XFX Radeon RX 5600 XT THICC 3 review
Fractal Design Define 7 XL review
Guru3D Rig of the Month - January 2020
Palit GeForce RTX 2070 SUPER GamingPro review
Team Group T-Force DarkZα 3600 MHz DDR4 review
Corsair Scimitar Elite RGB gaming mouse review
Cougar 700K Evo keyboard review
Aorus Liquid Cooler 360 review
AMD Ryzen Threadripper 3990X review
Corsair A500 CPU air cooler review
Nvidia Slide reveals numbers on Single and Double precision for Flagship Pascal GPU
A slide from an Nvidia presentation is stirring some things on the web as it seems as some info on the Flagship GPU on the new 16nm Pascal architecture (likely named GP100) would be able to perform numbers that are astonishing.
Slides from "The Future of HPC and The Path to Exascale" shows a roadmap with a DP GFLOPS/W value for Pascal. The presentation's date is between the GTC 2013 roadmap which does not contain Pascal and the GTC 2014 roadmap which does contain Pascal. But thus this presentation is dated (2015) and likely estimated, we do have to say that.
The slides from "Manuel Ujaldón CUDA Fellow @ Nvidia" however show performance numbers for both single and double precision, double-precision floating-point (DPFP) wise Nvidia seems to be reaching (or is aiming for) a 4 TFLOP/s throughput (at least on their HPC parts). That would be 3x over the current 1.3 TFLOP/s on the Tesla K20 which on it's end is based on "Kepler" GK110 silicon.
Single-precision then, it would be as high as 12 TFLOP/s. That is four times that a GK110, and roughly double a GM200 (6.4 TFLOPS for the 980 Ti). The slide does reveal one other thing, the GP100 is inidicated to use stacked HBM2 memory as the memory bandwidth is set at 1 TB/s.
Download the presentation here. Well .. yummie !?
Nvidia Shield Tablet K1 gets update towards Android 6.0 - 12/22/2015 09:36 AM
The Nvidia Shield Tablet K1 will be updated towards Android 6.0, the over the air update will bring the OS to Android 6.0 Marshmallow. The update update package is not yet available as standalone dow...
Nvidia silently launches GeForce GTX 960M, GTX 950M and GeForce 940M - 03/13/2015 07:45 AM
Nvidia silently has released their GeForce GTX 960M, GTX 950M and GeForce 940M mobile chips. The laptop GPUs bring a few features to enhance gaming experience and offer BatteryBoost for longer playtim...
Nvidia silently launches GeForce GTX 965M for laptops - 01/08/2015 09:14 AM
Without any noise Nvidia has launched a new mobile GPU, the GeForce GTX 965M for laptops. The GTX 965M is actually based upon a Maxwell GM204 GPU, yet with less shader processors compared to the 970M...
Samsung Files Complaint to Block Nvidia Selling Chips in the USA - 11/24/2014 09:28 AM
Samsung Electronics is seeking to block computer-graphics chips made by Nvidia Corp. (NVDA) from the U.S. market, escalating a battle begun after licensing talks failed....
Nvidia Starts selling Sturdy SLI Bridges with LED lighting - 09/25/2014 10:54 AM
Nvidia is now selling SLI Bridges with LED lighting. The sturdy version of these bridges can only be purchased through their website and seems to be limited to the USA region. The bridges are availab...
Denial
Senior Member
Posts: 12567
Joined: 2004-05-16
Senior Member
Posts: 12567
Joined: 2004-05-16
#5233133 Posted on: 02/17/2016 02:54 PM
For big Pascal the numbers don't seem that unreasonable given that they essentially skipped a generation with double precision. I doubt a card this big will launch this year though.
I wonder how Pascal is going to work anyway. Are they going to make gaming cards full precision, with cut down double like they do now? or are the gaming cards going to have mixed precision on them along with double and full? Isn't that like a waste of die space?
Maxwell doesn't have a card that has dedicated DP on it -- they are all cut down.
For big Pascal the numbers don't seem that unreasonable given that they essentially skipped a generation with double precision. I doubt a card this big will launch this year though.
