]

https://twitter.com/x/status/1130651430312824835

Idk what this is if not an AMD rumor, but OK 👌🏻

Look at the time stamp mate, not exactly a tapeout date, is it?

Navi was launched, what, 4 months later?

Also, I think it's techincally wrong, because you know, it's mostly a guess. Specifically the 32SIMD*2 part - pretty sure he's referring to cycles, and RDNA is Wave32 over a single cycle. So lol.

 

[Muhammed]

Idk what this is if not an AMD rumor, but OK 👌🏻

Don't try with them, they have fixated their mind on a preconcieved notion that NVIDIA is late, and their best evidence is an obscure twitter account. Some people never learn I guess.

They couldn't be more dead wrong.

 

[uzzi38]

Don't try with them, they have fixated their mind on a preconcieved notion that NVIDIA is late, and their best evidence is an obscure twitter account. Some people never learn I guess.

They couldn't be more dead wrong.

This is actually factually wrong, but ok then. Sure.

 

[Stuka87]

Don't try with them, they have fixated their mind on a preconcieved notion that NVIDIA is late, and their best evidence is an obscure twitter account. Some people never learn I guess.

They couldn't be more dead wrong.

You keep calling everybody wrong, but you do not appear to have anything to back it up. Its fine to not agree with somebody on a piece of speculation, but calling somebody wrong "just because" goes against how threads on this forum should be conducted.

 

[Muhammed]

You keep calling everybody wrong, but you do not appear to have anything to back it up. Its fine to not agree with somebody on a piece of speculation, but calling somebody wrong "just because" goes against how threads on this forum should be conducted.

Because random people on twitter are not sources, they are just rumor generators.

 

[Muhammed]

Moving on, the specs for the 3070 and the 3080 are supposedly leaked:

The GA103 will come with 10/20GB GDDR6 VRAM capacity. IT will also feature SM (Streaming Multiprocessors) count of 60, 3840 stream processors and 320-bit video memory.

The GA104, on the other hand, will feature 8/16GB GDDR6 VRAM memory capacity along with 48 sets of SM arrays. There will be 3072 stream processors and 256-bit video memory.

the GA103 points towards the Nvidia RTX 3080 GPU while the latter indicates RTX 3070

Nvidia RTX 3080, RTX 3070 Specifications Leaked for the First Time

Nvidia RTX 3080 and RTX 3070 has been leaked for the very first time. We now have the specifications of the Ampere based GPUs.

www.gizchina.com

 

[jpiniero]

That wouldn't be much of an SM count increase if real (esp given how big the increase in transistor density should be). Of course they could be spending a lot more on RT.

 

[Glo.]

Moving on, the specs for the 3070 and the 3080 are supposedly leaked:

Nvidia RTX 3080, RTX 3070 Specifications Leaked for the First Time

Nvidia RTX 3080 and RTX 3070 has been leaked for the very first time. We now have the specifications of the Ampere based GPUs.

www.gizchina.com

Yeah, it was leaked.

https://twitter.com/x/status/1218227064701997057

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

https://twitter.com/kopite7kimi/status/1218229502423314434
#not

Im not targeting this at anyone but: how stupid can you be to believe in this "leak"?

Read all three Twitts. There is NO Ampere other than 7 nm TSMC, and no other than GA100. Only GA exists.

 

[uzzi38]

Yeah, it was leaked.

https://twitter.com/x/status/1218227064701997057

#not

Im not targeting this at anyone but: how stupid can you be to believe in this "leak"?

Best bit is they just took Kopite's leak from months ago and made some diagrams based off of them hoping idiots would eat up the rumour. That what Kopite meant - that the diagrams aren't official ones.

 

[Muhammed]

The Sketch is not official, it's just an illustration, funny enough, the leak is coming from another twitter account too. So it's up to you to believe which one is true!

 

[beginner99]

The Sketch is not official, it's just an illustration, funny enough, the leak is coming from another twitter account too. So it's up to you to believe which one is true!

