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

#151
medi01
cucker tarlson
rdna1 refresh
Oh, dear god. I love we are not discussing "missing features", cough.

For the "why wouldn't they do it NOW", hell, why would they do it NOW?
GPUs released so far are on 7nm DUV process, which is:
a) incompatible
b) quite inferior

to 7nm EUV. Making major R&D on DUV made no sense and EUV wasn't there yet.

Now, would they roll out 350-400mm-ish chip in 2020? Maybe, but certainly not in Jan.
Posted on Reply
#152
cucker tarlson
medi01
Oh, dear god. I love we are not discussing "missing features", cough.

For the "why wouldn't they do it NOW", hell, why would they do it NOW?
GPUs released so far are on 7nm DUV process, which is:
a) incompatible
b) quite inferior

to 7nm EUV. Making major R&D on DUV made no sense and EUV wasn't there yet.

Now, would they roll out 350-400mm-ish chip in 2020? Maybe, but certainly not in Jan.
incompatible with what ?
Posted on Reply
#153
medi01
cucker tarlson
incompatible with what ?
7nm DUV.
Posted on Reply
#154
ratirt
medi01
7nm DUV.
Out of curiosity. Why would it be incompatible? I'm kinda stomped because I think that isn't right.
Posted on Reply
#155
cucker tarlson
ratirt
Out of curiosity. Why would it be incompatible? I'm kinda stomped because I think that isn't right.
leave it be.
it's like pigeon chess.
Posted on Reply
#156
ratirt
cucker tarlson
leave it be.
it's like pigeon chess.
No it isn't. if i don't know something i wish to know cause i might have missed it.
Posted on Reply
#157
efikkan
I don't get the point of this bickering.
Of course 7nm DUV => 7nm EUV requires some adjustments and a new tapeout, but the conversion shouldn't be hard.
And remember whatever improvements 7nm EUV will give, Nvidia get those too.
Posted on Reply
#158
cucker tarlson
efikkan
I don't get the point of this bickering.
Of course 7nm DUV => 7nm EUV requires some adjustments and a new tapeout, but the conversion shouldn't be hard.
And remember whatever improvements 7nm EUV will give, Nvidia get those too.
I think they're on samsung 7nm euv low power node
Posted on Reply
#159
efikkan
cucker tarlson
I think they're on samsung 7nm euv low power node
I believe some Nvidia official said the larger volume will be on TSMC. And I've not seen evidence yet of a Samsung node capable of ~250W chips, a low power node is not capable of that.
I would expect bigger dies on TSMC and smaller ones on Samsung.
Posted on Reply
#160
cucker tarlson
efikkan
I believe some Nvidia official said the larger volume will be on TSMC. And I've not seen evidence yet of a Samsung node capable of ~250W chips, a low power node is not capable of that.
I would expect bigger dies on TSMC and smaller ones on Samsung.
ga106 and even 104/103 could well be sub 200w cards.
Posted on Reply
#161
efikkan
I'll let Nvidia's engineers figure out where the sweet-spot is. But usually "low power" nodes are optimized for chips under 50W.
Posted on Reply
#162
cucker tarlson
efikkan
I'll let Nvidia's engineers figure out where the sweet-spot is. But usually "low power" nodes are optimized for chips under 50W.
maybe pci-e powered cards only
or mx gpus
Posted on Reply
#163
medi01
ratirt
Out of curiosity. Why would it be incompatible? I'm kinda stomped because I think that isn't right.
My knowledge in the area is too limited to answer this question, but that's what I got from Anand's article. Perhaps it has something to do with relative sizing of elements. I also couldn't find number of layers supported by N7 (non plus):
https://www.anandtech.com/show/14228/tsmc-reveals-6-nm-process-technology-7-nm-with-higher-transistor-density

N7 (7nm DUV) is not compatible with N7+ (7nm EUV).
However, it is compatible with N6 (6nm EUV), but its mass production, according to TSMC, is expected in 2021.
Posted on Reply
#164
cucker tarlson
medi01
My knowledge in the area is too limited to answer this question, but that's what I got from Anand's article. Perhaps it has something to do with relative sizing of elements. I also couldn't find number of layers supported by N7 (non plus):
https://www.anandtech.com/show/14228/tsmc-reveals-6-nm-process-technology-7-nm-with-higher-transistor-density

