Monday, February 3rd 2020

NVIDIA's Next-Generation "Ampere" GPUs Could Have 18 TeraFLOPs of Compute Performance

NVIDIA will soon launch its next-generation lineup of graphics cards based on a new and improved "Ampere" architecture. With the first Tesla server cards that are a part of the Ampere lineup going inside Indiana University Big Red 200 supercomputer, we now have some potential specifications and information about its compute performance. Thanks to the Twitter user dylan552p(@dylan522p), who did some math about the potential compute performance of the Ampere GPUs based on NextPlatform's report, we discovered that Ampere is potentially going to feature up to 18 TeraFLOPs of FP64 compute performance.

With Big Red 200 supercomputer being based on Cray's Shasta supercomputer building block, it is being deployed in two phases. The first phase is the deployment of 672 dual-socket nodes powered by AMD's EPYC 7742 "Rome" processors. These CPUs provide 3.15 PetaFLOPs of combined FP64 performance. With a total of 8 PetaFLOPs planned to be achieved by the Big Red 200, that leaves just a bit under 5 PetaFLOPs to be had using GPU+CPU enabled system. Considering the configuration of a node that contains one next-generation AMD "Milan" 64 core CPU, and four of NVIDIA's "Ampere" GPUs alongside it. If we take for a fact that Milan boosts FP64 performance by 25% compared to Rome, then the math shows that the 256 GPUs that will be delivered in the second phase of Big Red 200 deployment will feature up to 18 TeraFLOPs of FP64 compute performance. Even if "Milan" doubles the FP64 compute power of "Rome", there will be around 17.6 TeraFLOPs of FP64 performance for the GPU.
Sources: @dylan522p(Twitter), The Next Platform
Add your own comment

172 Comments on NVIDIA's Next-Generation "Ampere" GPUs Could Have 18 TeraFLOPs of Compute Performance

#101
renz496
gamefoo21
Nvidia definitely doesn't help the situation. They purposefully cripple OpenCL on their GPUs.

They refuse to support newer versions of it. Radeons have supported OpenCL 2.0 since the 7790.

It's like nV making an incompatible Variable Refresh tech, to the open standard in DisplayPort. Intel adopted the open spec, the HDMI group came on board, and adopted it. Suddenly nV made gSync work with Adaptive Vsync no NV port corruption necessary.

NV will block and slow down OpenCL as much as they can. I suspect you'll see their tune change if the Intel and AMD really start challenging NV.

I still hate that NV does everything they can to make PhysX all but impossible if your main GPU isn't NV.



Titan V?
Because nvidia don't relly have a reason to push opencl when their CUDA development will always outpace Opencl development due to their full control of CUDA. no need to purposely holding their performance back because it might be unfair to AMD. take this as an example: nvidia push for a feature that can speed up performance on their hardware only. But because it will only give nvidia the advantage the feature are being debate among IHV and being hold back or even refused to be part of API spec. This is part of the problem with open source API. That's why nvidia end up breaking some of opengl standard before for performance reason. If nvidia going all out with opencl and did similar thing what they did with opengl before people will still going to complain about it. So rather than wasting time to porting all their work to opencl they just continue their already solid CUDA foundation to move forward.
Posted on Reply
#102
efikkan
cucker tarlson
They're on 7 nm tsmc and new uarch,why dont they put up? Cause of what I wrote before.They would lose to nvidias +$500 cards cause of lack of features,so they're happy to milk mid range.
Let us see those chips,I am not buying a word their CEO or their fanbase say about high end Radeon.
I'm not sure if I'm reading you correctly, but I wouldn't call what they're doing "milking the mid-range", they are moving small volumes compared to Nvidia in the mid-range. I think they would love to have products in the high-end, those would offer good margins and good PR for the company. But their "mistake" is that their focus is on making custom chips for the consoles, and then port this to PCs to make some extra bucks. In order to become competitive, they need to put desktop PCs first, and I don't think they are willing to do that.

Navi 1x couldn't scale much further, and even if they had a 60 CU version, it could easily go over 350W if they didn't clock it much lower than RX 5700 XT, so Navi 1x couldn't even theoretically truely compete in the high-end.

