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

Ottonomous

Senior member
May 15, 2014
559
292
136
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.
 

swilli89

Golden Member
Mar 23, 2010
1,558
1,181
136
Yeah really can't wait for this next gen, should give us a nice boost. AMD should roll out big Navi by mid-Spring next year, and I expect NVIDIA's 7nm Samsung chips to roll out around that time or slightly after.
 

Guru

Senior member
May 5, 2017
830
361
106
We have no clue and they are quite far off. At least half a year away, so zero information so far.

I mean pure speculation is that just from going from 12nm to 7nm they are going to gain either 15% performance or 20% power efficiency at the same die size and stuff.

And we don't know the quality of Samsung's 7nm, if its even going to be Samsung, they don't sell their production to others, they just licence their process, so we have no clue really what is really going to happen.
 

jpiniero

Lifer
Oct 1, 2010
14,487
5,151
136
And we don't know the quality of Samsung's 7nm, if its even going to be Samsung, they don't sell their production to others, they just licence their process, so we have no clue really what is really going to happen.

Uh, Samsung does produce for others, they just don't have that much business. nVidia's GP107 and GP108 were fabbed by Samsung on their 14 nm node.
 
Mar 11, 2004
23,020
5,485
146
I'm not sure anyone outside of Samsung knows that. I'm also not sure if that would tell you much as their knowledge with regards to it is likely based around small mobile SoC and not large GPU.
Doubt anyone other than Nvidia knows that yet. Most people expect they'll be beefed up. We'll see how much. I honestly am not expecting a big increase as its already a significant part of the chip, and games are not transitioning that quickly to support ray-tracing, and although the new consoles will have ray-tracing hardware I doubt its really powerful, and Nvidia can just rely on gradual progression and they'll still handily top the consoles. Better utilization might help quite a bit (we'll see; ray-tracing seems to be pretty difficult to work around just the sheer processing needed).
Probably will have some, but what they might be and how much they apply to games is anyone's guess.
Dunno.
That's actually the biggest question. There's speculation that Ampere is just a replacement for Volta, which was not a consumer GPU.
Depends on the game and settings obviously.
Probably, but I doubt they will soon. We'll see how GDDR6 advances, but for now I think it offers enough bandwidth for their gaming needs.
Like? Given their history I'd guess they will, but they already have so many and some of them are starting to falter or transition from being proprietary (GSync, PhysX).
 

jpiniero

Lifer
Oct 1, 2010
14,487
5,151
136
The main issue with nVidia right now is the Datacenter product line, which is in dire dire need of a refresh. I could see nVidia doing a $$$$$ Titan as the first gaming card.
 
  • Like
Reactions: Ajay

happy medium

Lifer
Jun 8, 2003
14,387
480
126
NVIDIA Ampere Specs, Performance, Pricing Leaked – Massively Faster Ray Tracing, Higher Clocks, More vRAM & Lower TDP vs. Turing.



You must post personal commentary with
links.

AT Mod Usandthem
 
Last edited by a moderator:

GodisanAtheist

Diamond Member
Nov 16, 2006
6,685
6,944
136
I think every rumor is thinking the same thing right now: Ampere is to Turing what Pascall was the Maxwell. It's the tock in NV's tick tock - beefed up existing arch on a new node. The fact that it's even a different fab company than the one NV has always worked with would even more strongly reinforce this idea (don't want to introduce too many variables when doing such a low level change).

I think we can all expect a heartier mix of RT and tensor cores per CU. I don't think that hardware is going anywhere, and I wouldn't be surprised if the underlying HW itself was more capable than Turing.
 

Bouowmx

Golden Member
Nov 13, 2016
1,138
550
146

Read in the comments (always a good chuckle from WCCFTech comments) that this stuff is from a now-deleted Reddit post.

Side note: If you've been following Reddit, it just went through an episode of shit "anonymous source(s) say" with a really dumb conspiracy theory of NVIDIA paying YouTube reviewers for product placement in AMD reviews. Not going to further dignify it.
 

JSt0rm

Lifer
Sep 5, 2000
27,399
3,947
126
Read in the comments (always a good chuckle from WCCFTech comments) that this stuff is from a now-deleted Reddit post.

