Discussion NVIDIA Ampere RTX3000 series Discussion Thread(Updated Launch Specs)

[King Dranzer]

NVIDIA Ampere is rumored to be announced in August and launch in September at Computex. It is said to be decent improvement over Turing with around 40% performance improvement through out the lineup. With the launch of RTX3080Ti(most probably that is what the RTX2080Ti successor is going to be called) we can expect 4K 60fps+ in 90%+ titles at ultra settings which was not possible with present high end cards well still 4K 100fps+ will out of reach for majority of demanding titles.

GA102(TITAN, RTX3080Ti):
5376 CUDA Cores
24GB GDDR6

GA103(RTX3080):
3840 CUDA Cores
10GB GDDR6

GA104(RTX3070,RTX3060):
3072 CUDA Cores
8GB GDDR6

GA106(RTX2660Ti,RTX2660):
1920 CUDA Cores
6GB DDR6

GA107(RTX2650):
1280 CUDA Cores
4GB GDDR6

I doubt there will be non RTX cards even on low end, I guess all the cards will be Ray-tracing enabled and come packed with RT Cores.
The info out there is very little but the potential leaks of specs or rumors are interesting. We did expect even more than this but 40% performance improvement over Turing is not too bad. I hope there is at-least comparatively bigger jump in Ray-tracing performance compared to Raw performance making Ray-tracing bit more mainstream.

Are we finally getting to the point where a single GPU can handle 4K@60Hz without any limitations.

UPDATED SPECS AFTER LAUNCH
​

RTX3090(TITAN)
10496 CUDA Cores
24GB GDDR6X
$1499

RTX3080
8704 CUDA Cores
10GB GDDR6X
$699

RTX3070
5888 CUDA Cores
8GB GDDR6
$499

That is pure Insanity in comparison to what we were expecting. Interesting. Lets see how the Performance goes in real world. On paper the specs are pure Insanity. If it delivers similarly in real world then we are in for a big WIN.

 

[renz496]

it can be as high as 80% performance increase. GTX980ti to GTX1080ti is around 85% performance increase (at 4k where the CPU bottleneck is the most minimal). even with turing nvidia probably can go even faster than what they can with RTX titan but CPU bottleneck starts becoming real issue even on higher resolution.

 

[King Dranzer]

it can be as high as 80% performance increase. GTX980ti to GTX1080ti is around 85% performance increase (at 4k where the CPU bottleneck is the most minimal). even with turing nvidia probably can go even faster than what they can with RTX titan but CPU bottleneck starts becoming real issue even on higher resolution.

Nah, I cant think of a card that could get you day and night difference at 4K with simple CPU upgrade. There is improvement but is less than 5% in 90%+ games. Yes games like CSGO even at 4K will show noticeable improvement with CPU upgrade. For majority of games CPU bottleneck will not be an issue at 4K.

 

[renz496]

Nah, I cant think of a card that could get you day and night difference at 4K with simple CPU upgrade. There is improvement but is less than 5% in 90%+ games. Yes games like CSGO even at 4K will show noticeable improvement with CPU upgrade. For majority of games CPU bottleneck will not be an issue at 4K.

that's the problem. for our GPU not to be bottlenecked we need higher resolution. worse thing about bottleneck is they give use lower performance on faster gpu. first have my GPU bottleneck when my GTX460 (which is more or less perform on the same performance tier as GTX285) give me lower frame rate than my older GTS250. right now the only resolution that will not going to bottleneck RTX2080Ti is 4k. there probably be a time we wonder why in certain tittle RTX2080Ti can reach 144FPS (or even faster) when it supposed have no problem to get such performance on lower resolution.

my point is before we get excited with much more faster GPU we need even faster CPU first to really feed out GPU. personally i have no interest at 4k. given a choice i would take 1080p 144hz over 4k 60FPS. so i want a GPU that can maintain 144hz on max setting and hope CPU bottleneck will not going to lower my frame rate.

 

[King Dranzer]

that's the problem. for our GPU not to be bottlenecked we need higher resolution. worse thing about bottleneck is they give use lower performance on faster gpu. first have my GPU bottleneck when my GTX460 (which is more or less perform on the same performance tier as GTX285) give me lower frame rate than my older GTS250. right now the only resolution that will not going to bottleneck RTX2080Ti is 4k. there probably be a time we wonder why in certain tittle RTX2080Ti can reach 144FPS (or even faster) when it supposed have no problem to get such performance on lower resolution.

my point is before we get excited with much more faster GPU we need even faster CPU first to really feed out GPU. personally i have no interest at 4k. given a choice i would take 1080p 144hz over 4k 60FPS. so i want a GPU that can maintain 144hz on max setting and hope CPU bottleneck will not going to lower my frame rate.

