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

Page 167 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

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.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
I'm sure that's a big part of it. I also know a dozen other guys personally that haven't been able to get their hands on a 3080 either, and that's somewhat abnormal for a launch. It was cake to pickup 1080TIs, 2080s, and 2080TIs, but it really does feel like demand for 3080 this go around is through the roof and they won't be able to keep up with demand.

Its also possible the rumors of higher than average failure rates of chips has caused a shortage of cards. It will be interesting to see what the sales actually were after the fact. Although only a few retail outlets actually release sales numbers, and nVidia never ever does.

And, does it seem like there is a shortage going forward for a while. With most GPU releases its quite easy to get a card 2-3 weeks after launch.
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
Its first day of launch and bots around the globe and hammering sites to try and purchase. It is not an indicator of actual demand. Its like when some new high end car comes out, and they all sell for 20K over sticker. Doesn't mean the company will sell tons of them.

NOTE: Not that I am saying 3080 will be a flop. Just that you cannot determine demand based off today.

No doubt the demand is real. Reviews are overwhelmingly positive, as is forum buzz everywhere I have read.

Unfortunately, that will lead some people to resort to scalpers, which will feed the cycle. I hate scalpers, but the only time I remember them getting burned a bit in my memory was back when they tried to scalp the PS3 launch, but the high price led to weak inital demand on PS3, and even then I think that just meant they were getting close to MSRP.
 
  • Like
Reactions: ozzy702

Kenmitch

Diamond Member
Oct 10, 1999
8,505
2,248
136

For giggles I looked on Craigslist to see how many were flipping. I only saw 1 person selling a B&H pre-order.

In his description:

Latest news from Microcenter in Tustin (from 9/17/2020 2AM):
Microcenter in Tustin had a line forming around 9PM but some people would try to form a line as early as NOON the day before. 85 Cards are confirmed to be available for the first 85 people that lined up (that was before 9PM). As of now, there is a waitlist that has another 80 people on it but those people are not guaranteed to be able to purchase a RTX 3080. There might be another shipment of 3080's in the morning though. Roll calls every 2 hours to make sure people are in fact waiting in line. Many people had their switches with them to play alone or with their friends. 6AM Pacific time is when online sales of RTX 3080 start.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
No doubt the demand is real. Reviews are overwhelmingly positive, as is forum buzz everywhere I have read.

Unfortunately, that will lead some people to resort to scalpers, which will feed the cycle. I hate scalpers, but the only time I remember them getting burned a bit in my memory was back when they tried to scalp the PS3 launch, but the high price led to weak inital demand on PS3, and even then I think that just meant they were getting close to MSRP.

I would not say they are "overwhelmingly positive". nVidia does place restrictions on what reviewers can say and show, and if reviewers break any of those guidelines, they get black listed. There is also the naming switcharoo that nVidia did where they are calling the 3080Ti a 3080, and the Titan, a 3090. Which makes people compare the 3080 to the 2080, not the 2080Ti. Even though the 3080 uses a 102 chip, and the 2080 uses a 104 chip.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I would not say they are "overwhelmingly positive". nVidia does place restrictions on what reviewers can say and show, and if reviewers break any of those guidelines, they get black listed. There is also the naming switcharoo that nVidia did where they are calling the 3080Ti a 3080, and the Titan, a 3090. Which makes people compare the 3080 to the 2080, not the 2080Ti. Even though the 3080 uses a 102 chip, and the 2080 uses a 104 chip.

There might still be a 3080 Ti (and a Titan). The 3080 is much further cut down compared to a 2080 Ti btw.

If you were comparing it on cut downness, the 3090 is the new 2080 Ti and the 3080 doesn't have a comparable product.
 
  • Like
Reactions: ozzy702

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
I would not say they are "overwhelmingly positive". nVidia does place restrictions on what reviewers can say and show, and if reviewers break any of those guidelines, they get black listed. There is also the naming switcharoo that nVidia did where they are calling the 3080Ti a 3080, and the Titan, a 3090. Which makes people compare the 3080 to the 2080, not the 2080Ti. Even though the 3080 uses a 102 chip, and the 2080 uses a 104 chip.

There seems to be a lot wrong in that post.

