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.

 

[MrTeal]

Now, the really fun question. Is there a CPU covering the back side of the GPU because there is a coprocessor or something else proprietary there, or because there isn't and they don't want the rumours to stop?

 

[Konan]

All the leaks seems to be spilling over today!

I'm going to share a summary of leaks from Baidu forums. Make of them what you will - Translated = keep open mind but lots of salt. The source actually pre-dates KatCorgi and Kopitekimi's information by some time. The timespy notes below in relation to the twitter leakers were time stamped as coming out before theirs, just with extra information. So here is a translated summary for speculation.... We'll see how this turns out AND if any YouTubers pick it up....


On FP32 Performance –

Rumors point to GA104 throughput being the equivalent of Navi 21. For reference that is the Big Navi performance will equal RTX 3070 (Ti). The key changes related to 3 units of FP32 tensor RT which are internal unit replacement upgrades in the ALU cluster. The area, transistors and emitter width are almost doubled…. the engine type is different, CUDA count is the same but the FP32 throughput is through the roof. Therefore all cards got a memory bump up including GA104 to 16GB and 3080 to 20GB. The 3090 FP32 is more than 24T.

An RTX 3060 GA106 FP32 performance is very competitive compared to a PS5 and is more cost effective.

AMD have greatly underestimated Ampere performance. They will now focus on price and look to be competitive there.

Regarding TSE Scores:

The TimeSpy Extreme score of 10,000 was just the beginning of testing on less than 1900mhz.

The FE card should deliver around 2000mhz and board partners cards around 2100mhz.

The drivers have had another revision since.

The TSE score in reviewer testing should be more like 11,000 to 11,500 (at ~2100mhz OC)

TSE RTX 3090 Founders Edition just under 11k (10,800-11,000) | The partner cards should reach 11,200 - 11,500 score

TSE RTX 3080 Founders Edition just under 9k 8,800 - 9,000 | The partner cards should reach 9,200 - 9,700 score

TSE RTX 3070Ti (GA 104 full fat) founders edition 7-7,300 and partner cards 7,500 - 7,800 | Note: The 3070 Ti is apparently higher than the 2080 TI TSE (3070 Ti penciled in release in October)

Performance will be reminiscent of the 9 series in terms of the partner cards – they will have greater performance improvements vs stock NV ones.

Other General Performance Notes:

Top of the line Ampere performance - FP 32 doubled, Tensor's doubled, RT doubled and memory capacity doubled. DLSS2.0 performance is 200% +

The difference between 3070 Ti and a 3080 is 20-25% | The 3070 Ti at 4K will get 60FPS easy in games.

The 3090 is about ~25% better than the 3080 in games and about 1.7 times better than a stock 2080Ti

Founders Edition cards
The radiator is exceptionally well made and great performance however; the default frequency boost is not high in the bios. There is Low frequency in FE cards but temperature control is very powerful. Overclocking apparently isn't too bad because of the lower frequency.

Note: There are FE cards and Partner Cards. There is also a strong possibility there will also be a super high-end versions of Ampere from Nvidia.

VRAM changes:

Titan = 48GB

3090 = 24GB

3080 = 20GB

All previous 12GB and 10GB were rumors or test SKUs that didn't make the cut.

GA104 3070Ti will have 16GB

GA06 will have 12GB

There is currently no 3080Ti but that doesn’t mean one could come later.

Power Consumption:

The RTX 3080 power consumption will be just slightly over 300W but not more than 340W

The RTX 3070 power consumption is coming in around 200W if not a little lower and should end up under 200W

Competition –

Apparently Big Navi is 'dead before arrival'. AMD are flustered because they have grossly underestimated Ampere performance in top 3 cards in the stack. Source rumors are saying Nvidia will win by a large amount in traditional performance this time and then DLSS and RTX scale proportionally. AMD were full of confidence before but now are very anxious. They based their targets off Turing's generational improvements (being the smallest in Nvidia history) and grossly miscalculated the evolution capability that Nvidia has done with Ampere

RTX 6900 ROPs have been improved to 96 up from 64 that was leaked a while ago. However, FP32 performance will be nowhere near GA102.

