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.

 

[AtenRa]

- The 2080 was a cut down chip though.

The 2080S is a better example of what we're looking at full chip vs full chip and according to TPU, it looks like a 12% performance difference in favor of the 2080S and the gap only widens at higher resolutions. Given there is roughly a 13% difference in transistor count in favor of the 2080s, thats about as linear as scaling gets despite the added transistor load and new features.

I can see NV boosting performance through higher clocks afforded by the smaller process + whatever arch magic they seem to always seem to whip up to hit that magic 30% number.

GTX980Τι has 8.0B Xtros with a die size of 601mm2 @ 28nm
GTX1080 has 7.2B Xtors with a die size of 314mm2 @ 16nm

According to TPU at the date of release, GTX1080 with 10% less Xtors was 37% faster at 1440p

Anyone expecting the RTX3080 to be 30% faster than RTX2080Ti in raster performance is going to be very disappointed

ps. TU104 (RTX2080/S) has 15% more Xtors than GP102 (GTX1080Ti) and not 13%

 

[Bouowmx]

https://twitter.com/x/status/1234861928230653955

More signs pointing to 8192 cores (likely 8 GPC, 8 TPC per GPC, 2 SM per TPC, 64 cores per SM).

Additional details by appending .gb5 to the link show:
32 GB, 4096-bit, 2.4 GT/s HBM2: 1228 GB/s
32 MB L2 cache (4x L2 cache per GPC, compared to GV100)

 

[GodisanAtheist]

GTX980Τι has 8.0B Xtros with a die size of 601mm2 @ 28nm
GTX1080 has 7.2B Xtors with a die size of 314mm2 @ 16nm

According to TPU at the date of release, GTX1080 with 10% less Xtors was 37% faster at 1440p

Anyone expecting the RTX3080 to be 30% faster than RTX2080Ti in raster performance is going to be very disappointed

ps. TU104 (RTX2080/S) has 15% more Xtors than GP102 (GTX1080Ti) and not 13%

- I am willing to grant that I might not be the sharpest tool in the shed... but it looks like you're supporting my argument?

980Ti - GM102 - 100% Relative
1080 - GP104 - 130%
1080TI - GP102 - 170%

NV "simply" die shrunk, remixed, and upclocked Maxwell and got a 30% increase in performance over their prior generation big-die with their current generation mid-die. They were able to lose some transistors by stripping out extra compute resources for the lean GP104, but that doesn't fundamentally alter the argument.

So I'm saying:

2080TI - GT102 - 100%
3080 - GA104 - 130%
3080TI - GA103 - 170%

NV has been following this playbook with die shrinks for a while now:

GTX580 - GF102 - 100%
GTX680 - GK104 - 130%
GTX 780TI - GK102 - 170%

The only time NV doesn't do these types of massive generational increases is when they launch a new arch on the same node. Top Maxwell was roughly 40% faster than Kepler and Top Turing is roughly 40% faster than Top Pascal.

Given NV's track record, I do not think it is remotely outlandish to expect the GA104 die to outperform GT102 by 30%.

Is it going to be a 300mm2 die? I doubt it, but maybe NV finds a more efficient way to get their Raster/RT performance with less transistors. Is it going to be $500 bucks? I doubt it, but who knows maybe NV will try to get ahead of the next gen consoles.

I grant I did a bit of quick and dirty with the transistors using TPUs numbers, but I don't think 15% vs 13% is really shifting my argument in either direction.

 

[Hans Gruber]

- I am willing to grant that I might not be the sharpest tool in the shed... but it looks like you're supporting my argument?

980Ti - GM102 - 100% Relative
1080 - GP104 - 130%
1080TI - GP102 - 170%

NV "simply" die shrunk, remixed, and upclocked Maxwell and got a 30% increase in performance over their prior generation big-die with their current generation mid-die. They were able to lose some transistors by stripping out extra compute resources for the lean GP104, but that doesn't fundamentally alter the argument.

So I'm saying:

2080TI - GT102 - 100%
3080 - GA104 - 130%
3080TI - GA103 - 170%

NV has been following this playbook with die shrinks for a while now:

GTX580 - GF102 - 100%
GTX680 - GK104 - 130%
GTX 780TI - GK102 - 170%

The only time NV doesn't do these types of massive generational increases is when they launch a new arch on the same node. Top Maxwell was roughly 40% faster than Kepler and Top Turing is roughly 40% faster than Top Pascal.

Given NV's track record, I do not think it is remotely outlandish to expect the GA104 die to outperform GT102 by 30%.

Is it going to be a 300mm2 die? I doubt it, but maybe NV finds a more efficient way to get their Raster/RT performance with less transistors. Is it going to be $500 bucks? I doubt it, but who knows maybe NV will try to get ahead of the next gen consoles.