NDA's and embargoes do not prevent negative reviews. Go back and check Turing reviews in comparison. Negative reviews from nearly everyone.

3080 is not a misnamed x80 Ti. As I already pointed out earlier. 3080 is a appropriately named x80 series card.

Expect 3080 Ti with 3090 performance in the future. This will be a return to the pattern of Kepler, Maxwell, and Pascal, You get a more fully enabled Halo card, at a high price initially while yields are low. The only difference this time is that Halo card is using the x90 name instead of Titan. Titan name having been moved to a productivity for the last two releases.

In a few months when yields improve, producing a lower priced 3080 Ti, with 3090 performance, will be viable.

But that is all distraction, and not related to the point. Reviews are overwhelmingly positive and buzz online is very strong. People were lining up at B&M stores last night.

You can't convincingly argue demand isn't high.
 
  • Like
Reactions: xpea and ozzy702

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
3080 is not a misnamed x80 Ti. As I already pointed out earlier. 3080 is a appropriately named x80 series card.

x00 and x02 are big die chips. x04 are mainstream medium size chips. The 3080 very clearly is using big Ampere. Not mainstream Ampere like 3070. I stand by that nVidia marketing messed with naming to make it look more appealing.

980: GM204
1080: GP104
2080: TU104
3080: GA102

980 Ti: GM200
1080 Ti: GP102
2080 Ti: TU102
3090: GA102
 

CakeMonster

Golden Member
Nov 22, 2012
1,384
482
136
Those that bought 780Ti->980->980Ti->1080->1080Ti->2080Ti got their yearly 25% performance increase and most probably didn't care about the chip size at all.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,719
7,016
136
3080 is more like the 780 than any other recent release. Cut down 102 die with a substandard memory profile that maintains the non-TI x80 moniker.

NV had no need to do the same with the 9xx and 10xx series, and the TU104 is already so close in performance to the TU102 that even if NV wanted to there was really no room for an x80 102 part.

It appears pretty straightforward that NV couldn't get their GA104 = TU102 + 30% performance increase this time around, coupled with poor yields/stronger AMD/early launch, results in naming shens.
 

ozzy702

Golden Member
Nov 1, 2011
1,151
530
136
Those that bought 780Ti->980->980Ti->1080->1080Ti->2080Ti got their yearly 25% performance increase and most probably didn't care about the chip size at all.

Yep. At the end of the day chip size doesn't matter, it's price/performance and the 3080 delivers. Yes, the 2080TI pricing set the 3080 up to look even better than it is, but it's a good release despite being on a less than stellar process, and they are going to sell like hot cakes assuming NVIDIA can keep up with demand.
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
x00 and x02 are big die chips. x04 are mainstream medium size chips. The 3080 very clearly is using big Ampere. Not mainstream Ampere like 3070. I stand by that nVidia marketing messed with naming to make it look more appealing.

980: GM204
1080: GP104
2080: TU104
3080: GA102

980 Ti: GM200
1080 Ti: GP102
2080 Ti: TU102
3090: GA102

You are obsessing over chip names that don't matter.

The part that matters is that for Kepler, Maxwell, and Pascal, there was a Halo part (Titan in those cases) that shipped early at VERY high price, and low volume when yields were weak. This is a Halo part.

Months later when yields improved, they released x80 ti with ~Titan performance.

This has been the pattern since the Titan Debut with the Kepler line. Only broken by Turning. Ampere is a return to form, and this logical pattern could return.


Kepler: 780 and Halo Titan -> Followed Months later 780 ti with ~Titan performance.
Maxwell: 980 and Halo Titan -> Followed Months later 980 ti with ~Titan perf.
Pascal: 1080 and Halo Titan -> Followed Months later 1080 ti with ~Titan perf.

There is very sound logic to this launch pattern. You sell a Halo card early with early adopter pricing to reflect the low yield reality.

Later when yields improve you can offer that Halo performance at a lower price. 3090 is the that new Titan, which Jensen said during a reveal.

You have presented NO reason we won't see a sensible return to that pattern. Chip Names are not. Kepler used the same chip for 780, Titan and 780 Ti. Just as would be the case here.

