I just saw a story on a German news site about a new, power-efficient supercomputer, which claims 5.27 GFlops/watt; that makes it roughly 20% more efficient than the current leader of the Green 500 which has 4.39 GFlops/watt. The owner's press release is (also in German) here. Their website supports English as well, but currently not for this press release — you might want to check again later today.
I was thinking, maybe this is the new race in power-computing. In the past, energy costs played a substantial role in operation of a datacenter. Lowering this prohibitive cost makes power-computing suitable for lots of new organisations, which might have a much bigger impact than new local power-houses.
This discussion has been archived.
No new comments can be posted.
Supercomputing: Now about Efficiency rather than Raw Power
|
Log In/Create an Account
| Top
| 12 comments
| Search Discussion
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
(Score: 1) by WillAdams on Friday November 21 2014, @03:13PM
Haven't the Chinese had a hard time finding customers for their new #1 supercomputer 'cause of the energy costs necessary to keep it running for a given period of time?
(Score: 2) by Dunbal on Friday November 21 2014, @06:39PM
Yes and lets judge drag racers by which car gets the best MPG...
(Score: 3, Insightful) by bob_super on Friday November 21 2014, @07:18PM
If your algorithm requires weeks, the right analogy would be the 24h Le Mans, in which good gas mileage actually saves you stops.
Supercomputers are not drag strip machines. They sacrifice quite a bit of potential performance for flexibility, reliability and stability.
(Score: 1) by Entropy on Friday November 21 2014, @07:42PM
Businessperson: Where is it going?
Hippie: We don't know..The supercomputer is working on it but will require 3 weeks to do the calculations.
Businessperson: The storm will be over in 3 weeks.
Hippie: But we're getting 5GFlops per watt!
Businessperson: What?
Hippie: We're green!
Businessperson: Who cares where is the hurricane going?!?!?!
Europe: It'll make landfall at these GPS coordinates. Our computer isn't green, suck it.
Personally in the world of supercomputing, I'm pretty sure speed matters. It's like trying to save wattage on hard drives where the green drives use 3 watts and suck at the good drives use 5 watts and perform well. Why is everyone trying to make everything about performance per watt?
(Score: 2) by bob_super on Friday November 21 2014, @08:27PM
I guess it depends on your customer base. Some want their results ASAP and will pay for it. Others want humongous datasets crunched pretty quickly at a decent cost.
If your target is the second case, perf per watt is important, because it controls ongoing cycles/$.
Let's not forget that a Watt has to be reliably brought into the building, and then reliably taken out. Both operations are typically costly and require significant infrastructure. In some cases wielding 20% or 50% less power can decide on the feasibility of the project in your preferred location.
(Score: 0) by Anonymous Coward on Friday November 21 2014, @08:47PM
Yeah, the thing is that speed is great, but most of us drive a 25+ MPG sedan to work rather than a rocket car. Why?
Costs matter.
(Score: 3, Informative) by takyon on Friday November 21 2014, @11:15PM
Performance per watt now matters almost as much as speed, or more, across most computing sectors.
The largest supercomputers: they want 1 exaflops within power targets ranging from 20-67 megawatts. A Green500 1 exaflops per 50 MW supercomputer would be 20 GFLOPS/W.
The Green500 List - November 2014 [green500.org]
Powering Up Exascale (2011) [hpcwire.com]
Exascale Requires 25x Boost in Energy Efficiency, NVIDIA’s Dally Says [hpcwire.com]
“No Exascale for You!” An Interview with Berkeley Lab’s Horst Simon [hpcwire.com]
Graphics cards: gamers may not care about power efficiency as more as other sectors, and graphics card performance is more important than CPU performance for gaming, but improvements to graphics card efficiency can enable more performance at a given wattage or cooler cards and systems (with smaller power supplies or form factors).
Nvidia Launches GTX 980 And GTX 970 Maxwell Graphics Cards [tomshardware.com]
Phones/tablets: obviously the more power efficient each part is, the longer battery life you can get from a mobile device. Intel is pitching its Broadwell Core-M [anandtech.com] 4.5 W chips as suited for thinner "fanless" tablet and ultrabook designs that are a lot faster than Intel Atom. They are fanless because at that TDP they don't need it and supposedly won't heat skin to some temperature they picked (marketing or health science, you can decide). The power efficiency improvements have allowed them to bridge the gap between chips that perform well, and "mobile" chips. This may allow Intel to eliminate Atom and compete better [tomshardware.com].