I grant I did a bit of quick and dirty with the transistors using TPUs numbers, but I don't think 15% vs 13% is really shifting my argument in either direction.

A few months ago an Nvidia employee hinted that customers would be very happy with the next generation cards performance. I think it's mostly because of the die shrink but rumors are that the improvement will be like when the 1080 came out.

 

[Det0x]

More signs pointing to 8192 cores (likely 8 GPC, 8 TPC per GPC, 2 SM per TPC, 64 cores per SM).

Additional details by appending .gb5 to the link show:
32 GB, 4096-bit, 2.4 GT/s HBM2: 1228 GB/s
32 MB L2 cache (4x L2 cache per GPC, compared to GV100)

For reference:

Ampere 7936 cuda core : 220 000
Tesla V100-SXM2 5120 cuda : 212 000

 

[CakeMonster]

The only reason I could think of for them to have a slightly above average perf increase would be that the wait has been longer than average this time around.

 

[Hans Gruber]

Gamers are getting rolled by both AMD and Nvidia. The mining craze is over. Just because there are still miners doesn't mean demand has increased.

 

[MrTeal]

- I am willing to grant that I might not be the sharpest tool in the shed... but it looks like you're supporting my argument?

980Ti - GM102 - 100% Relative
1080 - GP104 - 130%
1080TI - GP102 - 170%

NV "simply" die shrunk, remixed, and upclocked Maxwell and got a 30% increase in performance over their prior generation big-die with their current generation mid-die. They were able to lose some transistors by stripping out extra compute resources for the lean GP104, but that doesn't fundamentally alter the argument.

So I'm saying:

2080TI - GT102 - 100%
3080 - GA104 - 130%
3080TI - GA103 - 170%

NV has been following this playbook with die shrinks for a while now:

GTX580 - GF102 - 100%
GTX680 - GK104 - 130%
GTX 780TI - GK102 - 170%

The only time NV doesn't do these types of massive generational increases is when they launch a new arch on the same node. Top Maxwell was roughly 40% faster than Kepler and Top Turing is roughly 40% faster than Top Pascal.

Given NV's track record, I do not think it is remotely outlandish to expect the GA104 die to outperform GT102 by 30%.

Is it going to be a 300mm2 die? I doubt it, but maybe NV finds a more efficient way to get their Raster/RT performance with less transistors. Is it going to be $500 bucks? I doubt it, but who knows maybe NV will try to get ahead of the next gen consoles.

I grant I did a bit of quick and dirty with the transistors using TPUs numbers, but I don't think 15% vs 13% is really shifting my argument in either direction.

The number of transistors between GP104 (7.2B) and GM102(8B) was quite close though, and a lot of the improvement came from a very substantial increase in clock rate with Pascal. Reference base clocks on the 1080 were 60% higher than 980Ti, and the boost was 50%-60% up. I don't think we'll see similar increases in the clock rate for GA104 vs GT102 considering the 2080 Ti boosts to 1800MHz already.

I don't know what kind of transistor density we'll see from GA104, but the Navi 10 is 41M/mm², Navi 14 is 40.5, and Vega 20 is 40. If Ampere is similar a GA104 with the same number of transistors as TU102 would be ballpark 450mm².

History has shown it's not wise to bet against nvidia with pushing performance generation after generation, but if it is a 400mm²+ die on 7nm, the economics of it will be interesting.

 

[DXDiag]

More signs pointing to 8192 cores (likely 8 GPC, 8 TPC per GPC, 2 SM per TPC, 64 cores per SM).

The full die is likely 128CU indeed.

 

[Fanatical Meat]

Gamers are getting rolled by both AMD and Nvidia. The mining craze is over. Just because there are still miners doesn't mean demand has increased.

Agreed, I am 1000% disappointed with most prices and the marginal performance increases.

 

[xpea]

For reference:

Ampere 7936 cuda core : 220 000
Tesla V100-SXM2 5120 cuda : 212 000

and 430Mhz difference in clock. Expect this Ampere to go over 300k points with final clock

 

[DXDiag]

and 430Mhz difference in clock. Expect this Ampere to go over 300k points with final clock

Also results are not directly comparable, Ampere is tested using CUDA 11 on Windows 20H1 with insider drivers, while Tesla is using CUDA 10 on Windows 1909, with public drivers.

The 118 CU + 24GB is using Driver version 445.01 with CUDA 11.0.0
The 108 CU + 48GB is using Driver version 445.35 with CUDA 11.0.28

CUDA 11 is still yet to be released

445.01 & 445.35 are both insider drivers using WDDM 2.7 for Windows 10 20H1

Current latest driver is 442.50 and CUDA 10.2

https://twitter.com/x/status/1233535708389875713

 

[Richie Rich]

I am comparing the highest Ampere chip (3080Ti) to the highest Turing chip (2080TI), in this comparison I expect NVIDIA to provide 70-80% performance uplift just like they did several times before.

