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

[candle_86]

oh no, where my strix 2080ti gonna head to? sub $400 after new gen?

Yes before going worthless then it will slowly rise in value in 30 years

 

[DeathtoGnomes]

Yes & pigs will also fly when that happens

Reminds me of Intel's PR these days with multiple ***

is that RTX ON?

 

[R0H1T]

Probably, will have to check with JHH once ~ ($)2k cards here we come

 

[Fluffmeister]

Hey hey logic and reasoning is offensive to my gut feelings and love for certain brands, take it back! /s

On a serious note, tech progression don’t give a flying fk about fans. Nvidia dominates GPU accelerated computing and AI in industry and academia. Nvidia has good reason to keep pushing large gen to gen performance boost. This is the best way to entrench themselves in the lucrative market.

I hear you fella, it's amusing reading this thread and the highly aggro and defensive comments (even the ones that have recently been deleted), becuase the shocking truth is Nvidia haven't even gone 7nm yet and they don't like that fact.

Hey, we all wish the fastest cards you can buy were cheaper, but those cards need competition to force the prices down.

All else is tears in the rain.

 

[prnsforum]

The bad is price 100% uplift

 

[Midland Dog]

it's gonna be feliz navi-dead for amd when ampere launches

nah,but seriously,I'm confused by this tsmc 7nm mention.
would not be surprised if smaller a107/108 dies went for tsmc,they segmented their production for big and small pascals between tsmc and glofo too IIRC.I bet rdna2 desktop cards would not like to see 3050 and 3060 production get in their way too.

rdna2 parts might end up competing with the 3050 and 3060 the way RTG seems to be going

 

[Aerpoweron]

I can imagine a 50% speed uplift in a best case scenario, or a 50% less power draw in such a scenario. But not both at the same time. We probably don't see this in the consumer market soon, since it will be most likely focused on the AI and Datacenter markets.

And for the competition from AMD. Keep in mind, AMD is focusing on CPUs, maybe a little on the GPUs for the consoles. AMD just has to do good enough in the GPU segment for now. They just don't have enough resources to go like Ryzen on the GPU front. Most of the money is made in the mid range GPU segment. I think AMD has some good offers there to compete with Nvidia.
If AMD should focus on GPUs in the near future, it will get very interesting

 

[Vayra86]

this is shameless promotion from "gurustud" ( )
a guy youtubing his commentary for TPU posts,how brave of him to avoid getting into a discussion.I bet that would go well for him.

We're all haters man! Like, OMG, really... *add sweet boy voice*

YouTube will put out (just about) anything people... just remember that.

Anything. Even avrona has a yt channel...I'm not surprised to see some other neophyte at it... is anyone?

Double standard my ass... get your head out of the sand.

Its almost like TV

No wait, its worse, its TV without filters

is that RTX ON?

RTX ON means the pigs will blot out the sun!

While I don't disagree with what you're saying entirely, claiming that 7nm EUV is a full node up from 7nm DUV is... dubious. It is definitely an improved node, but not by that much.

Well, it is kind of the proposed ultimate form of 7nm since ASML started doing EUV. DUV was never really cost effective or fantastic in the first place, its just a placeholder to ramp up 7nm. That is also why you don't see that much gain in absolute performance per watt; the major gain from 7nm DUV is density.

Its more accurate to say 7nm EUV is a real 7nm node and DUV is a bandaid.

 

[cucker tarlson]

We're all haters man! Like, OMG, really... *add sweet boy voice*



Its almost like TV

No wait, its worse, its TV without filters



RTX ON means the pigs will blot out the sun!



Well, it is kind of the proposed ultimate form of 7nm since ASML started doing EUV. DUV was never really cost effective or fantastic in the first place, its just a placeholder to ramp up 7nm. That is also why you don't see that much gain in absolute performance per watt; the major gain from 7nm DUV is density.

Its more accurate to say 7nm EUV is a real 7nm node and DUV is a bandaid.

this guy has done very poor research to begin with.

like he begins with "navi is better than polaris" since it competes with tu104 not like polaris competed with gtx1060.
this is plain wrong in the opening assumptions.
full tu106 is 2070,and that's what 5700xt is head to head against.full tu104 with 3072 cuda is 25% faster at least.
so navi goes against the same 106 die,but with a node down advantage this time.

 

[ppn]

N5P in 2020/21 uses more EUV layers than N7+ and provides 33% better performance and 45% smaller die size than N7. So I might as well wait for that. it all seems like temporary nodes to me, N7,N7P,N6,N7+,N5,N5P and what not.

