Friday, January 3rd 2020

NVIDIA's Next-Generation Ampere GPUs to be 50% Faster than Turing at Half the Power

As we approach the release of NVIDIA's Ampere GPUs, which are rumored to launch in the second half of this year, more rumors and information about the upcoming graphics cards are appearing. Today, according to the latest report made by Taipei Times, NVIDIA's next-generation of graphics cards based on "Ampere" architecture is rumored to have as much as 50% performance uplift compared to the previous generations of Turing GPUs, while using having half the power consumption.

Built using Samsung's 7 nm manufacturing node, Ampere is poised to be the new king among all future GPUs. The rumored 50% performance increase is not impossible, due to features and improvements that the new 7 nm manufacturing node brings. If utilizing the density alone, NVIDIA can extract at least 50% extra performance that is due to the use of a smaller node. However, performance should increase even further because Ampere will bring new architecture as well. Combining a new manufacturing node and new microarchitecture, Ampere will reduce power consumption in half, making for a very efficient GPU solution. We still don't know if the performance will increase mostly for ray tracing applications, or will NVIDIA put the focus on general graphics performance.
Source: Taipei Times
Add your own comment

180 Comments on NVIDIA's Next-Generation Ampere GPUs to be 50% Faster than Turing at Half the Power

#51
notb
Valantar
Turing was nowhere near 30% perf/SM over Pascal. More like 10%. Any further gains came from more SMs and higher clocks.
So it was or wasn't? Because I'm not sure what you mean.

I don't care about SMs, clocks and all that internal stuff (much like I don't care about IPC in CPUs). It's not what I'm paying for as a gamer.

1660Ti came out roughly 2.5 years after 1060.
It's slightly more expensive, the same power draw, similar form factor and features.
1660Ti is 30-40% faster.
Posted on Reply
#52
cucker tarlson
notb
So it was or wasn't? Because I'm not sure what you mean.

I don't care about SMs, clocks and all that internal stuff (much like I don't care about IPC in CPUs). It's not what I'm paying for as a gamer.

1660Ti came out roughly 2.5 years after 1060.
It's slightly more expensive, the same power draw, similar form factor and features.
1660Ti is 30-40% faster.
Valantar can't do simple maths
2070 super matches 1080Ti with 2560 cuda vs 3584 cuda
clocks go slightly in favor of 2070S by 100mhz (5-6%),bandwidth in favor of 1080Ti by 40GB/s (8%)
that's around 1.4x performance per CUDA on average,not 1.3x,not 1.1x
in some cases 2070s ends up 15% faster,in some a few perrcent slower.A right way to estimate it would be 1.25x-1.5x depending on a game,but certainly at least 1.3x on average
Posted on Reply
#53
HTC
notb
1660Ti came out roughly 2.5 years after 1060.
It's slightly more expensive, the same power draw, similar form factor and features.
1660Ti is 30-40% faster.
So 30 - 40% faster @ the same power draw for the 1660Ti VS 1060 but now they say you'll get even more performance uplift @ HALF the power draw for Ampere VS Turing?

See the problem?
Posted on Reply
#54
cucker tarlson
HTC
So 30 - 40% faster @ the same power draw for the 1660Ti VS 1060 but now they say you'll get even more performance uplift @ HALF the power draw for Ampere VS Turing?

See the problem?
except they're going down a node,and already to a tweaked version,with new uarch on top.
we never had this recently.
maxwell was just uarch,pascal was node,turing was uarch on same node but tweaked,with die sizes so big the power went to 250W on 104 die.
ampere will be new uarch on a small,efficient node.

I'm guessing $650-700 for the A106 die,and I'm not kidding.
mu gut tells me my 2070 Super will serve me for at least 6-10 months after this thing launches.
Posted on Reply
#55
notb
HTC
So 30 - 40% faster @ the same power draw for the 1660Ti VS 1060 but now they say you'll get even more performance uplift @ HALF the power draw for Ampere VS Turing?

See the problem?
Power consumption: new node.
Performance uplift: some from traditional improvements, (maybe) some from RTX polishing.

+30% performance and -40% power draw is the minimum we should expect.
But I won't be shocked if they managed +50%/-50%. They've been developing Ampere for a very long time.
Posted on Reply
#56
HTC
cucker tarlson
except they're going down a node,and already to a tweaked version,with new uarch on top.
we never had this recently.
maxwell was just uarch,pascal was node,turing was uarch on same node but tweaked,with die sizes so big the power went to 250W on 104 die.
ampere will be new uarch on a small,efficient node.

