[Supercomputers] 1st design wins for Nvidia Volta and IBM PowerPC

Toggle sidebar Toggle sidebar
X

xpea

Senior member
Feb 14, 2014
458
156
116
Already the 2 first design wins for Nvidia Pascal :awe:
The numbers are big and so is the news. The U.S. Department of Energy today unveiled plans to build two GPU-powered supercomputers. Each will deliver at least 100 petaflops of compute performance.

And one – the Summit system at Oak Ridge National Laboratory, designed for open science – is expected to be 150 petaflops. That’s more than three times the peak speed of today’s fastest supercomputer.

...

NVIDIA GPUs and IBM POWER CPUs, connected with the NVLink interconnect technology, will power both machines.
Click to expand...

It means many things:
1/ Intel is out of these deals ! Partnership between NV and IBM pays off
2/ Pascal is -at one way or other- real, ie at prototype stage
3/ with all these PetaFLOPs, finally crysis at 60fps / 8k becomes reality :biggrin:

more information at the source: http://blogs.nvidia.com/blog/2014/11/14/what-is-nvlink/

the video is also interesting: https://www.youtube.com/watch?v=nqERLsNTnXk

edit: it's Volta, not Pascal. Title has been corrected
 
Last edited:
KaRLiToS

KaRLiToS

Golden Member
Jul 30, 2010
1,918
11
81
Honestly, I wonder what kind of voltage is feeding this.

25 kv? :awe:

(No, honestly, it must be 220v )
 
N

nvgpu

Senior member
Sep 12, 2014
629
202
81
http://info.nvidianews.com/rs/nvidia/images/An%20Inside%20Look%20at%20Summit%20and%20Sierra%20Supercomputers-3-1.pdf

jRSlxYz.png


It's a big design win for Nvidia Volta, the next-gen architecture after Pascal.
 
Last edited:
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,219
54
91
There is most likely a goal to vastly outperform the Titan supercomputer while using the same or less energy to do it. Summit, as it looks to be called and according to the video, is 5 to 10 times more powerful than its Titan predecessor.
The next goal is to shrink all this down into the size of a smartphone. We can call them tri-corders.
:)
 
NTMBK

NTMBK

Lifer
Nov 14, 2011
10,448
5,831
136
Clearly HPC values unified address space. Hopefully AMD will be able to produce some competitive APUs for this market.
 
X

xpea

Senior member
Feb 14, 2014
458
156
116
NTMBK said:
Clearly HPC values unified address space. Hopefully AMD will be able to produce some competitive APUs for this market.
Click to expand...
From the posted pdf, these new supercomputers will already have HSA (well Nvidia version of it through NVLink)
Regarding AMD, I don't see them doing much in this space. They don't have the resource to address this market and I don't even talk about their shrinking R&D budget...
 
Last edited:
E

Erenhardt

Diamond Member
Dec 1, 2012
3,251
105
101
IBM? That is a slap in the intels face.

Apparently time is not money here. They can afford to wait for nv to catch with the others.
 
X

xpea

Senior member
Feb 14, 2014
458
156
116
Erenhardt said:
IBM? That is a slap in the intels face.

Apparently time is not money here. They can afford to wait for nv to catch with the others.
Click to expand...
to catch with what ? Do you know something that these engineers at Oak Ridge and Livermore don't ?
 
X

xpea

Senior member
Feb 14, 2014
458
156
116
witeken said:
It surely would've been nice if it used Knight's Landing.
Click to expand...
Obviously, Intel Xeon + KL was one contender but it was discarded. The million dollar question is Why ?
From Nvidia own PDF :
• The Heterogeneous computing model
• NVIDIA NVLink high-speed interconnect
• NVIDIA GPU accelerator platform
• IBM OpenPOWER platform
Click to expand...
but are these valid reasons ?
 
9

96Firebird

Diamond Member
Nov 8, 2010
5,742
340
126
Hopefully this does not delay the consumer version of big-Volta when the time comes. I'm not sure if the release of big-Volta will coincide with a new node like big-Kepler did, but I have a feeling all those needed dies delayed the release of GK110...

Edit - Referring to all the GK110 GPUs needed for the Titan Supercomputer...
 
Last edited:
F

Fox5

Diamond Member
Jan 31, 2005
5,957
7
81
Nvidia is offering is basically do HSA with discrete GPUs. AMD already showed how much of a performance boost having a unified memory bus between CPU and GPU could provide, and nvidia is the first one with an offering to provide a high speed coherent bus between external GPU and CPU. It's like AMD Torrenza, except it's actually making it to market.
 