I wonder how Pascal is going to work anyway. Are they going to make gaming cards full precision, with cut down double like they do now? or are the gaming cards going to have mixed precision on them along with double and full? Isn't that like a waste of die space?
Interesting that Maxwell's not mentioned in that slide but they show Fermi and Kepler... Wonder how this would rank, who knows, maybe very close to Pascal.
Maxwell doesn't have a card that has dedicated DP on it -- they are all cut down.
Krteq
Senior Member
Posts: 543
Joined: 2009-09-03
Senior Member
Posts: 543
Joined: 2009-09-03
#5233137 Posted on: 02/17/2016 02:56 PM
It's an estimated computing performance taken from slides from June 2015.
It's an estimated computing performance taken from slides from June 2015.
fantaskarsef
Senior Member
Posts: 11262
Joined: 2014-07-21
Senior Member
Posts: 11262
Joined: 2014-07-21
#5233141 Posted on: 02/17/2016 02:59 PM
Of course you are right, totally forgot about that!
Maxwell doesn't have a card that has dedicated DP on it -- they are all cut down.
Of course you are right, totally forgot about that!
Denial
Senior Member
Posts: 12567
Joined: 2004-05-16
Senior Member
Posts: 12567
Joined: 2004-05-16
#5233144 Posted on: 02/17/2016 03:03 PM
Yeah, it says that in the article...
It's an estimated computing performance taken from slides from June 2015.
Yeah, it says that in the article...
Noisiv
Senior Member
Posts: 6664
Joined: 2010-11-16
Senior Member
Posts: 6664
Joined: 2010-11-16
#5233166 Posted on: 02/17/2016 03:45 PM
Not bad, but these are pro cards, operating 24/7
SP flops need to be multiplied by
Volta (Sierra/Summit) as early as 2017.
https://www.olcf.ornl.gov/summit/
Not bad, but these are pro cards, operating 24/7
SP flops need to be multiplied by
~1.25
to account for higher clocks on consumer version.
Volta (Sierra/Summit) as early as 2017.
https://www.olcf.ornl.gov/summit/
svictorcc
Junior Member
Posts: 10
Joined: 2015-06-29
Junior Member
Posts: 10
Joined: 2015-06-29
#5233210 Posted on: 02/17/2016 04:39 PM
Imho, and considering some other infos, i'm almost sure Pascal GP100 will launch this year and before GP104.
Imho, and considering some other infos, i'm almost sure Pascal GP100 will launch this year and before GP104.
AnarchoPrimitiv
Junior Member
Posts: 3
Joined: 2015-10-03
Junior Member
Posts: 3
Joined: 2015-10-03
#5233249 Posted on: 02/17/2016 06:12 PM
14nm? I thought it was supposed to be 16nm?
Pascal is 14nm? I thought it was supposed to be 16nm (and Polaris is supposed to be 14nm)? That's what all the news on other sites has been confirming for the last several months.
14nm? I thought it was supposed to be 16nm?
Pascal is 14nm? I thought it was supposed to be 16nm (and Polaris is supposed to be 14nm)? That's what all the news on other sites has been confirming for the last several months.
Lane
Senior Member
Posts: 6361
Joined: 2005-02-25
Senior Member
Posts: 6361
Joined: 2005-02-25
#5233268 Posted on: 02/17/2016 06:47 PM
A slide from an Nvidia presentation is stirring some things on the web as it seems as some info on the Flagship GPUÂ*in the new 14nm Pascal architecture (likely GP100) would be able to perform numbers...
Nvidia Slide reveals numbers on Single and Double precision for Flagship pascal GPU
Damn old slide, it seems ... June 2015 and based on even older projection.
Not that thoses number are not believable. its around 50% SP gain.
This said the author is not part of Nvidia. ( even if it cite Nivida engineer )
A slide from an Nvidia presentation is stirring some things on the web as it seems as some info on the Flagship GPUÂ*in the new 14nm Pascal architecture (likely GP100) would be able to perform numbers...
Nvidia Slide reveals numbers on Single and Double precision for Flagship pascal GPU
Damn old slide, it seems ... June 2015 and based on even older projection.