Wearables: much smaller batteries than phones/tablets. One of the major complaints against Google Glass, body cameras, non-epaper smartwatches? Not enough battery life for more than 6-12 hours of continuous or heavy use
Now think of it this way: increasing power efficiency does not necessarily mean sacrificing performance. If you have a 200 W chip with 1 teraflops of performance, and you replace that with a 200 W chip with 2 teraflops of performance, you just doubled efficiency from 5 GFLOPS/W to 10 GFLOPS/W. There are other considerations including cooling that factor in, but power efficiency is a good trend. Further improvements could enable 3D stacked chips which would be too hot to operate now, or wearable/IoT/medical devices that are smaller and more useful than they are now.
(Score: 1) by jmorris on Saturday November 22 2014, @02:04AM
Green is not the only factor but it is A factor when designing a supercomputer. Even when raw performance is the goal.
Break it down; you have X budget to build a supercomputer and keep it going for the typical three year primary life of one before a refresh will be justified. You want the maximum grunt you can afford but going overbudget in the real world brings problems. You can buy fast, hot nodes but those raise the power and cooling budget, remember that over three years the cost of keeping a flaming node lit up and cooled is very non-trivial. You can buy really green nodes but you will need more physical space for more nodes and probably spend more on interconnect hardware and even more nodes to make up the losses in compute efficiency over and above the rated speed. The sort of job you are running will also impact your decision, some jobs will scale over a lot of wimpy nodes better, some not so much.
This is why supercomputers are still one off products, each one carefully designed for their task.
(Score: 2) by maxwell demon on Saturday November 22 2014, @04:48PM
Businessperson: "You say the computing cluster costs us a million per day. How come?"
Speed geek: "Most of it is electricity cost."
Businessperson: "Aren't there solutions which consume less power?"
Speed geek: "Yes, there are those so-called green computing clusters. But they provide less computing power."
Businessperson: "But according to my data our computing cluster is idle 90% of the time. So why do we need that computing power?"
Speed geek: "Because we get our results faster that way. For example, remember the last week when you asked us to calculate the aerodynamics of the new car design?"
Businessperson: "Yes, of course. I asked you to give me the data by the next week."
Speed geek: "And we could give you the result in less than two days. That clearly shows the power of our cluster."
Businessperson: "And how long would it have taken with one of these green computing clusters you mentioned?
Speed geek: "Three days. That's significantly longer."
Businessperson: "But still very much below the week I had given you. So how much power cost would be safe with a state of the art green cluster?"
Speed geek: "About four hundred thousand dollars per day."
Businessperson: "So, the green cluster would cost us less, and yet deliver the results fast enough. So why don't we have a green cluster?"
Speed geek: "Because it's slower."
Businessperson: "So we pay four hundred thousand dollars per day just so you can feel proud of the speed of the computing cluster which we don't actually need?"
Speed geek: "Err ... I wouldn't say it that way ..."
Businessperson: "Well, OK, let me reformulate it: You're fired for wasting the company's money. Security, please lead this man to the door!"
The Tao of math: The numbers you can count are not the real numbers.
(Score: 2) by takyon on Friday November 21 2014, @10:33PM
The November 2014 [green500.org] list is out. The leader?
5,271.81 MFLOPS/W - GSI Helmholtz Center L-CSC - ASUS ESC4000 FDR/G2S, Intel Xeon E5-2690v2 10C 3GHz, Infiniband FDR, AMD FirePro S9150
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Saturday November 22 2014, @10:13AM
1) It doesn't work at all without JavaScript
2) It serves PDF files as Content type: text/html
(Score: 2) by maxwell demon on Saturday November 22 2014, @04:59PM
I have no idea which web site you speak of. The summary links to three different web pages. I could load each of those without enabling JavaScript or access to third-party sites (which I disable by default), although for the GSI one I had to scroll down quite a bit to get at the content). None of the linked pages are PDF files (I even looked at the page source code to be sure).
The Tao of math: The numbers you can count are not the real numbers.