 

[cucker tarlson]

N5P in 2020/21 uses more EUV layers than N7+ and provides 33% better performance and 45% smaller die size than N7. So I might as well wait for that. it all seems like temporary nodes to me, N7,N7P,N6,N7+,N5,N5P and what not.

5nm+ cards are not coming in 2020 or 2021.

 

[ppn]

N7 risk production Q2/2017, 5700XT Q3/2019, 3080 Q3/2020 +50% perf or -50% power
N5 risk production Q2/2019, 6700XT Q3/2021, 4080 Q3/2022 +25% perf or -50% power

 

[efikkan]

this guy has done very poor research to begin with.

like he begins with "navi is better than polaris" since it competes with tu104 not like polaris competed with gtx1060.
this is plain wrong in the opening assumptions.
full tu106 is 2070,and that's what 5700xt is head to head against.full tu104 with 3072 cuda is 25% faster at least.
so navi goes against the same 106 die,but with a node down advantage this time.

Just because someone is able to post a YouTube video about a technical topic doesn't make them an authority on the subject matter, any fool can sit in their basement and make a video where they just ramble about "technical" stuff.

"Why Nvidia Is Stuck with Tensor/RT till 2021"
Well, at least to me this title really says it all. It's really incredible how clueless and misguided some of these opinionators can be, even lacking the ability to do logical reasoning. There is just so much wrong here I don't even know where to begin. But let's just say that Nvidia have been heavily investing into RT for over a decade. What we see in Turing was not thrown quickly together, and when their next generation arrives, it will take it quite a bit further, it's not like they are "stuck" with some "poor" implementation.

I would be much more concerned about AMD's RT ventures. They have obviously been researching too, but were caught off guard when Turing brought it to market so quickly. It will be really interesting to see if they have been able to integrate it properly in such a short timeframe.

 

[Vayra86]

N7 risk production Q2/2017, 5700XT Q3/2019, 3080 Q3/2020 +50% perf or -50% power
N5 risk production Q2/2019, 6700XT Q3/2021, 4080 Q3/2022 +25% perf or -50% power

Holy crap so by 2022 we get 175% performance at 0% power!!!

I can't wait!

 

[Midland Dog]

maybe
Ampere Compute Series
TSMC 7EUV
GA100 8GPC*8TPC*2SM 6144bit
GA101 4GPC*8TPC*2SM 3072bit
Ampere Gaming Series
SAMSUNG 8EUV
GA102 7GPC*6TPC*2SM 384bit NVLINK
GA103 6GPC*5TPC*2SM 320bit
GA104 6GPC*4TPC*2SM 256bit
GA106 3GPC*5TPC*2SM 192bit
GA107 2GPC*5TPC*2SM 128bit
ganked from twitter and according to my numbers that makes ga100 an 8192 cuda core part

 

[cucker tarlson]

Just because someone is able to post a YouTube video about a technical topic doesn't make them an authority on the subject matter, any fool can sit in their basement and make a video where they just ramble about "technical" stuff.

"Why Nvidia Is Stuck with Tensor/RT till 2021"
Well, at least to me this title really says it all. It's really incredible how clueless and misguided some of these opinionators can be, even lacking the ability to do logical reasoning. There is just so much wrong here I don't even know where to begin. But let's just say that Nvidia have been heavily investing into RT for over a decade. What we see in Turing was not thrown quickly together, and when their next generation arrives, it will take it quite a bit further, it's not like they are "stuck" with some "poor" implementation.

I would be much more concerned about AMD's RT ventures. They have obviously been researching too, but were caught off guard when Turing brought it to market so quickly. It will be really interesting to see if they have been able to integrate it properly in such a short timeframe.

yup.exactly.I don't get the "stuck" part either.

maybe
Ampere Compute Series
TSMC 7EUV
GA100 8GPC*8TPC*2SM 6144bit
GA101 4GPC*8TPC*2SM 3072bit
Ampere Gaming Series
SAMSUNG 8EUV
GA102 7GPC*6TPC*2SM 384bit NVLINK
GA103 6GPC*5TPC*2SM 320bit
GA104 6GPC*4TPC*2SM 256bit
GA106 3GPC*5TPC*2SM 192bit
GA107 2GPC*5TPC*2SM 128bit
ganked from twitter and according to my numbers that makes ga100 an 8192 cuda core part

tbh I was expecting 320-bit parts.
what are gpc/tpc ?

 

[Midland Dog]

yup.exactly.I don't get the "stuck" part either.


tbh I was expecting 320-bit parts.
what are gpc/tpc ?