Not that thoses number are not believable. its around 50% SP gain.
This said the author is not part of Nvidia. ( even if it cite Nivida engineer )
geogan
Senior Member
Posts: 466
Joined: 2010-01-04
Senior Member
Posts: 466
Joined: 2010-01-04
#5233272 Posted on: 02/17/2016 06:52 PM
From what I know Nvidia cripped or removed double precision form their consumer level Geforce cards so they didn't cut into lucrative sales of their way more expensive pro cards, but AMD did not do this.
Hence my humble dual GPU HD7990 card has a DP performance of over 2000 whereas the Tesla K20X only has 1310 and how much is a Tesla K20X these days? Over two grand? and originally costing?
From what I know Nvidia cripped or removed double precision form their consumer level Geforce cards so they didn't cut into lucrative sales of their way more expensive pro cards, but AMD did not do this.
Hence my humble dual GPU HD7990 card has a DP performance of over 2000 whereas the Tesla K20X only has 1310 and how much is a Tesla K20X these days? Over two grand? and originally costing?
Denial
Senior Member
Posts: 12567
Joined: 2004-05-16
Senior Member
Posts: 12567
Joined: 2004-05-16
#5233280 Posted on: 02/17/2016 07:11 PM
From what I know Nvidia cripped or removed double precision form their consumer level Geforce cards so they didn't cut into lucrative sales of their way more expensive pro cards, but AMD did not do this.
Hence my humble dual GPU HD7990 card has a DP performance of over 2000 whereas the Tesla K20X only has 1310 and how much is a Tesla K20X these days? Over two grand? and originally costing?
I think it's more apt to compare the HD7990 to the Titan Black, which was also $1000, is only ~150Gflops less DP performance, but ~100w less power - though it did launch a year later.
Nvidia shifted away from DP performance because most of the HPC market shifted towards neural networks, which perform best in mixed precision. I mean yeah, differentiating their products was also probably a large factor, but it's not like Maxwell even had a DP card.
That's why I'm kind of curious to see what they do with Pascal. They are going offer mixed (FP16) full (FP32) and Double (FP64). I wonder if they will come out with a card marketed specifically towards neural networks with FP16/32, a gaming card with FP32, and a workstation card with 32/64.
From what I know Nvidia cripped or removed double precision form their consumer level Geforce cards so they didn't cut into lucrative sales of their way more expensive pro cards, but AMD did not do this.
Hence my humble dual GPU HD7990 card has a DP performance of over 2000 whereas the Tesla K20X only has 1310 and how much is a Tesla K20X these days? Over two grand? and originally costing?
I think it's more apt to compare the HD7990 to the Titan Black, which was also $1000, is only ~150Gflops less DP performance, but ~100w less power - though it did launch a year later.
Nvidia shifted away from DP performance because most of the HPC market shifted towards neural networks, which perform best in mixed precision. I mean yeah, differentiating their products was also probably a large factor, but it's not like Maxwell even had a DP card.
That's why I'm kind of curious to see what they do with Pascal. They are going offer mixed (FP16) full (FP32) and Double (FP64). I wonder if they will come out with a card marketed specifically towards neural networks with FP16/32, a gaming card with FP32, and a workstation card with 32/64.
Lane
Senior Member
Posts: 6361
Joined: 2005-02-25
Senior Member
Posts: 6361
Joined: 2005-02-25
#5233303 Posted on: 02/17/2016 08:04 PM
I think it's more apt to compare the HD7990 to the Titan Black, which was also $1000, is only ~150Gflops less DP performance, but ~100w less power - though it did launch a year later.
Nvidia shifted away from DP performance because most of the HPC market shifted towards neural networks, which perform best in mixed precision. I mean yeah, differentiating their products was also probably a large factor, but it's not like Maxwell even had a DP card.
That's why I'm kind of curious to see what they do with Pascal. They are going offer mixed (FP16) full (FP32) and Double (FP64). I wonder if they will come out with a card marketed specifically towards neural networks with FP16/32, a gaming card with FP32, and a workstation card with 32/64.