Is possible that nVidia is gaining architectural advantage thanks to mobile devices? I kind of suspect that Tegra and Nintendo Switch gave nVidia hard time to develop but on the other hand they could learn and boost power efficiency in desktop GPUs. A new Nintendo Switch should be released this year so it's maybe related to Ampere too (bigger efficiency step than expected).

 

[DXDiag]

Looks like the release of Ampere for gamers is quite imminent:

wait for new TITAN, 3090, 3080 and others.

https://twitter.com/x/status/1235161517512740864

 

[Hitman928]

Looks like the release of Ampere for gamers is quite imminent:

https://twitter.com/x/status/1235161517512740864

What's the source? It says I'm not authorized to view this status.

 

[DXDiag]

Is possible that nVidia is gaining architectural advantage thanks to mobile devices? I kind of suspect that Tegra and Nintendo Switch gave nVidia hard time to develop but on the other hand they could learn and boost power efficiency in desktop GPUs.

Yes, during the Maxwell era, NVIDIA introduced Tile Rendering, which significantly boosted their power efficiency and clocks, tile rendering is often associated with mobile GPUs, which they definitely got from tampering with Tegra.

 

[Veradun]

30% on a new node? that's not going to happen, even a 2080Ti is 40% faster than a 1080Ti on the same node!

Check die size of GP102 and TU102.

 

[amrnuke]

Check die size of GP102 and TU102.

Yep.
60% increase in die area, 50+% more transistors.
The performance gain 1080 Ti -> 2080 Ti is due to brute force, not node differences.

 

[DXDiag]

Check die size of GP102 and TU102.

The 3080Ti will have both large die size and a node advantage, this will not result in a mere 40% uplift.

 

[uzzi38]

Looks like the release of Ampere for gamers is quite imminent:

https://twitter.com/x/status/1235161517512740864

Citing Kopite now?

Ignoring the fact that he was talking about TAPEOUT now in March/April just a month and a bit ago?

No, gaming Ampere is not imminant, and if you checked the post history of the very guy you're quoting, you'd know that.

 

[Glo.]

Citing Kopite now?

Ignoring the fact that he was talking about TAPEOUT now in March/April just a month and a bit ago?

No, gaming Ampere is not imminant, and if you checked the post history of the very guy you're quoting, you'd know that.

I find it funny that Kopite now is suddenly reliable source of Nvidia information, because it confirms beliefs of some people.

P.S. He always was. Whether people liked his information or not.

The thing that bothers me however is that according to Kopite's info, there will not be another HPC-based Titan, despite the fact way too many people heard about it, that there will be Titan A. Even adored in his last video about AMD Radeon upcoming GPUs said that, according to his sources, Nvidia truly fears what AMD is readying and are considering release of GA100 chip for consumers.

Confusing to say the least. I consider Kopite more reliable in terms of Nvidia info than any other source, but who knows? Maybe he hasn't heard about Titan A, yet?

 

[GodisanAtheist]

History has shown it's not wise to bet against nvidia with pushing performance generation after generation, but if it is a 400mm²+ die on 7nm, the economics of it will be interesting.

- I can see NV dropping a ~450mm2 GA104 @ $700 with 2080ti +30% Raster and maybe even higher RT performance and people calling that a steal after this generation's outlandish pricing...

There is no doubt that Turing moved the scale on die size and relative cost of performance.

 

[Tup3x]

I find it funny that Kopite now is suddenly reliable source of Nvidia information, because it confirms beliefs of some people.

P.S. He always was. Whether people liked his information or not.

The thing that bothers me however is that according to Kopite's info, there will not be another HPC-based Titan, despite the fact way too many people heard about it, that there will be Titan A. Even adored in his last video about AMD Radeon upcoming GPUs said that, according to his sources, Nvidia truly fears what AMD is readying and are considering release of GA100 chip for consumers.

Confusing to say the least. I consider Kopite more reliable in terms of Nvidia info than any other source, but who knows? Maybe he hasn't heard about Titan A, yet?

To be honest, I don't see what's the point of Titan when there's Quadro lineup. For those that do not need full blown Quadro, RTX 3080 Ti should be enough.

 

[Glo.]

To be honest, I don't see what's the point of Titan when there's Quadro lineup. For those that do not need full blown Quadro, RTX 3080 Ti should be enough.

Bragging rights for the users, and Nvidia's opportunity for earning more money?

Is that enough?

 

[Tup3x]

Bragging rights for the users, and Nvidia's opportunity for earning more money?

Is that enough?

Probably peanuts. Ti lineup will sell much better and perform about the same in games (custom versions especially - might be even faster). I don't think that they are selling many Titans at the moment.