Expect:
Ampere: 3080 and Halo 3090 -> Followed Months later 3080 ti with ~3090 perf.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,719
7,016
136
Yep. At the end of the day chip size doesn't matter, it's price/performance and the 3080 delivers. Yes, the 2080TI pricing set the 3080 up to look even better than it is, but it's a good release despite being on a less than stellar process, and they are going to sell like hot cakes assuming NVIDIA can keep up with demand.

-I'd disagree, it absolutely does matter as it fundamentally affects a number of characteristics about the chip that will directly affect us: How many can be made, how low can the price go, and how much room is there to slot in a mid-generation upgrade above it.

It mattered when AMD brings a 450mm2 Vega to a 300mm2 GTX1080 fight, or when NV brought a 576mm GTX 280 vs a 250mm HD4870 fight. It will matter again now.
 

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
Yep. At the end of the day chip size doesn't matter, it's price/performance and the 3080 delivers. Yes, the 2080TI pricing set the 3080 up to look even better than it is, but it's a good release despite being on a less than stellar process, and they are going to sell like hot cakes assuming NVIDIA can keep up with demand.
It's interesting to us here who like to debate all the intricacies of shader counts and memory buses and transistors per mm², but for the consumer really all the matters is that they're introducing an x80 at the same price point as the previous x80s, and it performs significantly better than the card it's replacing. That they had to use the big die for it doesn't really matter.

Now, next year it will matter because we're not getting a 3080Ti released at the same price point with +50% of the functional units. Maybe we'll get something in between the 3080 and 3090 at a reasonable price point, but there isn't going to be a 1080Ti calibre value this generation.

As an aside, there's a lot of people here who complain that Nvidia has been selling small die xx4 chips as x80s for awhile, you think they'd be happy to have a big die in the x80.
 

Hitman928

Diamond Member
Apr 15, 2012
5,182
7,633
136
As an aside, there's a lot of people here who complain that Nvidia has been selling small die xx4 chips as x80s for awhile, you think they'd be happy to have a big die in the x80.

Yep, I have been one of those people (maybe not so much complain but more just disappointed). Seeing 3080 being a big die again without increasing price from previous gen is a good thing, but at the same time it also does kind of show that Turing isn't really as big an improvement as an architecture/full stack that Nvidia tried to paint it as. You can argue that the 3080 being big doesn't matter and to wait for the 3080Ti, but the problem is (as has been pointed out) is that because the 3080 is a big die, the 3080Ti will now not be such a big deal and will be more like a 3080 Super which is what they may end up calling it anyway. All of this to say, the 3080 is a good value for Ampere, but at the same time it kind of exposes the rest of the lineup as not so great.
 
Last edited:
  • Like
Reactions: Elfear

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
Same as you can't convincingly argue that supply was plentiful.

Reading comment from users who tried to buy them I'd lean towards very little supply....Bot's get blamed. JHH snickers!

To stop bots and scalpers on the NVIDIA store, we’re doing everything humanly possible, including manually reviewing orders, to get these cards in the hands of legitimate customers.
Shouldn't take Doug the intern long to make it through the list of orders today.
 

antihelten

Golden Member
Feb 2, 2012
1,764
274
126
You are obsessing over chip names that don't matter.

The part that matters is that for Kepler, Maxwell, and Pascal, there was a Halo part (Titan in those cases) that shipped early at VERY high price, and low volume when yields were weak. This is a Halo part.

Months later when yields improved, they released x80 ti with ~Titan performance.

This has been the pattern since the Titan Debut with the Kepler line. Only broken by Turning. Ampere is a return to form, and this logical pattern could return.


Kepler: 780 and Halo Titan -> Followed Months later 780 ti with ~Titan performance.
Maxwell: 980 and Halo Titan -> Followed Months later 980 ti with ~Titan perf.
Pascal: 1080 and Halo Titan -> Followed Months later 1080 ti with ~Titan perf.

There is very sound logic to this launch pattern. You sell a Halo card early with early adopter pricing to reflect the low yield reality.

Later when yields improve you can offer that Halo performance at a lower price. 3090 is the that new Titan, which Jensen said during a reveal.

You have presented NO reason we won't see a sensible return to that pattern. Chip Names are not. Kepler used the same chip for 780, Titan and 780 Ti. Just as would be the case here.

Expect:
Ampere: 3080 and Halo 3090 -> Followed Months later 3080 ti with ~3090 perf.