Side note: If you've been following Reddit, it just went through an episode of shit "anonymous source(s) say" with a really dumb conspiracy theory of NVIDIA paying YouTube reviewers for product placement in AMD reviews. Not going to further dignify it.

nvidia paid a guy to post on these forums. Why wouldnt they pay youtubers?
 

mopardude87

Diamond Member
Oct 22, 2018
3,348
1,575
96
Article states a first half of 2020 release? Is this set in stone? I got a 3gb 1060 that i want to upgrade to a 5700xt/1080ti/2080 Super. Its been crossing my mind to wait as far out as the end of next year for a upgrade. Would be nice if Amperes release is sooner then that.

May just jump on a cheap 5700xt or used 2070 and ride it out. Sell it off to help fund the Ampere upgrade or put into my spare i5 4670 1080p based htpc rig. Would be overkill for both 1080p and the 4670 though.
 

Bouowmx

Golden Member
Nov 13, 2016
1,138
550
146
My slightly educated opinion:

Core count is 1.33x Turing. Assuming no RT-'incapable' successor to GeForce 16 series:
TU116 -> A106 (GeForce RTX 3060): 2048 cores
A104 (3070): 3072
A104 (3080): 4096
A102 (3080 Ti): 5888
A102 (TITAN): 6144

In architecture, Ampere needs to be more than a straight shrink of Turing, since there is no large frequency increase (~40%) as was going from planar to FinFET.

An interesting thing to watch is transistor density. AMD GPUs on TSMC 7 nm are not close to theoretical max density (41.0 vs 91.2 M/mm^2). For simplicity in calculating die area, assume transistor count A102 is 1.33x TU102. 24800 M = 1.33 * 18600 M.
Same as AMD Navi 10 (41.0 M/mm^2): 604 mm^2
0.75x TSMC 7 nm (68.4 M/mm^2): 363 mm^2
0.875x Samsung 7 nm HP (67.4 M/mm^2): 368 mm^2
 

jpiniero

Lifer
Oct 1, 2010
14,487
5,151
136
Article states a first half of 2020 release? Is this set in stone? I got a 3gb 1060 that i want to upgrade to a 5700xt/1080ti/2080 Super. Its been crossing my mind to wait as far out as the end of next year for a upgrade. Would be nice if Amperes release is sooner then that.

Seems pretty likely that the first product will be released in the first half.
 

amenx

Diamond Member
Dec 17, 2004
3,842
2,003
136
"first half" of 2020 is a 6 month span. I dont think Nvidia is on record as saying first half of 2020, these are just rumors from rumor sites. Nvidia would be very tight-lipped about release dates until the last couple months or so, for obvious reasons.
 

Bouowmx

Golden Member
Nov 13, 2016
1,138
550
146
Why would we assume that? Is this a historical trend for NV?

For the *02-class die:
For core count:
Maxwell - Pascal: 1.25x
Pascal - Turing: 1.2x
I sort of arbitrarily picked 1.33x, which also leads to core counts being multiples of 1024.

For transistor count:
M - P: 1.5x
P - T: 1.55x
Transistor count estimate should be more like 1.5x then.
 

jpiniero

Lifer
Oct 1, 2010
14,487
5,151
136
"first half" of 2020 is a 6 month span. I dont think Nvidia is on record as saying first half of 2020, these are just rumors from rumor sites. Nvidia would be very tight-lipped about release dates until the last couple months or so, for obvious reasons.

True. I do think what will happen is that they will announce the Ampere Tesla first in Q1 and then a Titan (or Titan priced) soon thereafter. Whether you will see any normally priced cards in the first half - that I am unsure about.
 

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
For the *02-class die:
For core count:
Maxwell - Pascal: 1.25x
Pascal - Turing: 1.2x
I sort of arbitrarily picked 1.33x, which also leads to core counts being multiples of 1024.

For transistor count:
M - P: 1.5x
P - T: 1.55x
Transistor count estimate should be more like 1.5x then.

Not sure how much core count will go up. Like the transition to Turing, I'd expect much more ray tracing hardware (dunno about tensor bits - seems like a feature ppl didn't like).
Samsung's 7nm partial EUV node could deliver close to twice the xtor density of TSMC 12FFN node - so we'll see how big a die NV want's for it's consumer cards. If it goes big, that would be allot of graphics and ray tracing power.