I dont know if your report is wrong, but is really a majority of the HPC market, so a majority of HPC servers are really dedicated on Neural Network analysis ?
This a new growing market, but im not really sure that a majority of HPC servers are now dedicated to research about it. Big data, scientist market, medical research.. i mean, HPC servers are used in nearly one thousand of different " research " type.
When discussing about it, on BY3d as example, if important in term of research, it was still a niche market in term of servers and center dedicated to it.
Then come the link with Maxwell, ... neural network computing was not even really used when Maxwell 1 and 2 ( 980 have launch in 09/2014, the maxwell 1 is even older ) .. so even if this arch was developped in something like 2012 ... neural network is too much new, for make it a cause of not have FP64 ..
The thing is Nvidia was more need an architecture linked to mobile, power efficient, for bring it in the Tegra line up.
Its even a complete contradiction with pascal return to FP64.. ( Now that Neural network research becomre way more important )
I think it's more apt to compare the HD7990 to the Titan Black, which was also $1000, is only ~150Gflops less DP performance, but ~100w less power - though it did launch a year later.
Nvidia shifted away from DP performance because most of the HPC market shifted towards neural networks, which perform best in mixed precision. I mean yeah, differentiating their products was also probably a large factor, but it's not like Maxwell even had a DP card.
That's why I'm kind of curious to see what they do with Pascal. They are going offer mixed (FP16) full (FP32) and Double (FP64). I wonder if they will come out with a card marketed specifically towards neural networks with FP16/32, a gaming card with FP32, and a workstation card with 32/64.
I dont know if your report is wrong, but is really a majority of the HPC market, so a majority of HPC servers are really dedicated on Neural Network analysis ?
This a new growing market, but im not really sure that a majority of HPC servers are now dedicated to research about it. Big data, scientist market, medical research.. i mean, HPC servers are used in nearly one thousand of different " research " type.
When discussing about it, on BY3d as example, if important in term of research, it was still a niche market in term of servers and center dedicated to it.
Then come the link with Maxwell, ... neural network computing was not even really used when Maxwell 1 and 2 ( 980 have launch in 09/2014, the maxwell 1 is even older ) .. so even if this arch was developped in something like 2012 ... neural network is too much new, for make it a cause of not have FP64 ..
The thing is Nvidia was more need an architecture linked to mobile, power efficient, for bring it in the Tegra line up.
Its even a complete contradiction with pascal return to FP64.. ( Now that Neural network research becomre way more important )
Denial
Senior Member
Posts: 12567
Joined: 2004-05-16
Senior Member
Posts: 12567
Joined: 2004-05-16
#5233304 Posted on: 02/17/2016 08:16 PM
I dont know if your report is wrong, but is really a majority of the HPC market, so a majority of HPC servers are really dedicated on Neural Network analysis ?
This a new growing market, but im not really sure that a majority of HPC servers are now dedicated to research about it. Big data, scientist market, medical research.. i mean, HPC servers are used in nearly one thousand of different " research " type.
When discussing about it, on BY3d as example, if important in term of research, it was still a niche market in term of servers and center dedicated to it.
I mean it's the majority of the growing market. All kinds of companies, large and small, are getting into it, in all different types of disciplines. And I think it's pretty obvious that Nvidia has shifted it's attention. They wouldn't waste resources on redesigning their architecture specifically for Neural Networks if it was only 1/1000s of the market.
It's obviously larger then that and has demand in the industry.
Edit: You Keep editing your post. Neural Networks have been around for way longer then 2012.. There was CUDA accelerated Neural Network applications like Aquila since 2010. In fact I'm pretty sure Nvidia had people from Matlab talk about their Neural Network implementation when the Kepler variants of Tesla launched.
I dont know if your report is wrong, but is really a majority of the HPC market, so a majority of HPC servers are really dedicated on Neural Network analysis ?
This a new growing market, but im not really sure that a majority of HPC servers are now dedicated to research about it. Big data, scientist market, medical research.. i mean, HPC servers are used in nearly one thousand of different " research " type.