I'm guessing $650-700 for the A106 die,and I'm not kidding.
notb
Power consumption: new node.
Performance uplift: some from traditional improvements, (maybe) some from RTX polishing.

+30% performance and -40% power draw is the minimum we should expect.
But I won't be shocked if they managed +50%/-50%. They've been developing Ampere for a very long time.
It's possible but i still think it's very VERY unlikely.

I suppose i'll have to eat my words if i'm wrong: we shall see ...
Posted on Reply
#57
cucker tarlson
HTC
It's possible but i still think it's very VERY unlikely.

I suppose i'll have to eat my words if i'm wrong: we shall see ...
it is a rumor after all.
but the fact it's new uarch on a new,tweaked node,I don't think 50/50 is impossible.I think it's just the best case scenario. on average it'll be slightly lower.

sometimes I think people don't even read the article,they've got their comments pre-written, saved and ready to roll depending on what brand we're discussing in the news.
If utilizing the density alone, NVIDIA can extract at least 50% extra performance that is due to the use of a smaller node. However, performance should increase even further because Ampere will bring new architecture as well. Combining a new manufacturing node and new microarchitecture, Ampere will reduce power consumption in half, making for a very efficient GPU solution.
I don't know if you get what the 50/50 numbers mean too.
It means the new 3070 (a106 die if the trend continues) matches 2080Ti at ~150W.
Posted on Reply
#58
Vya Domus
For one thing I cannot see how the SM could be improved to get 50% more raster performance out of it or make it more power efficient in any significant way, they've already pulled pretty much every trick in the book with Turing, huge caches, more ILP, etc. It's just not within the realm of possibility.

With the newer node they'll either get a higher transistor count at the same speeds or get higher clocks but they'll have to maintain more or less the same core counts to stay within the same power. Suffice to say they'll go with the former, so 50% less power from the node is out of the question. Nothing can possibly explain 50% more performance for 50% less power or anywhere close to that, it simply isn't doable. There are two possibilities here, rumors are wrong, or this comes with a million caveats involving RTX, DLSS, variable shading and whatever else they come up with to mitigate the cost of all the extra silicon and justify increased prices.
Posted on Reply
#59
64K
I'm a gamer. I buy GPUs to play games. I don't care about Synthetics. I care about real-world performance in gaming and going from reviews here I find that the original RTX 2070 on average was 39% faster than it's Pascal counterpart GTX 1070. The test bed of games are mostly not raytracing games and even if they had some element of raytracing you could just turn it off.

The original RTX 2080 was on average 37% faster than it's Pascal counterpart GTX 1080.

The RTX 2080 Ti was on average 39% faster than it's Pascal counterpart GTX 1080 Ti.

Granted the Turings are using around 50 more watts than the Pascals.

Now if Nvidia can make those gains going from the 16nm process to the 12nm process then why is a 50% performance increase for the same watts as Turing not possible going from the 12nm process to the 7nm process?

Remember that I'm not comparing watts used by Ampere to Pascal. I'm comparing watts used by Ampere to Turing.

I think a 50% increase in performance by Ampere over Turing using the same watts is very possible.

Posted on Reply
#60
cucker tarlson
Vya Domus
For one thing I cannot see how the SM could be improved to get 50% more raster performance out of it or make it more power efficient in any significant way, they've already pulled pretty much every trick in the book with Turing, huge caches, more ILP, etc. It's just not within the realm of possibility.

With the newer node they'll either get a higher transistor count at the same speeds or get higher clocks but they'll have to maintain more or less the same core counts to stay within the same power. Suffice to say they'll go with the former, so 50% less power from the node is out of the question. Nothing can possibly explain 50% more performance for 50% less power or anywhere close to that, it simply isn't doable. There are two possibilities here, rumors are wrong, or this comes with a million caveats involving RTX, DLSS, variable shading and whatever else they come up with to mitigate the cost of all the extra silicon and justify increased prices.
for ampere they've had good foundations to build on,a narrow focus,plenty of time and huge rd budget.
Posted on Reply
#61
HwGeek
64K
I'm a gamer. I buy GPUs to play games. I don't care about Synthetics. I care about real-world performance in gaming and going from reviews here I find that the original RTX 2070 on average was 39% faster than it's Pascal counterpart GTX 1070. The test bed of games are mostly not raytracing games and even if they had some element of raytracing you could just turn it off.