You do know that you're just proving him correct right?

Stuka87's point is that the underlying tech (die size, and die classification) plus price, matches up with the Ti cards of the past, but that Nvidia changed their naming strategy, to encourage a more favorable match up against the non-Ti cards of the past.

You're then basically saying that this isn't the case, but the only evidence you're using is Nvidia's past naming strategy, which Stuka87 specifically already pointed out was changed for the Ampere generation.
 
  • Like
Reactions: Stuka87

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
Stuka87's point is that the underlying tech (die size, and die classification) plus price, matches up with the Ti cards of the past, but that Nvidia changed their naming strategy, to encourage a more favorable match up against the non-Ti cards of the past.


No I clearly showed how this release matches exactly the Kepler, Maxwell, and Pascal releases, when they released Titan followed months later by x80 Ti of essentially equal performance.

What he, and apparently you, seem to be missing, is that all the x80 Ti cards from all those generations match Titan performance.

For 3080 to be this generations 3080 Ti, it would have to match 3090 (this generations Titan) performance. It clearly does not.

Months from now when yields improve, we will almost certainly get 3080 Ti with 3090 performance.

Both of you are missing the basic relation between Titan and x80 Ti cards. It's not which chips they use. It's x80 Ti delivering Titan performance after yields improve to make it viable.
 

antihelten

Golden Member
Feb 2, 2012
1,764
274
126
No I clearly showed how this release matches exactly the Kepler, Maxwell, and Pascal releases, when they released Titan followed months later by x80 Ti of essentially equal performance.

What he, and apparently you, seem to be missing, is that all the x80 Ti cards from all those generations match Titan performance.

For 3080 to be this generations 3080 Ti, it would have to match 3090 (this generations Titan) performance. It clearly does not.

Months from now when yields improve, we will almost certainly get 3080 Ti with 3090 performance.

Both of you are missing the basic relation between Titan and x80 Ti cards. It's not which chips they use. It's x80 Ti delivering Titan performance after yields improve to make it viable.

Actually what your missing is the relationship between the 102 die and the 104 die which has been broken with Ampere.

Normally the 102 is 50% bigger than the 104 die, when looking at CUDA core count. The XX80 Ti is cut down by about 6-8%, thus making it effectively about 40% bigger than the full 104 die.

With Ampere this pattern no longer holds since the GA102 die is a whooping 75% larger than the GA104. So to maintain the normal pattern of the XX80 Ti card being 40% larger than the 104 die, Nvidia have had to cut down the 102 die more than usual. By 20% to be exact, which the 3080 matches almost perfectly by being cut down by 19%, thus making it 42% larger than the full GA104 die. So spec wise the 3080 fits in perfectly as a Ti card n the traditional XX80 to XX80 Ti pattern.

And yes this would make the 3090 as the Titan card significantly bigger than usual relatively speaking. But then again Nvidia has probably had a hard time justifying the position of the Titan cards after they dropped the whole prosumer argument (by dropping double precision support), with memory size thus being the only real differentiator. Now there

Now exactly why the 102 die is 75% larger this time around instead of the normal 50% is anyone's guess. Some rumors were that Nvidia was also planning a 103 die to sit in between the two, but I don't know how reliable those are.

With all that being said even if we ignore all of the above, there's still one fatal flaw with you explanation. If the 3080 fits in with the pattern we saw with Kepler, Maxwell and Pascal, then why in the world did Nvidia price it at $700, the price point occupied by the XX80 TI card in all those generations (except Maxwell which was actually cheaper), instead of the $500-600 that was the historical price of a XX80 card.

Thus if we go with your theory and ignore the underlying specs and die used, and look purely at the market positioning and pricing of the Kepler, Maxwell and Pascal generations, then we are left with the rather unappealing conclusion that Nvidia decided to give us XX80 non-Ti performance at XX80 Ti prices.
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,222
5,224
136
With all that being said even if we ignore all of the above, there's still one fatal flaw with you explanation. If the 3080 fits in with the pattern we saw with Kepler, Maxwell and Pascal, then why in the world did Nvidia price it at $700, the price point occupied by the XX80 TI card in all those generations (except Maxwell which was actually cheaper), instead of the $500-600 that was the historical price of a XX80 card.