The transition from 7nm DUV to 7nm EUV alone is not going to be enough to stay competitive with high-end Turing, there have to be at least some decent improvements to the architecture. But for a theoretical 80 CU card, or something to be competitive with Nvidia's next gen, AMD would have to pull off some major improvements to their architecture.
Posted on Reply
#103
cucker tarlson
efikkan
I'm not sure if I'm reading you correctly, but I wouldn't call what they're doing "milking the mid-range", they are moving small volumes compared to Nvidia in the mid-range.
but still very decent by their standards and with very good margins
efikkan
But their "mistake" is that their focus is on making custom chips for the consoles, and then port this to PCs to make some extra bucks. In order to become competitive, they need to put desktop PCs first, and I don't think they are willing to do that.
that's my main problem with them too.
nvidia may charge a lot,but they're clearly pc oriented,and most (all) new technologies that made it into pc gaming,like variable refresh rate,ulmb,vxao,htfs,gpu based physx or recently rtx,dlss and vrs were pushed by nvidia not amd.

I think we're at a time when two upgrade paths are intertwining.
In the recent years gpu manufacturers have been pushing higher clocks,core counts and vram capacity with little to no improvements in the quality of the image we're getting except for higher resolution.Some peple like me are already caught up with performance a long time ago and want to see some changes to the quality too,while an even bigger group of people are mainly after gpu performance first since they're sitting on older cards.
Posted on Reply
#104
ratirt
efikkan
But their "mistake" is that their focus is on making custom chips for the consoles, and then port this to PCs to make some extra bucks. In order to become competitive, they need to put desktop PCs first, and I don't think they are willing to do that.
I wouldn't call it a mistake. Consoles have a lot of potential in terms of sales and cash returns. I guess they have decided to get consoles first because it is more lucrative and not much of a competition for AMD. I'm sure when the consoles are sorted PC will get something new as well.
Posted on Reply
#105
cucker tarlson
ratirt
I guess they have decided to get consoles first because it is more lucrative
than what ?
Posted on Reply
#106
medi01
cucker tarlson
This is a GPU thread,Intel and epyc have nothing to do with it.
It clearly demonstrates AMD's capabilities.

cucker tarlson
They're on 7 nm tsmc and new uarch,why dont they put up?
They are gradually rolling out superior products, starting with the sure-bet-most-profitable.

cucker tarlson
Cause of what I wrote before.They would lose to nvidias +$500 cards cause of lack of features,so they're happy to milk mid range.
That is one hell of a conjecture.
Are you from the same universe?
In our universe, a 2070 super, a 550mm2 13600M transistors card struggles with 10,300 million transistor 250mm2 5700XT.
Yeah, different process node, but wait, that number of transistors diff, even if you write off 8% of it to Ray Tracing lols, what are tehy?

There is no fact at hand, that even remotely hints at AMD not being able to upset The Leather Man with bigger chips. Oh, and they don't need the biggest chips on the planet either.
Posted on Reply
#107
cucker tarlson
medi01
It clearly demonstrates AMD's capabilities.
it clearly does not demonstrate nothing about their gpu division.
show me the clear link how a cpu architecture translates to a better gpu.
medi01
They are gradually rolling out superior products, starting with the sure-bet-most-profitable.
well one navi 10 die really,then sliced into three cards,and that's about it.

medi01
In our universe, a 2070 super, a 550mm2 13600M transistors card struggles with 10,300 million transistor 250mm2 5700XT.
one line below

medi01
Yeah, different process node
write off 8% of it to Ray Tracing
:roll: :roll:
medi01
There is no fact at hand, that even remotely hints at AMD not being able to upset The Leather Man with bigger chips.
there's nothing that hints at them being able either.
and more importantly willing to.
Posted on Reply
#108
medi01
cucker tarlson
it clearly does not demonstrate nothing about their gpu division.
show me the clear link how a cpu architecture translates to a better gpu.
Let's leave it at that.

cucker tarlson
:roll: :roll:
I'll take it as "I can't make a coherent argument, so let me blow it up". I hope you are ok.
With 8% figure I was being generous:



cucker tarlson
there's nothing that hints at them being able either.
They've rolled out 2 times smaller chip with 30% smaller number of transistors and the same ballpark performance. If that is not a hint enough, perhaps you need to do something about the green reality distortion field around you.
Posted on Reply
#109
cucker tarlson
medi01
Let's leave it at that.


