ORNL to acquire new supercomputer in 2017

[DrMrLordX]

https://www.olcf.ornl.gov/2014/11/14/oak-ridge-to-acquire-next-generation-supercomputer/

So ORNL is switching to IBM's POWER uarch for the per-node host processor (assuming it follows that pattern). Titan is Opteron + Tesla per node.

Volta is documented and on Nvidia roadmaps, so that comes as no surprise, but I have heard very little about POWER9. It's been mentioned, just . . . not in great detail, at least not that I can tell.

 

[Arachnotronic]

I'm pretty excited to see what IBM did with POWER9. POWER8 is a beast.

 

[witeken]

I'm pretty excited to see what IBM did with POWER9. POWER8 is a beast.

I didn't even notice it uses Power9, interesting.

 

[NTMBK]

I didn't even notice it uses Power9, interesting.

POWER9 is, as far as I'm aware, the only thing which will support NVLink. (Especially since Denver for servers seems to be dead.)

 

[jdubs03]

This just popped up:

http://www.anandtech.com/show/8729/nvidia-launches-tesla-k80-gk210-gpu

It's interesting that Nvidia has remained on 28nm with Kepler yet still has been able to get some extra performance out of those transistors. It makes me excited to see what a 20nm Maxwell could do, let alone 16nm Volta. There is plenty to be optimistic about on a performance basis.

I know extrapolating performance is quite silly at this juncture but still a 20nm Maxwell M80 or whatever could be near 10 TFLOPS DP, at least in terms of the highest-end Tesla GPU. Volta I think will put the HPC space into exascale around 2020, because it seems like Nvidia is keeping their uArch's on a 2-3 year time-scale.

 

[III-V]

Over 136 petaflops... Nice. Hitting an exaflop by 2020 seems very achieveable.

 

[NTMBK]

It's interesting that Nvidia has remained on 28nm with Kepler yet still has been able to get some extra performance out of those transistors.

Actually, they've added more transistors... double register file and double L1 cache means that this is going to be even bigger than GK110.

 

[Ajay]

I know extrapolating performance is quite silly at this juncture but still a 20nm Maxwell M80 or whatever could be near 10 TFLOPS DP, at least in terms of the highest-end Tesla GPU. Volta I think will put the HPC space into exascale around 2020, because it seems like Nvidia is keeping their uArch's on a 2-3 year time-scale.

Volta looks slated for 2017 from the link in the OP. So Maxwell or Pascal or both are being dropped - unless NV is bifurcating it's consumer and HPC lines.

 

[III-V]

Well they just did that today, Ajay, lol.

 

[witeken]

Over 136 petaflops... Nice. Hitting an exaflop by 2020 seems very achieveable.

Back in 2013 we'd reach exascale computing in 2018, so it has moved 2 years further. Disappointing.

 

[jdubs03]

Back in 2013 we'd reach exascale computing in 2018, so it has moved 2 years further. Disappointing.

There is always variance in these assumptions, even conceding that point, the last time I was looking at this topic, a lot of predictions were focusing on 2020 as the point in time in which we would have the technological capability.

 

[III-V]

Back in 2013 we'd reach exascale computing in 2018, so it has moved 2 years further. Disappointing.

I was just saying by the end of the decade, not trying to pinpoint a particular year that we'd hit it.

 

[witeken]

There is always variance in these assumptions, even conceding that point, the last time I was looking at this topic, a lot of predictions were focusing on 2020 as the point in time in which we would have the technological capability.

If you do some extrapolation, you come out at 2018.

It should be possible, but with the slow down of Moore's Law, it might be challenging, maybe 10nm Xeon Phi could be used.

 

[NTMBK]

It should be possible, but with the slow down of Moore's Law, it might be challenging, maybe 10nm Xeon Phi could be used.

BTW, Intel has officially announced that 10nm Phi is codenamed Knight's Hill: http://www.v3.co.uk/v3-uk/news/2381...dated-interconnect-fabric-to-xeon-phi-roadmap

 

[witeken]

BTW, Intel has officially announced that 10nm Phi is codenamed Knight's Hill: http://www.v3.co.uk/v3-uk/news/2381...dated-interconnect-fabric-to-xeon-phi-roadmap

Nice, very exciting, the first 10nm procuct announced! Will probably launch in 2017, though. Intel here really blew TSMC away, whose equivalent node is still 7 years or so away...

 

[jdubs03]

BTW, Intel has officially announced that 10nm Phi is codenamed Knight's Hill: http://www.v3.co.uk/v3-uk/news/2381...dated-interconnect-fabric-to-xeon-phi-roadmap

Of course the slide is cut off Gotta leave enough up to the imagination for a while.

This article states 2020-2022 for exascale, I'd like to think it'll be sooner rather than later.

http://www.itworld.com/article/2848755/intel-says-its-supercomputing-chips-wont-suffer-larrabees-fate.html

 

[witeken]

Don't want to open a new thread, but it seems not much has changed in the Top500: still the same fastest supercomputer for the 4th time (30PFLOPS).

 

[III-V]

Of course the slide is cut off Gotta leave enough up to the imagination for a while.

This article states 2020-2022 for exascale, I'd like to think it'll be sooner rather than later.

http://www.itworld.com/article/2848...mputing-chips-wont-suffer-larrabees-fate.html

Google, bro:

 

[Nothingness]

Nice, very exciting, the first 10nm procuct announced! Will probably launch in 2017, though. Intel here really blew TSMC away, whose equivalent node is still 7 years or so away...

There is a mismatch in your use of future and past :biggrin:

 

[Nothingness]

Don't want to open a new thread, but it seems not much has changed in the Top500: still the same fastest supercomputer for the 4th time (30PFLOPS).

Yes, the top 9 didn't change, disappointing

 

[witeken]

Google, bro:

I hope it isn't a coincidence that I haven't seen anything about this topic on AnandTech or ExtremeTech, contrary to the Nvidia coverage.

 

[Nothingness]

I hope it isn't a coincidence that I haven't seen anything about this topic on AnandTech or ExtremeTech, contrary to the Nvidia coverage.

It seems like there was basically no new information during the presentation.

 

[witeken]

It seems like there was basically no new information during the presentation.

Oh, then Intel PR did some lying when they said they revealed new things.


http://newsroom.intel.com/community...as-momentum-builds-for-intel-xeon-phi-product

 

[mavere]

It seems like there was basically no new information during the presentation.

As far as I could tell, we now know the name "Knights Hill" and some vague marketing comparisons regarding the new interconnect's port density and port-to-port latency.

 

[Nothingness]

As far as I could tell, we now know the name "Knights Hill" and some vague marketing comparisons regarding the new interconnect's port density and port-to-port latency.

Yes, but that certainly isn't as interesting as details about KNL, and IMHO explains why the press coverage wasn't the same as for NVIDIA new GPGPU chip.