N21 = PCB is in final stages of being completed. The core has been taped out but drivers and PCB not complete yet. Expected to be completed end of September. The release and distribution are estimated to be after mid-October. Performance estimations are 1.5 times that of 5700 XT with a TSE score of less than 8000. (Will trade blows with GA104) Big Navi power consumption is estimated to be between 330-375W

On Pricing –

It is tough as no AIC has no news yet on Ampere pricing. It is expected that the Ampere cards will be expensive, and this is due to the overwhelming performance that will bring.

Pricing estimations are around 20% more than what Turing was at launch especially on high end Ampere. The 3070/Ti will be around 2080 Super to 2070 Super prices.

 

[jpiniero]

So now I am thinking there will be only two cards at launch - the 3090 and the 3080.

The 3070 and 3060 would be released in October.

The 3080 Ti would be released right before the biggest of the Big Navi, whenever that is.

 

[Glo.]

Wasn't Coreteks youtuber claiming that on the back of the PCB is a coprocessor that is doing the RT stuff?

 

[Tup3x]

If it's Colorful then that 3x8-pin makes sense. They've done it before.

 

[Glo.]

So now I am thinking there will be only two cards at launch - the 3090 and the 3080.

The 3070 and 3060 would be released in October.

The 3080 Ti would be released right before the biggest of the Big Navi, whenever that is.

There is no RTX 3080 Ti if there is 3090.

 

[Bouowmx]

On FP32 Performance –

Rumors point to GA104 throughput being the equivalent of Navi 21. For reference that is the Big Navi performance will equal RTX 3070 (Ti). The key changes related to 3 units of FP32 tensor RT which are internal unit replacement upgrades in the ALU cluster. The area, transistors and emitter width are almost doubled…. the engine type is different, CUDA count is the same but the FP32 throughput is through the roof. Therefore all cards got a memory bump up including GA104 to 16GB and 3080 to 20GB. The 3090 FP32 is more than 24T.

So, the GeForce RTX 3090 will still be marketed with 5376 cores.
But the FLOPS calculation would be like 4 FLOP per cycle? 5376 * 1.7 GHz * (4 FLOP per cycle) = 36 TFLOPS
It's a lot more than "more than 24 TFLOPS", so the more conservative estimate is implying not all the doubled FP32 throughput can be used at once?
A rumor first mentioned back in March (Twitter Khatcorgi) is coming back full force.

At this point, rumors as absurd as a co-processor might not be so absurd (like GDDR6X memory).

 

[blckgrffn]

So, I can basically ignore the rumor-mill spec sheets that list anything greater than 12GB of vram on these guys now, right?

I didn't realize that was what the Anandtech article was implying, but that GDDR6X leak seems really solid.

 

[aigomorla]

So, I can basically ignore the rumor-mill spec sheets that list anything greater than 12GB of vram on these guys now, right?

I didn't realize that was what the Anandtech article was implying, but that GDDR6X leak seems really solid.

20GB of ram on 3080ti

UPDATED SAT, AUG 8 2020 10:29 AM CDT
Read more: https://www.tweaktown.com/news/7005...070-leaked-specs-up-20gb-gddr6-ram/index.html

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

So, the GeForce RTX 3090 will still be marketed with 5376 cores.

I think the 3090 has been debunked, as it will not come out.
Or the leaker mistook it for a possibly "Titan" class edition.

 

[Bouowmx]

So, I can basically ignore the rumor-mill spec sheets that list anything greater than 12GB of vram on these guys now, right?

The PCB images you see are of the back side. So, an upcoming graphics card will be using double-sided memory: 12 on back, 12 on front, 1 GB each, 24 GB total.

NVIDIA GeForce RTX 3090 features twelve GDDR6X memory modules - VideoCardz.com

It has been discovered that the PCB leaked earlier today features, in fact, twelve memory chips instead of eleven, as we reported in a previous post. NVIDIA GeForce RTX 3090 features 24GB of memory This has been discovered by @stixalot who noticed the missing information on a newly posted RTX...

videocardz.com

You can only count 11 but there is one more hidden at bottom by the SODIMM.

 

[linkgoron]

I don't see nvidia launching both a 3080 and 3090/3080ti at the same time if they completely destroy the competition and their older cards. IMO They would return to releasing a 80 one year and ti the next, similar to the Maxwell and Pascal releases).

The only reason that the 2080ti and 2080 were released at the same time was because the 2080 was a relatively small upgrade from the 1080ti.