The original RTX 2080 was on average 37% faster than it's Pascal counterpart GTX 1080.

The RTX 2080 Ti was on average 39% faster than it's Pascal counterpart GTX 1080 Ti.

Granted the Turings are using around 50 more watts than the Pascals.

Now if Nvidia can make those gains going from the 16nm process to the 12nm process then why is a 50% performance increase for the same watts as Turing not possible going from the 12nm process to the 7nm process?

Remember that I'm not comparing watts used by Ampere to Pascal. I'm comparing watts used by Ampere to Turing.

I think a 50% increase in performance by Ampere over Turing using the same watts is very possible.
you got those uplift because NV tricked all of us and changed the names of the cards, in real life you should compare 1070 vs 2060, 1080 vs 2070, 1080Ti vs 2080.
the 2080Ti isn't 1080Ti replacement, it's the Titan, that why the prices shifted.
Posted on Reply
#62
DeeJay1001
TheDeeGee
50% Speed, Half the Power, double the cost.

That's Nvidia alright!
More like
50% more speed (in very specific use cases)
Half the power (in extremely limited lab test scenarios)
Double the cost ALL THE TIME
Posted on Reply
#63
64K
HwGeek
you got those uplift because NV tricked all of us and changed the names of the cards, in real life you should compare 1070 vs 2060, 1080 vs 2070, 1080Ti vs 2080.
the 2080Ti isn't 1080Ti replacement, it's the Titan, that why the prices shifted.
I'm not going to derail this thread with a lengthy discussion on Nvidia naming conventions, memory bus width etc. but if the RTX 2080 Ti is the replacement for the Titan Xp then what is the RTX Titan? Don't just look at pricing. That's not how to compare a counterpart with Nvidia.

Also my best guess is that we will see a 50% increase in performance with Ampere over Turing using the same watts across the board but that's just speculation on my part until the cards actually wind up in W1zzard's hands to bench and then we will know for sure.
Posted on Reply
#64
lexluthermiester
R0H1T
Yes & pigs will also fly when that happens :rolleyes:
As a general rule, when NVidia makes a claim, they come very close to it if not actually achieving it.
TheDeeGee
double the cost.
That's not going to happen. The price is very likely to come down..
64K
Don't just look at pricing. That's not how to compare a counterpart with Nvidia.
Exactly right. Prices fluctuate and are not a reliable indicator of generational steppings, but replacement model naming conventions generally are.
Posted on Reply
#65
Valantar
Xaled
Then then the title should've included words as : Claim or say ( Nvidia says, or claims) at worst, just like did in the Intel news.
Saying that IT WILL BE %50 FASTER this way is just wrong and biased.
I'd say the headline ought to be along the lines of "Nvidia's Next-Generation Ampere GPUs reportedly 50% faster than Turing at Half the Power". I'll agree that a headline with zero reservations or considerations taken as to this being second-hand information with dubious sourcing making bombastic claims at a very early point is poor journalism, but I don't see it as a double standard or bias - inaccuracies like this are quite par for the course across the board for TPU, sadly.
Tsukiyomi91
if Ampere really has a 50% gain over Turing in all benchmarks/real world use while using less power is a good thing. Problem here is many bought the "refreshed" RTX20 Series Super cards & GTX16 Series cards... so those folks might be at a loss-ish?
Logic like that is always wrong. If you buy a card today and a better card comes out for the same price tomorrow, you have lost nothing whatsoever. Sure, you could have gotten a better deal if you waited, but that is always the case. There will always be better hardware in the future for a better price, so you just have to pull the trigger at some point, and your purchase will always be made to look less than ideal at some point. That doesn't mean it was a bad purchase, nor does it change the performance/dollar that you tacitly agreed to pay when you made the purchase decision.
notb
So it was or wasn't? Because I'm not sure what you mean.

I don't care about SMs, clocks and all that internal stuff (much like I don't care about IPC in CPUs). It's not what I'm paying for as a gamer.