Yes you are correct when we are discussing about competitive gaming. I don't know a single competitive gamer who would go for a 4K@60Hz monitor over 1080p@240Hz or 1440p@120Hz. To be honest till now I have only recommended 1080p@144Hz/240Hz or 1440p@120Hz/144Hz for any kind of gaming and never recommended a 4K@60Hz monitor unless one specifically asks for. But as the New GPU lineup from NVIDIA shows up and if it is 40%+ improvement over present gen I will start recommending 4K@60Hz monitors for casual and non competitive gamers as they would like to have better visual experience over faster refresh as they will just be playing to have fun and pleasure.

 

[renz496]

btw seeing turing is fully compatible with DX12 ultimate spec we might not going to see big changes on ampere like how it go from pascal to turing. IMO nvidia probably taking the same approach they are doing with pascal for ampere. refining turing for maximum efficiency in both die size and clock speed. the only major improvement probably on RT core efficiency. if they can pull another "pascal" that alone can already be quite big for nvidia.

 

[King Dranzer]

btw seeing turing is fully compatible with DX12 ultimate spec we might not going to see big changes on ampere like how it go from pascal to turing. IMO nvidia probably taking the same approach they are doing with pascal for ampere. refining turing for maximum efficiency in both die size and clock speed. the only major improvement probably on RT core efficiency. if they can pull another "pascal" that alone can already be quite big for nvidia.

If they pull another Pascal that will not be quite big. That will be insane. But I guess they will be mainly focusing on Ray Tracing. But still no matter how big of improvement Ray-Tracing will be I think it will be one more generation wait(after Ampere) for it to become mainstream.

 

[cryoburner]

Are we finally getting to the point where a single GPU can handle 4K@60Hz without any limitations.

I doubt it. Raytracing will likely be the norm for "ultra" settings very soon, and its already here in some capacity in a number of titles. I do expect significant improvements to raytracing performance on the new hardware, but considering the 2080 Ti can't even manage 60fps at native 1440p with raytraced lighting effects turned up in last-year's games, it seems a bit questionable whether the new top-end card will be pushing more than twice as many pixels at significantly better frame rates at native 4K. Especially in games targeting the next generation of consoles, which are likely to be more graphically demanding overall.

That goes for CPU performance too though. While it probably won't be as much of a concern at 4K, due to games being limited more by graphics performance at that resolution, on high-refresh screens at lower resolutions we will probably be seeing a hit to frame rates in many of the titles targeting the new consoles. The Zen 2 processors utilized by the Xbox Series X and PS5 should be much closer to current PC hardware than what consoles have been using for quite a while, and if developers aim for 60fps on those CPUs, you might not be seeing over 100fps on the PC versions of many of those games, even with the fastest desktop CPUs.

Back to higher resolutions, at least upscaling and sharpening techniques have been improving lately, so it's arguably not as bad to be running games at non-native resolutions. Rendering games at something like 1440p upscaled to 4K at ultra settings with raytracing enabled is probably going to look better than turning graphics options down just to get a slightly sharper image at native 4K. Of course, I would still rather go for a high-refresh 1440p screen, at least until high-refresh 4K screens start appearing at reasonable price points.

 

[King Dranzer]

I doubt it. Raytracing will likely be the norm for "ultra" settings very soon, and its already here in some capacity in a number of titles. I do expect significant improvements to raytracing performance on the new hardware, but considering the 2080 Ti can't even manage 60fps at native 1440p with raytraced lighting effects turned up in last-year's games, it seems a bit questionable whether the new top-end card will be pushing more than twice as many pixels at significantly better frame rates at native 4K. Especially in games targeting the next generation of consoles, which are likely to be more graphically demanding overall.

That goes for CPU performance too though. While it probably won't be as much of a concern at 4K, due to games being limited more by graphics performance at that resolution, on high-refresh screens at lower resolutions we will probably be seeing a hit to frame rates in many of the titles targeting the new consoles. The Zen 2 processors utilized by the Xbox Series X and PS5 should be much closer to current PC hardware than what consoles have been using for quite a while, and if developers aim for 60fps on those CPUs, you might not be seeing over 100fps on the PC versions of many of those games, even with the fastest desktop CPUs.

Back to higher resolutions, at least upscaling and sharpening techniques have been improving lately, so it's arguably not as bad to be running games at non-native resolutions. Rendering games at something like 1440p upscaled to 4K at ultra settings with raytracing enabled is probably going to look better than turning graphics options down just to get a slightly sharper image at native 4K. Of course, I would still rather go for a high-refresh 1440p screen, at least until high-refresh 4K screens start appearing at reasonable price points.