E

Erenhardt

Diamond Member
Dec 1, 2012
3,251
105
101
Fox5 said:
Nvidia is offering is basically do HSA with discrete GPUs. AMD already showed how much of a performance boost having a unified memory bus between CPU and GPU could provide, and nvidia is the first one with an offering to provide a high speed coherent bus between external GPU and CPU. It's like AMD Torrenza, except it's actually making it to market.
Click to expand...

Well, nv is not in HSA fundation. But I know who is: Samsung. They need to make extra sure to not get suit.
 
njdevilsfan87

njdevilsfan87

Platinum Member
Apr 19, 2007
2,342
265
126
KL isn't as mature either. NV dominates the compute market right now. Go to any university or government lab and what do they have in their nodes? A bunch of Xeons + Teslas. Intel sort of entered this game a little bit late in the sense that everyone who was interested in the potential of GPU compute went the AMD or NV route. Now everyone that knows how well GPU compute can work either knows AMD or NV and passes that knowledge down.

I wonder if Intel's sub $200 promo one their KL cards is going to help change that. I've ordered one just to see what I can do with it. At that price, why not? Won't put too much time into it, just whatever spare time I have to continue to expand my knowledge in this field.
 
Last edited:
Keysplayr

Keysplayr

Elite Member
Jan 16, 2003
21,219
54
91
Please folks I hate to be a pain, but can you take and extra sec or two to proofread your posts? It's making my brain hurt.
Not trying to be grammar nazi, but it only takes an extra few seconds.
-Keys :)
 
M

mavere

Member
Mar 2, 2005
194
14
81
xpea said:
Obviously, Intel Xeon + KL was one contender but it was discarded. The million dollar question is Why ?
From Nvidia own PDF :

but are these valid reasons ?
Click to expand...

Intel: "We already had a high profile design win, so no need to give a low-cost bid this time. We only have so much fab capacity."

IBM/Nvidia: "We need a high profile design win. We must win this bid."

DoE: "Intel already had a recent win. Don't want to look like we're favoring any one company."

I don't want to dismiss the technical nuances, but trying to analyze a governmental contracting decision from an engineering standpoint will just give you headaches. ;)
 
RussianSensation

RussianSensation

Elite Member
Sep 5, 2003
19,458
765
126
Lots of other data to take away here:

1) HBM confirmed for Volta
2) NV-Link confirmed for Volta (and possibly it might come out during Pascal)
3) Since this supercomputer is expected to be online in 2017, that means the time between Pascal and Volta would be relatively short (or the 2017 timeline is wrong).
4) It's nearly 2015 and GM200 (Maxwell flagship) isn't even out. Does this mean that NV plans on releasing flagship Maxwell, flagship Pascal and flagship Volta in a span of just 2-3 years (early 2015 to say late 2017)? Insane if true. Doesn't sound believable to me.

Big question: Does this mean NV will go straight from 28nm Maxwell GM200 (2015) to 16nm Pascal GP200 (2016) and then 14-16nm Volta (2017)? Not likely, which means there is likely a confusion in the article regarding what Volta is. Unless in this case the name "Volta" and "Pascal" are interchangeable, then it would make sense because originally NV had Volta+HBM for its 2016 roadmap.

Old NV roadmap -- Volta follow Maxwell, which makes sense for it to be used in this supercomputer around 2017.
OldRoadmap.jpg


New NV roadmap as of March 2014 -- Volta will follow Pascal, and Pascal is slated for 2016-2017.
PascalRoadmap.jpg


I think the source got it incorrect as in this case Pascal = Volta. 3D stacked memory (HBM), NV-Link and 2017 for Maxwell flagship successor all point to Pascal, not Volta. I think the real Volta is going to follow way later in 2018-2019 as a Pascal successor.
 
Last edited:
ShintaiDK

ShintaiDK

Lifer
Apr 22, 2012
20,378
146
106
2017? That kinda took the air out of the PR.

The chips going into it may not even exist out of simulators yet.
 
A

Alatar

Member
Aug 3, 2013
167
1
81
Another article on this subject also talked about per node flops numbers. Apparently ~40TFLOPs per node. Each node probably has 4 volta teslas.

njdevilsfan87 said:
Sorry not Knights Landing. The one that came before = Knights Corner. They're clearing out this specific model. Sorry.

http://www.sabrepc.com/intel-bc31s1p-xeon-phi-31s1p-8gb-knights-corner-coprocessor-passive.html

Has been less in other places but usually OOS.
Click to expand...

That's actually quite interesting. Might pick one up even though I'd probably really struggle with actually using it with the skills I have. Maybe in the future?

But it's cheap enough that it doesn't really matter, would be cool even as a collectible.
 
You must log in or register to reply here.

TRENDING THREADS

COMPANY
RESOURCES
FOLLOW
Top Bottom