I'll take it as "I can't make a coherent argument, so let me blow it up". I hope you are ok.
With 8% figure I was being generous:




They've rolled out 2 times smaller chip with 30% smaller number of transistors and the same ballpark performance. If that is not a hint enough, perhaps you need to do something about the green reality distortion field around you.
tell me,in what "universe" do you compare die sizes of chips made on different nodes and with different hardware features but you don't compare efficiency.
tu106 is still 1.12x more efficient on an older node,tu104 1.25x more
and if rt cores alone are 6%,then adding tensor would add up to more than 8%



medi01
I'll take it as "I can't make a coherent argument, so let me blow it up". I hope you are ok.
no,you can't.
you contradict yourself all the time

medi01
Let's leave it at that.
of course,there's nothing to be said about epyc vs intel here.
"with zen they took over intel,so they'll do the same with navi vs ampere" is a silly argument

medi01
They've rolled out 2 times smaller chip with 30% smaller number of transistors and the same ballpark performance. If that is not a hint enough, perhaps you need to do something about the green reality distortion field around you.
but they can't match tu104 and tu102 with 25% worse efficiency like they did with navi 10 vs tu106,pretty easy to understand.
they'd probably need a card with a higher power conspumption than 2080ti to match the 2080.aib 5700xt's are already more power hubngry than both 2080 and 2080 super



here's an aib 2070 for comparison



efikkan
The transition from 7nm DUV to 7nm EUV alone is not going to be enough to stay competitive with high-end Turing, there have to be at least some decent improvements to the architecture.
they'll be able to make a 2080 super without going above 280W for the aib cards,but that's about it.
Posted on Reply
#110
ratirt
cucker tarlson
than what ?
You mean "then what"?
Then, they will hold hands in a circle around the fire, dance, sing songs, drink milk and eat cookies like you in your fantasy world. :)

medi01 gave you a good example and procured arguments by the conclusion he's got. Funny though I got the same.
You instead jump from one thing to the other disregarding what the conversation is about. He mentioned die size and difference of the dies sizes between NV and AMD GPUs.
He didn't talk about efficiency that you have jumped to disregarding everything he said about size, transistor and performance wise. saying that efficiency is the variable determining possibility of bigger die is nuts.
The 7nmEUV will be more efficient and less power hungry. Depends how AMD is going to play this card though. 250mm2 and same performance with twice the size chip from NV's is impressive. If you think AMD can't go bigger you are a lost cause. Since AMD didn't release RDNA big navi, it doesn't mean it is not coming or that AMD couldn't do it. I already mentioned why but of course you disregarded what I said and changed the subject (classic). RDNA2 will be the BIG NAVI now. AMD's more than capable of doing this. How will this work with the 7nm EUV time will tell since we don't have any equivalent die size based on RDNA to have an idea. We can have only guesses.
BTW, the Intel mention, was to show you that nothing is certain and AMD can disrupt market with their products. If they did it in the CPU market (even though a lot said it is stupid and it won't happen) they may do it in GPU (even though you say it is stupid and won't happen). Ryzen and RDNA iterations are quite similar like a pattern but Ryzen started earlier. Priority I'd say. Just like the console market is a priority over PC at this moment.
So buckle up sit tight and wait because you are going for a crazy ride my friend. Sooner or later.
Posted on Reply
#111
cucker tarlson
ratirt
You mean "then what"?
Then, they will hold hands in a circle around the fire, dance, sing songs, drink milk and eat cookies like you in your fantasy world. :)