Hahahaha...................
Yes definitely not expecting Ray-tracing at 4K. Or Ray-tracing itself to become mainstream to extent of being considered as main feature. For now Ray-tracing is limited to being additional feature provided and supported by some new titles and will remain the case for at-least another generation or two(2-4Yrs from now).

Yes Consoles are getting Ray-tracing implemented into them as well but obviously they will not be great at doing that and not many titles will be initially supporting it. By the time Ray-tracing becomes mainstream Consoles will be ready to drop in their Refresh models.

Anyhow I am not expecting 4K Raytracing any time soon.

 

[larsv8]

I am super excited.

Going to make the next flagship Ampere card my next upgrade.

 

[King Dranzer]

Hahahaha...................
Yes definitely not expecting Ray-tracing at 4K. Or Ray-tracing itself to become mainstream to extent of being considered as main feature. For now Ray-tracing is limited to being additional feature provided and supported by some new titles and will remain the case for at-least another generation or two(2-4Yrs from now).

Yes Consoles are getting Ray-tracing implemented into them as well but obviously they will not be great at doing that and not many titles will be initially supporting it. By the time Ray-tracing becomes mainstream Consoles will be ready to drop in their Refresh models.

Anyhow I am not expecting 4K Raytracing any time soon.

That will be an insane jump in performance from GTX1080 if you get RTX3080Ti

 

[renz496]

time to get amped soon.

 

[King Dranzer]

time to get amped soon.

40% performance jump throughout the lineup is not bad at all.

 

[renz496]

to be honest i don't know how much jump we going to see. will be interesting to see what nvidia will going to show us later at GTC 2020 pre-recorded event.

 

[King Dranzer]

to be honest i don't know how much jump we going to see. will be interesting to see what nvidia will going to show us later at GTC 2020 pre-recorded event.

Yeah but it will be limited to AI, Workstation and will not be anything related to mainstream Gaming GPUs.

 

[renz496]

Yeah but it will be limited to AI, Workstation and will not be anything related to mainstream Gaming GPUs.

it depends. if they somehow show us their new line of quadro then we might be able to get some glimpse of what next gen geforce can do. in GTC 2018 nvidia first show case RTX Quadro in that event. we immediately know geforce turing will have RT core and tensor cores inside them that capable of running RT scene using a single GPU which took 4 titan volta to run when DXR being first showcased earlier that year.

but it is also possible nvidia will not want to talk about their quadro on upcoming GTC videos to build unnecessary hype. because right now AIB are working hard to get rid of existing RTX 20 series inventory.

 

[King Dranzer]

it depends. if they somehow show us their new line of quadro then we might be able to get some glimpse of what next gen geforce can do. in GTC 2018 nvidia first show case RTX Quadro in that event. we immediately know geforce turing will have RT core and tensor cores inside them that capable of running RT scene using a single GPU which took 4 titan volta to run when DXR being first showcased earlier that year.

but it is also possible nvidia will not want to talk about their quadro on upcoming GTC videos to build unnecessary hype. because right now AIB are working hard to get rid of existing RTX 20 series inventory.

I doubt that they will release any kind of info even on Quadro lineup. This time everything got held back a bit because of Covid-19. I would be surprised if they give us any kind of info on Quadro

 

[renz496]

but even if they only talk about compute Ampere it might be interesting. already heard some crazy rumors out there about GA100

 

[King Dranzer]

RTX3060 to have equivalent Ray-Tracing performance to RTX2080Ti if not better. Hmmmm.....
Ampere to have 10-20% IPC improvement over Turing. Hmmmmmmmmmmmm.................

This is interesting. 10-20% improvement on IPC alone and good increase in core count. We can expect decent upgrade in performance over Turing.

 

[King Dranzer]

I don't get the reason why people are expecting NVIDIA to throw in their GA100 having 8192CUDA cores, 1024Tensor cores and 256RT cores as RTX3080Ti offering. Do people really expect NVIDIA to release a card which is a good 2x or even more performance than RTX2080Ti. Then what RTX3080 will be 40% more powerful than RTX2080Ti. If people have to expect some thing it should at-least be practical enough to dream of. This is so practically impractical that I can't even dream of it. NVIDIA is never gonna do that. They are not mad or have lost their mind completely.

 

[renz496]

well those GA100 is more complicated than what many think. JHH said both compute and gaming will use one architecture (ampere) but looking at GA100.....that thing is more complicated than nvidia past compute card. and i don't know what does it mean for gaming ampere. will it more like traditional GPU...or will nvidia make clever use of tensor core to redefine how GPU being design.

 

[King Dranzer]

well those GA100 is more complicated than what many think. JHH said both compute and gaming will use one architecture (ampere) but looking at GA100.....that thing is more complicated than nvidia past compute card. and i don't know what does it mean for gaming ampere. will it more like traditional GPU...or will nvidia make clever use of tensor core to redefine how GPU being design.