When discussing about it, on BY3d as example, if important in term of research, it was still a niche market in term of servers and center dedicated to it.
I mean it's the majority of the growing market. All kinds of companies, large and small, are getting into it, in all different types of disciplines. And I think it's pretty obvious that Nvidia has shifted it's attention. They wouldn't waste resources on redesigning their architecture specifically for Neural Networks if it was only 1/1000s of the market.
It's obviously larger then that and has demand in the industry.
Edit: You Keep editing your post. Neural Networks have been around for way longer then 2012.. There was CUDA accelerated Neural Network applications like Aquila since 2010. In fact I'm pretty sure Nvidia had people from Matlab talk about their Neural Network implementation when the Kepler variants of Tesla launched.
Lane
Senior Member
Posts: 6361
Joined: 2005-02-25
Senior Member
Posts: 6361
Joined: 2005-02-25
#5233309 Posted on: 02/17/2016 08:27 PM
I mean it's the majority of the growing market. All kinds of companies, large and small, are getting into it, in all different types of disciplines. And I think it's pretty obvious that Nvidia has shifted it's attention. They wouldn't waste resources on redesigning their architecture specifically for Neural Networks if it was only 1/1000s of the market.
It's obviously larger then that and has demand in the industry.
Completely right with this.. Thats a big growing market, but the industry dont shift to it.. They addition this to their research panel.
Thats the big 2016 thing, in term of "attention" around the appliccations of the research... in term of units sales.. thats a bit different.
I mean it's the majority of the growing market. All kinds of companies, large and small, are getting into it, in all different types of disciplines. And I think it's pretty obvious that Nvidia has shifted it's attention. They wouldn't waste resources on redesigning their architecture specifically for Neural Networks if it was only 1/1000s of the market.
It's obviously larger then that and has demand in the industry.
Completely right with this.. Thats a big growing market, but the industry dont shift to it.. They addition this to their research panel.
Thats the big 2016 thing, in term of "attention" around the appliccations of the research... in term of units sales.. thats a bit different.
Lane
Senior Member
Posts: 6361
Joined: 2005-02-25
Senior Member
Posts: 6361
Joined: 2005-02-25
#5233310 Posted on: 02/17/2016 08:29 PM
I mean it's the majority of the growing market. All kinds of companies, large and small, are getting into it, in all different types of disciplines. And I think it's pretty obvious that Nvidia has shifted it's attention. They wouldn't waste resources on redesigning their architecture specifically for Neural Networks if it was only 1/1000s of the market.
It's obviously larger then that and has demand in the industry.
Edit: You Keep editing your post. Neural Networks have been around for way longer then 2012.. There was CUDA accelerated Neural Network applications like Aquila since 2010. In fact I'm pretty sure Nvidia had people from Matlab talk about their Neural Network implementation when the Kepler variants of Tesla launched.
I have add a part yes ... because, as Anandtech, i really doubt that Maxwell have been designed only or specially with Neural network in mind, but more as mobile design in mind. ( the lack of FP64 was not because neural network, but it was not affect neural network. )
I mean it's the majority of the growing market. All kinds of companies, large and small, are getting into it, in all different types of disciplines. And I think it's pretty obvious that Nvidia has shifted it's attention. They wouldn't waste resources on redesigning their architecture specifically for Neural Networks if it was only 1/1000s of the market.
It's obviously larger then that and has demand in the industry.
Edit: You Keep editing your post. Neural Networks have been around for way longer then 2012.. There was CUDA accelerated Neural Network applications like Aquila since 2010. In fact I'm pretty sure Nvidia had people from Matlab talk about their Neural Network implementation when the Kepler variants of Tesla launched.
I have add a part yes ... because, as Anandtech, i really doubt that Maxwell have been designed only or specially with Neural network in mind, but more as mobile design in mind. ( the lack of FP64 was not because neural network, but it was not affect neural network. )
Click here to post a comment for this news story on the message forum.
Senior Member
Posts: 11262
Joined: 2014-07-21
Interesting that Maxwell's not mentioned in that slide but they show Fermi and Kepler... Wonder how this would rank, who knows, maybe very close to Pascal.