Given that it's a known fact Samsung is having issues with EUV especially on larger HPC dies and the suggested cards and code names make 0 sense, its' clear in what direction I would lean.

There never has been a xx1 or xx3 version. It's always been 100, 102, 104 and 106. That's already a clue the rumor is wrong. Then it's simply too many chips. With 7nm development costs, no way are they going to make that many chips that close to each other. Especially since the chips are probably rather big again and yields not great so there will be lots of binning opportunities.

 

[NostaSeronx]

I'm ready for that GH104 w/ a bajillion kajillion wazjillion flops of power.

 

[uzzi38]

Look at the time stamp mate, not exactly a tapeout date, is it?

Navi was launched, what, 4 months later?

Also, I think it's techincally wrong, because you know, it's mostly a guess. Specifically the 32SIMD*2 part - pretty sure he's referring to cycles, and RDNA is Wave32 over a single cycle. So lol.

Well if nothing else, it seems we were wrong and he also thinks the Navi 21 rumours are reliable.

So er, that comment about Big Navi not taping out yet seems to have aged really poorly.

Seems like he just doesn't care enough about AMD stuff, that's all.

https://twitter.com/x/status/1219098065912520704

 

[DisEnchantment]

The problems with 7LPP for NVIDIA could probably be attributed to die sizes. 7LPP is fine for Mobile dies. Ian also mentioned N5 yields for HPC are crappy too.
But I bet NVIDIA will concentrate on HPC. Gaming is probably an afterthought.

Anyway, WhyCry said NV is expected to launch a new GPU in Q2 around the same time as Big Navi.

https://twitter.com/x/status/1219171309298561024

Not sure if the GPU is Ampere/Next Gen.
Because according to @KOMACHI_ENSAKA the next gen Gaming GPUs will be earliest Q4 this year

https://twitter.com/x/status/1219209670620901378

Also the new MX series are all Pascal based.

NVIDIA GeForce MX350 and MX330 to feature Pascal architecture - VideoCardz.com

NVIDIA preparing low-power dedicated graphics for 2020 notebooks. NVIDIA MX300 Series: Pascal is back While we await more news on NVIDIA’s 7nm architecture, the company is still planning graphics cards based on Pascal. As a reminder, NVIDIA Turing is based on 12nm TSMC FinFET architecture, while...

videocardz.com

 

[DisEnchantment]

NVIDIA will want to thwart any AMD's attempt to steal the show around Computex which by AIB's account will launch Big Navi and I think WhyCry's source are quite consistent so far.
Question is which GPU will it be from NVIDIA to counter Big Navi. 2080Ti die size is already too big to compete with a much bigger RDNA(2) die. For comparison a Navi10 die on N7+ will be only 200+ mm2.
AMD will want to ride the Ryzen 4K hype to stir up interest on new Radeon products too.

Could it be the mesh shader thing for Multi GPU or a total new arch?
Can't imagine NVIDIA prioritizing DC over Gaming.
Right now the fabless companies have no choice but to start using Samsung as well, at least for the smaller dies.

 

[uzzi38]

The problems with 7LPP for NVIDIA could probably be attributed to die sizes. 7LPP is fine for Mobile dies. Ian also mentioned N5 yields for HPC are crappy too.
But I bet NVIDIA will concentrate on HPC. Gaming is probably an afterthought.

From my understanding (and I'm not taking a guess here), the issue with 7LPP/8LPP is pretty much for dies with a large number of logic levels. Power/perf and yields go to crap.

Anywho, server stuff is still intact, so that's pretty much exactly what they'll be doing.

 

[EXCellR8]

different source

NVIDIA GeForce RTX 3080, RTX 3070 leaked specs: up to 20GB GDDR6 RAM

NVIDIA Ampere: GA103/GA104 with 8/16GB and 10/20GB GDDR6 memory on 7nm.

www.tweaktown.com