medi01 gave you a good example and procured arguments by the conclusion he's got. Funny though I got the same.
You instead jump from one thing to the other disregarding what the conversation is about. He mentioned die size and difference of the dies sizes between NV and AMD GPUs.
He didn't talk about efficiency that you have jumped to disregarding everything he said about size, transistor and performance wise. saying that efficiency is the variable determining possibility of bigger die.
The 7nmEUV will be more efficient and less power hungry. Depends how AMD is going to play this card though. 250mm2 and same performance with twice the size chip from NV's is impressive. If you think AMD can't go bigger you are a lost cause. Since AMD didn't release RDNA big navi, it doesn't mean it is not coming or that AMD couldn't do it. I already mentioned why but of course you disregarded what I said and changed the subject (classic). RDNA2 will be the BIG NAVI now. AMD's more than capable of doing this. How will this work with the 7nm EUV time will tell since we don't have any equivalent die size based on RDNA to have an idea. We can have only guesses.
BTW, the Intel mention, was to show you that nothing is certain and AMD can disrupt market with their products. If they did it in the CPU market (even though a lot said it is stupid and it won't happen) they may do it in GPU (even though you say it is stupid and won't happen). Ryzen and RDNA iterations are quite similar like a pattern but Ryzen started earlier. Priority I'd say. Just like the console market is a priority over PC at this moment.
So buckle up sit tight and wait because you are going for a crazy ride my friend. Sooner or later.
no,I mean than what.
more lucrative than what.
you're making a comparison with no reference.

medi01 didn't talk about efficiency cause that's the elephant in the room he knew he couldn't touch,and got his chops busted again.amd can't compete with big turing even on 7nm when their 250mm cards break 250w,any other point is irrelevant including fanstasizing about die sizes.

tl,dr for the rest of your post.please write less but more fact less rumors and opinions.

the Intel mention was purely cause you have nothing else to cling to so you're quoting the example on how an incompetent company can lose performance crown in sereval years time.

ratirt
Ryzen and RDNA iterations are quite similar like a pattern
how ?
Posted on Reply
#112
ratirt
cucker tarlson
how ?
Brand new designs for starters. Starting from scratch designs.
cucker tarlson
no,I mean than what.
more lucrative than what.
I see. Then you didn't read what I wrote. Than the PCs at this point.
cucker tarlson
medi01 didn't talk about efficiency cause that's the elephant in the room he knew he couldn't touch,and got his chops busted again.amd can't compete with big turing even on 7nm when their 250mm cards break 250w,any other point is irrelevant including fanstasizing about die sizes.
He didn't talk about efficiency, cause that is not defining the size of the chip that can be manufactured.
cucker tarlson
tl,dr for the rest of your post.please write less but more fact less rumors and opinions.
There are no facts cause these cards are not out yet. You need to understand the difference between rumors and having your conclusions.
7nm EUV is a fact and will be more efficient than 7nm process. The performance is oscillating around 20% better. Considering the previous node advancements, the power efficiency reaches higher percentages than performance. Why I can't use this to say that RDNA2 will be more efficient and AMD will make RDNA2 Big Navi?
Plus we don't know what the RDNA2 architecture will bring in terms of arch improvements because it is not out yet we can only base out conclusions on the RDNA Navi.
Posted on Reply
#113
medi01
Intel's problems are not only architecture (although AMD's chiplet approach kicked their arse in a major way), it is also Fab advantage that came out of nowhere and is only growing, thank you TSMC.

AMD's weapons against NV are more limited, coming down to a 45% margin vs 60%, with plenty of room to harm The Leather Man.

Unless they pull "Chiplet Design 2, GPUs or Tears of the Leather Man", well, I don't see it happening.

cucker tarlson
You compare die sizes of chips made on different nodes
Yep, that is why I've included number of transistors.

cucker tarlson
and if rt cores alone are 6%,then adding tensor would add up to more than 8%
Does 2060 have em? Why? Even if you max that figure at 14%, it still doesn't cover 10 million vs 13 million transistor gap.

cucker tarlson
25% worse efficiency
Eh, whah?
Ah, figure rigging, hehe.

cucker tarlson
5700xt's are already more power hubngry than both 2080 and 2080 super
Nah. Stock thing is between 2080 and 2080s, and the former is what, 8% faster in TPU charts?
Anyhow, card pushed way up on clock-rates isn't indicative of actual perf/watt.