Graphics Processing Clusters and Texture Processing Clusters, my numbers assume a volta esque partioning of sms tho

Graphics Processing Clusters and Texture Processing Clusters, my numbers assume a volta esque partioning of sms tho

if any thing changes in the sm its likely to be cache wise, with a chip that wide cache and rops will be the biggest concern, nv engineers will cook up some new compression method to get further on the same if not slightly higher (14-16gbps) speeds

Graphics Processing Clusters and Texture Processing Clusters, my numbers assume a volta esque partioning of sms tho


if any thing changes in the sm its likely to be cache wise, with a chip that wide cache and rops will be the biggest concern, nv engineers will cook up some new compression method to get further on the same if not slightly higher (14-16gbps) speeds

assuming a 200-300mhz boost as aibs have reported, GA100 could be a 36 teraflop part at 2.2ghz, amdead

 

[cucker tarlson]

Graphics Processing Clusters and Texture Processing Clusters, my numbers assume a volta esque partioning of sms tho

hard to predict the CUDA count based on that since they use various SM per GPC count.

 

[Midland Dog]

hard to predict the CUDA count based on that since they use various SM per GPC count.

i still expect 64, then again intel is going with smaller partitions with the dg1 design maybe theres merit to it

 

[cucker tarlson]

i still expect 64, then again intel is going with smaller partitions with the dg1 design maybe theres merit to it

sm per gpc,not cuda per sm

 

[Midland Dog]

sm per gpc,not cuda per sm

cuda per sm fp32 and int32

 

[Super XP]

50%? I don't see that.
The GTX 980ti was approx: 34% faster over the GTX 780ti in 12 games on average.
The GTX 1080ti was approx: 23% faster over the GTX 980ti in 12 games on average.
The RTX 2080ti was approx: 25% faster over the GTX 1080ti in 12 games on average.
The RTX 3080ti to be approx: 50% faster over the RTX 2080ti on average? I Don't Think So, lol

What big Navi? Released when?

Big Navi is a rumor at this point, just like Ampere or whatever name Nvidia may call it.

Big Navi is not a rumour as it was confirmed by Dr. Lisa Su and its coming. The question that remains is when? Looking at AMD's past announcements and the fact Next Generation Gaming Consoles from Micro$oft and Sony are coming Christmas 2020, I can see AMD launching RDNA2 and ZEN3 at Computex 2020, at the same time, which availability shortly there after.

 

[64K]

Let's look at W1zzard's benches for some perspective on reality.

Using 1440p as a base for comparison. 4K often offers even more of a performance boost over the previous generation in some cases but I'm trying to keep this simple. Max quality settings.

GTX 980 Ti using a 21 games suite for benching showed an average of 41% increase in performance over a GTX 780 TI

GTX 1080 Ti using a 22 game suite for benching showed an average of 75% increase in performance over a GTX 980 Ti

RTX 2080 Ti using a 23 game suite for benching showed an average 33% increase in performance over a GTX 1080 Ti

Now with Ampere going from the 12nm process node to the 7nm process node and the resultant increase in efficiency and a new architecture then why in the world can't there be a 50% increase in performance over the Turing GPUs?

I have put all of the Relative Performance percentages behind the spoiler below. Please don't subtract performances from each other. That's not how percentages work. If GPU #1 is rated as 100% and GPU #2 is rated at 50% then that doesn't mean that GPU #1 is 50% faster than GPU #2. It means that GPU #1 is 100% faster or twice as fast as GPU #2

Simply divide GPU #1 by GPU #2 to get a number.

i.e. GPU #1 is rated at 100% and GPU #2 is rated at 75% then 100/75=1.33 which means that GPU #1 is 33% faster than GPU #2. If you doubt that then multiply GPU #2 at 75% times 1.25 and you get 93.75 and not 100.

Spoiler: Relative Performance One Generation Over The Last

 

[Fluffmeister]

50%? I don't see that.
The GTX 980ti was approx: 34% faster over the GTX 780ti in 12 games on average.
The GTX 1080ti was approx: 23% faster over the GTX 980ti in 12 games on average.
The RTX 2080ti was approx: 25% faster over the GTX 1080ti in 12 games on average.
The RTX 3080ti to be approx: 50% faster over the RTX 2080ti on average? I Don't Think So, lol

There is no way a GTX 1080 Ti was just an average of 23% faster than a GTX 980 Ti, which was the last big node change for Nvidia.

Looking at W1zz review (over 22 games) comparing reference cards:

At 1080P the 1080 Ti is 53% faster:

At 1440P the 1080 Ti is 75% faster:

And at 4K the 1080 Ti is 85% faster:

 

[Super XP]

I was basing the results on UserBenchmark results. As they compare 1,000's of gaming PCs.