 

[CastleBravo]

Adored thinks the upcoming GeForce cards will be made on Samsung 5LPE, which is an evolution of the Samsung 7LPE process that had terrible yields, and is supposedly design compatible.

 

[aigomorla]

The PCB images you see are of the back side. So, an upcoming graphics card will be using double-sided memory: 12 on back, 12 on front, 1 GB each, 24 GB total.

i do not trust that source.. or that site especially since the top has in big boxed statement:

Please note that this post is tagged as a rumor.

From all the vines ive heard from, the 3090 is a golden hostess Twinky. Its only meant to make the little kids like us drool at the possibility, but the 3090 is entirely rumor sauce as a upcoming prediction / vaporware of possibly a new Titan.
However i heard Nvidia is very iffy about making another Titan as the sales on them are very dismal. People rather go full Quadro line then go Titan as Quadros have better work performance, and the marginally better titan gaming experience is not worth it.

Titans just don't make sense anymore, hence why they were clashed even in mythology.,

 

[jpiniero]

I don't see nvidia launching both a 3080 and 3090/3080ti at the same time if they completely destroy the competition and their older cards. IMO They would return to releasing a 80 one year and ti the next Similar to the Maxwell and Pascal releases).

The only reason that the 2080ti and 2080 were released at the same time was because the 2080 was relatively small upgrade from the 1080ti.

Perhaps it's a hedge. Say Big Navi is a bust, the 3080 Ti won't get released. nVidia has typically kept Titan branded cards for themselves, the 3090 will probally be available by AIBs.

 

[CakeMonster]

I see very detailed specs from various sites and personalities, but they don't agree with each other.

What I'm mostly curious about:

1) If NVidia is launching 2 cards that are ahead of the 2080Ti, that would be a very ballsy move, and I'm really interested in the thought process behind it. They don't need to unless they know AMD has something up their sleeve. I fear pricing will be rather crazy though.

2) If the rumors that suggest memory capacity of 16Gb and up for these cards are true, they are really taking on the consoles and futureproofing for the console ports as well. Probably too good to be true but one can hope.

3) If the 3090 is 1.7x 2080Ti performance as suggested in one previous post they can pretty much charge however much they want. Prepare for prices north of $1500.

 

[aigomorla]

Prepare for prices north of $1500

i am prepared for prices already north of 1.5k on a 3080ti.... lol...
If the golden Twinky comes live, i wouldn't be surprised at a tag of upwards of 3k like how the Titan RTX is near priced.

 

[Glo.]

20GB of ram on 3080ti

UPDATED SAT, AUG 8 2020 10:29 AM CDT
Read more: https://www.tweaktown.com/news/7005...070-leaked-specs-up-20gb-gddr6-ram/index.html

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

I think the 3090 has been debunked, as it will not come out.
Or the leaker mistook it for a possibly "Titan" class edition.

Yeah it was debunked to the degree that Micron, the company that makes GDDR memory chips for those GPUs, confirmed existence of RTX 3090, today.

 

[Karnak]

Performance estimations are 1.5 times that of 5700 XT
[...]
Big Navi power consumption is estimated to be between 330-375W

Obviously fake, why do you even bother? 1.5x faster than a 5700XT but >=1.47x higher power consumption at the same time? How's that in line with AMD's +50% Perf/Watt claim? Just lol. Probably like 90% about Ampere a lie then too after that.

Adored thinks the upcoming GeForce cards will be made on Samsung 5LPE, which is an evolution of the Samsung 7LPE process that had terrible yields, and is supposedly design compatible.

Nah, 5nm won't happen. Why? Because 5nm entered mass production in Q2 2020, so like 2-3 months ago. Ampere will launch next month and that's just impossible in that short of a time frame. Definitely won't happen. I know rumors are cool but Adored needs to sort out what could happen and what is litereally impossible. 5nm is the latter.

 

[Stuka87]

Competition –

Apparently Big Navi is 'dead before arrival'. AMD are flustered because they have grossly underestimated Ampere performance in top 3 cards in the stack. Source rumors are saying Nvidia will win by a large amount in traditional performance this time and then DLSS and RTX scale proportionally. AMD were full of confidence before but now are very anxious. They based their targets off Turing's generational improvements (being the smallest in Nvidia history) and grossly miscalculated the evolution capability that Nvidia has done with Ampere

RTX 6900 ROPs have been improved to 96 up from 64 that was leaked a while ago. However, FP32 performance will be nowhere near GA102.