Even last minute OCed 5600XT is quite a bit "cooler" than a number of NV offerings:



How does 2080Ti handle "the heat"? Well, uh, doh, much larger die that runs at LOWER clock. Yay, that's cool ain't it?

The thought that AMD "cannot produce large chips" is... weird.
I still wonder what are the missing "features" on AMD chips. Is it that wonderful G-Sync think, that adds "$200+ more expensive" feature to monitors? :D :D :D Or is it that missing in action RT thing, found in under 1% of games, almost exclusively only if NV paid for it, that is anyhow switched off by most gamers due to performance hit? Let us know please.
Posted on Reply
#114
ratirt
medi01
How does 2080Ti handle "the heat"? Well, uh, doh, much larger die that runs at LOWER clock. Yay, that's cool ain't it?
If you consider the OC'ed versions of the 2080TI, you get closer to 400Watt consumption peak.
Posted on Reply
#115
cucker tarlson
Waste of time.Efficiency charts at 4k are irrelevant to you,so you're bringing up 5600 xt efficiency at 1080p in a discussion about big dies.
Cards pushed to max clock rates (which is your definition of aib cards for some reason ) are irrelevant either.
Have fun boys waiting for that big navi then if you believe its gonna be that fantastic.
Posted on Reply
#116
efikkan
cucker tarlson
In the recent years gpu manufacturers have been pushing higher clocks,core counts and vram capacity with little to no improvements in the quality of the image we're getting except for higher resolution.Some peple like me are already caught up with performance a long time ago and want to see some changes to the quality too,while an even bigger group of people are mainly after gpu performance first since they're sitting on older cards.
Higher resolution and higher refresh rates is still something, something that matters to many gamers. Fluid gaming is at least important to me.

But many games seems to demand more performance vs. what higher fidelity they offer in return. Some of this may be due to ever-increasing advanced lighting techniques, but most of it is due to mass produced "low quality" games.

ratirt
I wouldn't call it a mistake. Consoles have a lot of potential in terms of sales and cash returns. I guess they have decided to get consoles first because it is more lucrative and not much of a competition for AMD. I'm sure when the consoles are sorted PC will get something new as well.
I did use quotes, didn't I? ;)
They chose custom chips for consoles as the priority because it's safe money, not because it's good money. But this strategy might face some competition from Intel in the long run.

My point is that in order for AMD to become competitive and catch up with Nvidia, they need to put PC first. What they could have done instead of their custom-first route is to make a good CPU (like they finally have) and a good GPU, and then just bundle it instead of customizing it heavily. But it's too late now…
Posted on Reply
#117
ratirt
efikkan
My point is that in order for AMD to become competitive and catch up with Nvidia, they need to put PC first.
Everything has its timing. Guess PC time isn't up yet :) Gotta wait a bit more for that. I don't think it is too late. The world isn't over yet :)
Posted on Reply
#118
efikkan
ratirt
Everything has its timing. Guess PC time isn't up yet :) Gotta wait a bit more for that. I don't think it is too late. The world isn't over yet :)
We're not running out of time, but Nvidia is marching ahead, so catching up is only going to get harder.
Posted on Reply
#119
medi01
efikkan
My point is that in order for AMD to become competitive and catch up with Nvidia,
Because 5600 or 5700 series are not competitive! Oh wait...


efikkan
they need to put PC first. What they could have done instead of their custom-first route is to make a good CPU (like they finally have) and a good GPU, and then just bundle it instead of customizing it heavily. But it's too late now…
Then need to show commitment to the market that has basically saved them.
Besides, there nothing that hints at that console development is somehow harming "other" development. If anything, it's more R&D money for AMD.

efikkan
We're not running out of time, but Nvidia is marching ahead, so catching up is only going to get harder.
I think I've heard that before., this narrative is sooo Vega times...
Vega => Navi jump was a major architectural step forward by AMD, they have shrunk chip by a third, switched to a cheaper mem while keeping the same performance.

There is a clear parity on transistor/perf side of things. NV might be doing power savings a bit better, that is pretty much the only gap that we could see when Ampere faces RDNA2.