I missed the keynote as I was busy with something else. But there is mixed argument all over internet I tried to go through multiple sites to get accurate info and it made it worse. Some are saying it is a big improvement don't know on what context they were claiming it. Some are very disappointed by the numbers as they were expecting 8000+ CUDA Cores on the GA100 and only getting 6000+ cores

 

[renz496]

I missed the keynote as I was busy with something else. But there is mixed argument all over internet I tried to go through multiple sites to get accurate info and it made it worse. Some are saying it is a big improvement don't know on what context they were claiming it. Some are very disappointed by the numbers as they were expecting 8000+ CUDA Cores on the GA100 and only getting 6000+ cores

as i said it is complicated. in the past many people just make speculation about the core configuration. the rumor has it that in FP64 A100 performance can reach around 18Tflops. from there they try to connect all other rumor and speculate that it will have around 8000 cuda cores to reach that kind of performance. but in reality nvidia 3rd generation tensor cores makes things much more complicated. nvidia has been trying to make tensor core useful beyond AI for quite sometime now. this is a good read about it:

https://devblogs.nvidia.com/tensor-cores-mixed-precision-scientific-computing/

from traditional standpoint the improvement made on A100 is not that impressive. in comparison it was like this:

GV100 (highest configuration)
FP32 16Tflops
FP64 8Tflops

A100
FP32 19.5Tflops
FP64 9.7Tflops

but on A100 once you add tensor core "acceleration" they can boost performance significantly. and this is where it gets complicated. from nvidia wording using tensor core the can boost FP32 performance to 156 tensorflops (TF). this for sure is not our traditional FP32. and in nvidia wording it can accelerate AI related things that much when they are using FP32. so in my mind that boosted performance only benefit AI related work loads. but tensor cores also boost FP64 performance. in A100 the FP64 can be boosted to 19.5Tflops. and if i understand it right all application that use FP64 can benefit from this. not just specific AI workload like it did with FP32. this is what it meant accelerating scientific computing with tensor core mixed precision. now the question is can similar boost applied on our gaming GPU?

for A100 the main star most likely the new tensor core. the new tensor core is so much more effective that nvidia actually reduced the amount of tensor core in A100 vs GV100 (432 vs 640).

 

[King Dranzer]

I went through notes from different Youtube sources and one thing I got from it is that A100 has cutoff GA100 implementation which has limited core count of 6912 CUDA Cores and 432 Tensor Cores. While the actual GA100 packs 8192 CUDA Cores and 512 Tensor Cores. That is insanely huge jump in Core count itself. With IPC improvement things get interesting. Yes the Tensor Core count decreased but the New Gen implementation and Performance increase and the change in functional utilization may be more than enough to overcome the cut-off in Numbers.
Hmmmmm.......
Lets see how RTX series ends up.

 

[King Dranzer]

as i said it is complicated. in the past many people just make speculation about the core configuration. the rumor has it that in FP64 A100 performance can reach around 18Tflops. from there they try to connect all other rumor and speculate that it will have around 8000 cuda cores to reach that kind of performance. but in reality nvidia 3rd generation tensor cores makes things much more complicated. nvidia has been trying to make tensor core useful beyond AI for quite sometime now. this is a good read about it:

https://devblogs.nvidia.com/tensor-cores-mixed-precision-scientific-computing/

from traditional standpoint the improvement made on A100 is not that impressive. in comparison it was like this:

GV100 (highest configuration)
FP32 16Tflops
FP64 8Tflops

A100
FP32 19.5Tflops
FP64 9.7Tflops

but on A100 once you add tensor core "acceleration" they can boost performance significantly. and this is where it gets complicated. from nvidia wording using tensor core the can boost FP32 performance to 156 tensorflops (TF). this for sure is not our traditional FP32. and in nvidia wording it can accelerate AI related things that much when they are using FP32. so in my mind that boosted performance only benefit AI related work loads. but tensor cores also boost FP64 performance. in A100 the FP64 can be boosted to 19.5Tflops. and if i understand it right all application that use FP64 can benefit from this. not just specific AI workload like it did with FP32. this is what it meant accelerating scientific computing with tensor core mixed precision. now the question is can similar boost applied on our gaming GPU?

for A100 the main star most likely the new tensor core. the new tensor core is so much more effective that nvidia actually reduced the amount of tensor core in A100 vs GV100 (432 vs 640).

Well from the Notes I came across I got the info that A100 packs cut-off version of GA100 which actually packs 8192 CUDA Cores and 512 Tensor Cores.
Interesting.
If all the chips through out the lineup that kind of increase in core count.
RTX3000 series will be largest jump in performance.