1660Ti came out roughly 2.5 years after 1060.
It's slightly more expensive, the same power draw, similar form factor and features.
1660Ti is 30-40% faster.
The problem is that you weren't talking about absolute performance in the post I responded to, you were talking about architectural performance improvements specifically. While there are (very broadly) two ways for these to work (increased clock speeds not due to node changes, and "IPC" for lack of a better term), most clock speed improvements come from node changes, and most arch improvements are normally down to improving IPC. There are exceptions to this, such as Maxwell clocking significantly higher than previous architectures, but for the most part this holds true. If you're talking about perf/$ on an absolute level, you are right, but that's another matter entirely. So, if you don't care about how one arrives at a given performance level, maybe don't get into discussions about it?

cucker tarlson
Valantar can't do simple maths
2070 super matches 1080Ti with 2560 cuda vs 3584 cuda
clocks go slightly in favor of 2070S by 100mhz (5-6%),bandwidth in favor of 1080Ti by 40GB/s (8%)
that's around 1.4x performance per CUDA on average,not 1.3x,not 1.1x
in some cases 2070s ends up 15% faster,in some a few perrcent slower.A right way to estimate it would be 1.25x-1.5x depending on a game,but certainly at least 1.3x on average
2070S is a tuned and binned half-gen refresh SKU based on mature silicon, not really directly comparable to the 1st-gen 1080Ti. That doesn't mean it doesn't have good absolute performance, it just isn't a 1:1 comparison. The fact that the 2070S performs within a few percent of the 2080 makes this pretty clear. So, if you want a like-for-like comparison, the 2080 is the correct card. And then you suddenly have a rather different equation:
2080 beats the 1080Ti by about 8% with 2944 CUDA cores vs. 3584.
Clocks go slightly in favor of the 2080 by 100MHz (5-6%), and memory bandwidth is essentially a tie.
In other words the clock speed increase and performance increase pretty much cancel each other out, leaving us with ~22% more perf/CU. That of course ignores driver improvements and other uncontrollable variables, to which some improvements should reasonably be attributed. My 10% number might have been a bit low, but your 40% number is silly high.
Posted on Reply
#66
cucker tarlson
Valantar
2070S is a tuned and binned half-gen refresh SKU
how is 2070S more tuned and binned than 1080ti really ? that's a made up point.
pointless discussion.

bandwidth on 1080ti is 8% higher than 2080,like 2070S,they share 14bgps 256-bit ddr6.
Posted on Reply
#67
Valantar
lexluthermiester
As a general rule, when NVidia makes a claim, they come very close to it if not actually achieving it.
...but this isn't Nvidia making a claim, it's some analyst from an investment consulting agency.

cucker tarlson
how is 2070S more tuined and binned than 1080ti really ?
By being a mid-cycle release with a differently binned die? That's kind of how this works.
cucker tarlson
if you don't want to compare 2070s,comparre 2080.you'll end up with the same numbers.
...that's what I did?
cucker tarlson
bandwidth on 1080ti is 8% higher than 2080,like 2070S,they share 14bgps 256-bit ddr6.
Sorry, got that mixed up what with all the Turing SKUs out there. Doesn't make much of a difference though. 22%+8%= 23,8%. Nonetheless nowhere near 40.
Posted on Reply
#68
xkm1948
Nvidia is not Intel. This is nice
Posted on Reply
#69
cucker tarlson
Valantar
...but this isn't Nvidia making a claim, it's some analyst from an investment consulting agency.


By being a mid-cycle release with a differently binned die? That's kind of how this works.

...that's what I did?

Sorry, got that mixed up what with all the Turing SKUs out there. Doesn't make much of a difference though. 22%+8%= 23,8%. Nonetheless nowhere near 40.
1080ti is not first pascal either,it is mid-cycle too.

no,you didn't get the numbers correctly.if you did you'd end up with 1.32x not 1.22x.that's 0.08 (8%) from 2070S at 40%.Still,nowhere near your 10% and more than 30%.
Posted on Reply
#70
the54thvoid
Well, without using any maths or forum pseudo-engineering, it's quite probable, at such a die shrink with a better arch, the chips will be a generational leap, like we used to see. 50%, perhaps in best of circumstances. But despite the naysayers, Nvidia will hit it out the park, again. And yeah, it'll cost a fortune.
Posted on Reply
#71
phanbuey
ZoneDymo
^ this, I was just going to say the same thing.

Also, if any of this is true, Im not expecting prices to return to normal any time soon....in fact it will probably just get worse.
I think the prices will depend more on AMD than anything -- it's actually looking like a GPU war is on the horizon (about time).
Posted on Reply
#73
SaLaDiN666
God, I love the internet, a guy writes down complete FUD about 10%, compares 1080ti with 2080 instead of 2080Ti and then he even feels entitled to pester others how their more realistic estimations are silly. What a time to be alive.
Posted on Reply
#74
tomc100
I'll believe it when I see it.
Posted on Reply
Add your own comment