But when Ampere comes, you'll either see competing AMD cards, or The Leather Man going full greetard again. Either way, AMD would be fine. As for "bought by 1% or below" market of excessively expensive cards, from my POV it's a purely theoretical "harm green ego" thing, not something I would care about.
Posted on Reply
#120
kapone32
Can we not all agree that we need a high end inexpensive card from AMD to bring prices back down from the stratosphere. As far as if they can or not no one knows what AMD is currently doing. They only thing we can say is that they released variants of Navi10 but their R&D GPU department has not given us any carrots on what they are currently working on. All we have to go on is Dr Su's words. It appears to me that the mitigating factor for GPUs seems to be clock and memory speed.
Posted on Reply
#121
medi01
kapone32
Can we not all agree that we need a high end inexpensive card from AMD to bring prices back down from the stratosphere.
You think AMD rolling out 2080Ti type card will somehow reduce pricing of mid/low end cards???
Posted on Reply
#122
ratirt
efikkan
We're not running out of time, but Nvidia is marching ahead, so catching up is only going to get harder.
Yes it is marching ahead. AMD has plan so lets wait and see what AMD will do. Releasing something just because others are ahead to try catching up, rushing, that would be a mistake. Intel was ahead too and look now. It took a bit of time but AMD managed. So, lets see what AMD will do in GPU department. It is not like NV is 10 gens away. I think AMD GPUs are more closer to NV than AMD's CPUs were with Intel's before ZEN.
medi01
You think AMD rolling out 2080Ti type card will somehow reduce pricing of mid/low end cards???
That depends. If AMD launches a card 2080 TI, with performance and efficient we may not see the drop. Unless AMD will be satisfied with the lower price they mark the card with. Seriously doubt that would happen. It may be lower but it wont bring 2080 TI's price to $600 or even $800. On the other hand if the performance is there but the efficiency lacks a bit. It may happen. That would also depend on how much manufacturing of the card will cost and that is the major factor for the price point here.
Posted on Reply
#123
cucker tarlson
ratirt
Yes it is marching ahead. AMD has plan so lets wait and see what AMD will do. Releasing something just because others are ahead to try catching up, rushing, that would be a mistake. Intel was ahead too and look now. It took a bit of time but AMD managed. So, lets see what AMD will do in GPU department. It is not like NV is 10 gens away. I think AMD GPUs are more closer to NV than AMD's CPUs were with Intel's before ZEN.

That depends. If AMD launches a card 2080 TI, with performance and efficient we may not see the drop. Unless AMD will be satisfied with the lower price they mark the card with. Seriously doubt that would happen. It may be lower but it wont bring 2080 TI's price to $600 or even $800. On the other hand if the performance is there but the efficiency lacks a bit. It may happen. That would also depend on how much manufacturing of the card will cost and that is the major factor for the price point here.
they have a plan to but they have to wait cause catching up now is a mistake ?
how about catching up last year ?
for amd to make faster card would be a mistake,although they totally absolutely could :rolleyes: things the red fanbase say and somehow Im not surprised that's what the gpu situation looks like now

medi01
You think AMD rolling out 2080Ti type card will somehow reduce pricing of mid/low end cards???
no cause they'd sell it at $900
Posted on Reply
#124
ratirt
cucker tarlson
they have a plan to but they have to wait cause catching up now is a mistake ?
how about catching up last year ?
for amd to make faster card would be a mistake,although they totally absolutely could :rolleyes: things the red fanbase say and somehow Im not surprised that's what the gpu situation looks like now
Listen. I understand you have problems with ego and attitude. You are so "smart" you don't know which way is up. If you don't get it believe me nobody will explain nor try to convince you. I don't care about last year. AMD doesn't either. Mistake would be to release a card that is not up for the task and that is why you didn't get it last year. Simple?
Posted on Reply
#125
cucker tarlson
ratirt
Listen. I understand you have problems with ego and attitude. You are so "smart" you don't know which way is up. If you don't get it believe me nobody will explain nor try to convince you. I don't care about last year. AMD doesn't either. Mistake would be to release a card that is not up for the task and that is why you didn't get it last year. Simple?
why should it be not up for the task though ?
Posted on Reply
Add your own comment