N7 (7nm DUV) is not compatible with N7+ (7nm EUV).
However, it is compatible with N6 (6nm EUV), but its mass production, according to TSMC, is expected in 2021.
but you're talking node design,not gpu design
Posted on Reply
#165
medi01
cucker tarlson
but you're talking node design,not gpu design
You said something... elusive....
Posted on Reply
#166
cucker tarlson
medi01
You said something... elusive....
I meant they can make a 350mm navi on rdna1 and 7nm duv now if they wanted to,but that would really be pushing it on power draw side.
Posted on Reply
#167
medi01
cucker tarlson
I meant they can make a 350mm navi on rdna1 and 7nm duv now if they wanted to,but that would really be pushing it on power draw side.
7nm DUV is inferior to 7nm EUV (20% worse power consumption AND worse density at the same time).
So it would be a very short lived product, that would make sense ONLY if it was really really chip to make and could be pulled off quickly.
Two very big IFs.
Also questionable on the PR side of things.
And when product would be out, people would anyway wait for Ampere first.
Posted on Reply
#168
cucker tarlson
medi01
7nm DUV is inferior to 7nm EUV (20% worse power consumption AND worse density at the same time).
So it would be a very short lived product, that would make sense ONLY if it was really really chip to make and could be pulled off quickly.
Two very big IFs.
Also questionable on the PR side of things.
And when product would be out, people would anyway wait for Ampere first.
true
though it's rather sad amd are waiting for 7nm euv to compete with tu104

looks like both companies are running independent schedeules,competing mostly with their own products.
Posted on Reply
#169
ratirt
cucker tarlson
true
though it's rather sad amd are waiting for 7nm euv to compete with tu104

looks like both companies are running independent schedeules,competing mostly with their own products.
it's not sad. Maybe disappointing a little depends how you look at it. What would you do if this was your only way to compete? It is better to wait and get something good enough instead making a mistake that may trash the reputation to bits. I think AMD is doing the right thing. Besides, whats the point of releasing something and spend money on it when you know you can have something better in few/several months. Especially if NV is releasing new card which will be faster. Better to focus on this one than dwell in the past gens just to release something. Skip, forget, move on.
Posted on Reply
#170
cucker tarlson
ratirt
it's not sad. Maybe disappointing a little depends how you look at it. What would you do if this was your only way to compete? It is better to wait and get something good enough instead making a mistake that may trash the reputation to bits. I think AMD is doing the right thing. Besides, whats the point of releasing something and spend money on it when you know you can have something better in few/several months. Especially if NV is releasing new card which will be faster. Better to focus on this one than dwell in the past gens just to release something. Skip, forget, move on.
Sad from consumers perspective.
Posted on Reply
#171
John Naylor
I never would have imagined the tech world following the political journalists ... while CUDA can serve you will in certain apps ... I'm not interested in what a new card "cuda" done, could do, might do, only what it was proved to be capable of ... we don't need imagined facts to go with alternative facts. It's just nerd porn. We have about as much of a chance seeing what is projected in these articles "in real life" as we do a playboy centerfold in real life.

I really don't care how many cores, what size in nm or anythung else. If a CPU is 7nm / 16 cores and it can't go as fast as 14nm 8 core CPU in the applications I use, i don't want it. With a GFX card, don't care about any of the specs other than how fast it runs apps, how hot, how noisy and how much power. The rest is irrelevant. Video editing benefits from more cores .... so if a 12 core AMD CPU is 10% slower than an 8 core Intel CPU, why would i want the one w/ more cores when running Photoshop ?

PCs are a tool ... not a statement of who we are. Buying a pair of Air Jordans, isn't going to make you a NBA player. If you are choosing a PC component by any other criteria than how well it performs, and then only the apps you actually will use, you're doing it wrong.
Posted on Reply
#172
Bronan
Anymal
Turing will be discounted though. 2070super for 350eur, njami!
Still too much for my taste, these ampere seem to be pro gpu only for now.
However we will see what they have cooked up, i would not be surprised that its gonna cost 2 arms and a leg to get them.
So no sale to me for sure, i no longer want to spend a small fortune into any gpu ever again.
Because the chance is high that within 6 months they release a faster one pretty quick
Especially if amd is able to compete, but time will tell what happens if intel does start making similar products
Posted on Reply
Add your own comment