N21 = PCB is in final stages of being completed. The core has been taped out but drivers and PCB not complete yet. Expected to be completed end of September. The release and distribution are estimated to be after mid-October. Performance estimations are 1.5 times that of 5700 XT with a TSE score of less than 8000. (Will trade blows with GA104) Big Navi power consumption is estimated to be between 330-375W

I know you didn't write this. But who ever did is taking the sensationalism to the next level. *NO* tech company can estimate what the competition is going to do YEARS before they come out. A companies comes up with a design, and moves forward on it.

Actually that entire bit reads like somebody that just wanted to make things up for people to feed off. Which we get every time a new GPU line comes out. The majority of what they said just doesn't line up with either other rumors, or the facts that have come out today.

 

[Bouowmx]

Adored thinks the upcoming GeForce cards will be made on Samsung 5LPE, which is an evolution of the Samsung 7LPE process that had terrible yields, and is supposedly design compatible.

What a timely video. It's times like this when this ridiculous idea actually as a probability of truth. Still feels unlikely.

 

[Glo.]

Competition –

Apparently Big Navi is 'dead before arrival'. AMD are flustered because they have grossly underestimated Ampere performance in top 3 cards in the stack. Source rumors are saying Nvidia will win by a large amount in traditional performance this time and then DLSS and RTX scale proportionally. AMD were full of confidence before but now are very anxious. They based their targets off Turing's generational improvements (being the smallest in Nvidia history) and grossly miscalculated the evolution capability that Nvidia has done with Ampere

RTX 6900 ROPs have been improved to 96 up from 64 that was leaked a while ago. However, FP32 performance will be nowhere near GA102.

N21 = PCB is in final stages of being completed. The core has been taped out but drivers and PCB not complete yet. Expected to be completed end of September. The release and distribution are estimated to be after mid-October. Performance estimations are 1.5 times that of 5700 XT with a TSE score of less than 8000. (Will trade blows with GA104) Big Navi power consumption is estimated to be between 330-375W

Completely wrong.

 

[Konan]

Obviously fake, why do you even bother? 1.5x faster than a 5700XT but >=1.47x higher power consumption at the same time? How's that in line with AMD's +50% Perf/Watt claim? Just lol. Probably like 90% about Ampere a lie then too after that.

I didn't write it just re-posting it. Yes 50% faster than 5700XT : I can believe that. As for power consumption on that card and with 2x performance wouldn't that put it around 300W+ ? (including 50% per/watt claim)

I know you didn't write this. But who ever did is taking the sensationalism to the next level. *NO* tech company can estimate what the competition is going to do YEARS before they come out. A companies comes up with a design, and moves forward on it.

Actually that entire bit reads like somebody that just wanted to make things up for people to feed off. Which we get every time a new GPU line comes out. The majority of what they said just doesn't line up with either other rumors, or the facts that have come out today.

Yes, thank you Stuka, I didn't write it. Agree lots of sensationalism in places. I personally work in tech and we roadmap out competitor potential capabilities and direction 1-3 years out, and sometimes more even. How that turns out is above average with anticipation, never 100% accurate but tech trends are there in hardware and software. Doesn't always influence the end product, but it can.

 

[CP5670]

3) If the 3090 is 1.7x 2080Ti performance as suggested in one previous post they can pretty much charge however much they want. Prepare for prices north of $1500.

I think $2000 is easily possible if this is actually true, maybe even $2500 in line with the Titan RTX. Even the 2080Ti is close to $1500 these days due to virus-related demand. I can't see why they would go any lower, as there is clearly a market for them even at those prices.

 

[amenx]

RedGaming Tech discussing some of the leaks mentioned above.

 

[TESKATLIPOKA]

I didn't write it just re-posting it. Yes 50% faster than 5700XT : I can believe that. As for power consumption on that card and with 2x performance wouldn't that put it around 300W+ ? (including 50% per/watt claim)

If big navi has 80CU as mentioned in leaks, then It won't be just 50% faster than 5700xt. If the performance is 2x of 5700xt the TDP can be 300W including 50% better perf/w.

About the summary of leaks about Ampere It's an interesting read but I don't think everything is correct, especially the performance.
Luckily we don't need to wait for long, only a few weeks remain for Ampere presentation.