Thus if we go with your theory and ignore the underlying specs and die used, and look purely at the market positioning and pricing of the Kepler, Maxwell and Pascal generations, then we are left with the rather unappealing conclusion that Nvidia decided to give us XX80 non-Ti performance at XX80 Ti prices.

Sorry, but a loaf of bread isn't 10 cents anymore either. Prices have gone up permanently, wafers are much more expensive with each fab steps, with the biggest jumps on the most recent ones.

What matters, and you really can't argue against, is performance.

Every x80 Ti cards perform like Titans and always has as long as there were Titan cards.

The 3080 clearly does not perform like this generations Titan.

In a few months when the 3080 Ti is released you will know you were wrong, though I doubt you will admit it even then.
 

Saylick

Diamond Member
Sep 10, 2012
3,084
6,184
136
Sorry, but a loaf of bread isn't 10 cents anymore either. Prices have gone up permanently, wafers are much more expensive with each fab steps, with the biggest jumps on the most recent ones.

What matters, and you really can't argue against, is performance.

Every x80 Ti cards perform like Titans and always has as long as there were Titan cards.

The 3080 clearly does not perform like this generations Titan.

In a few months when the 3080 Ti is released you will know you were wrong, though I doubt you will admit it even then.
Just figured I drop this in to help sell your point about wafer prices. Chiplet-based GPUs pretty much has to happen next generation to bring down the cost of manufacturing or else we'll either A) not get our usual performance gains, B) pay out the wazoo for our expected performance gains, C) have to wait longer than usual before next generation, or D) some combination of the above.

TSMC Wafer Prices (Not Official, but Rigorous Estimate):

EiFmaWJWsAE9M2v

Source: @RetiredEngineer
 

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
Sorry, but a loaf of bread isn't 10 cents anymore either. Prices have gone up permanently, wafers are much more expensive with each fab steps, with the biggest jumps on the most recent ones.

What matters, and you really can't argue against, is performance.

Every x80 Ti cards perform like Titans and always has as long as there were Titan cards.

The 3080 clearly does not perform like this generations Titan.

In a few months when the 3080 Ti is released you will know you were wrong, though I doubt you will admit it even then.
While you're right that a 3080 Ti will likely be released, and it likely will perform about the same as the 3090, I think you're missing the bigger picture. Traditionally the x80 is about 2/3 (67%) of the Titan, and the x80 Ti is ~93% of the Titan. Maxwell was 91.7, Pascal 93.3, Turing 94.4 for shaders. That gave Ti performance close to Titan levels, and was still a massive step up over the x80. Here the 3080 is already 83% of the 3090. It's way more performant relative to the 3090 than any historical x80 to Titan, and that makes the Ti less compelling. Even if they give us a 78 SM 3080 Ti vs the 82 SM 3090 (95%), that's still less the 15% more shaders than the 3080.

A 20GB 3080 Ti with 15% more shaders than a 3080 for $800 next summer might not be a bad card, but it's sure not the kind of increase we're used to seeing with the mid cycle Ti refresh.
 

CakeMonster

Golden Member
Nov 22, 2012
1,384
482
136
A 20GB 3080 Ti with 15% more shaders than a 3080 for $800 next summer might not be a bad card, but it's sure not the kind of increase we're used to seeing with the mid cycle Ti refresh.

This sounds like a rather likely scenario if this cycle is 2 years similarly to Turing. This will probably work out rather well for NVidia (barring a surprise from AMD). Oh and BTW unless there is that AMD surprise, a Ti or 20Gb in 2-6 months is just wishful thinking, ~10-12 months is what makes sense.

There might not be a significant performance gap to fill between 3080 and 3090 but there will be a huge marketing gap to fill in price and VRAM.

A 3080Ti/20Gb and more shaders will make people pay more because the 3080 is underspecced on VRAM. The amount might not matter in actual games, but unless AMD can upset things, the 10Gb is quite a brilliant marketing move. Everyone got the message that its fast, but they will be tempted by cards with higher VRAM numbers because people are not rational. So NVidia didn't really make the 3080 "cheap", the lower amount of VRAM simply reflects the price. They can easily add $100 or possibly $200 to that price for a 20Gb card and people will buy it. But it won't arrive for Black Friday.