World’s top supercomputer from ‘09 is now obsolete, will be dismantled
arstechnica.com
arstechnica.com
And these data centers can be located wherever is convenient. And clearly, forward planning is not an issue when we are talking $100mm+ investments.
[1] http://www.reedconstructiondata.com/rsmeans/models/data-cent...
Add to that the fact that utilization in most DCs is terrible. A super computer will throw out significantly more heat than your "traditional" DC.
Let's not even consider the op. ex. of moving the staff working who would work on the new facilities.
Edit: And LOL at those figures being accurate. That's just the building. What about the costs for the fibre, all that copper, power infrastructure, cooling, false floors, air filtering, special construction needs such as sunken well floors (to contain any leaks from the cooling systems, for example).
Assuming they already paid for the replacement super-computer too etc. Otherwise, sure it matters. Instead of finding $xx-$xxx million to build a more efficient one, you might be better off to pay the electricity bill difference. This is still in the top 25 computers in the world, so it can be used for a lot of things, but we don't know all the details.
I do not know if RR will be physically shredded, but I'm hoping that it is retired to the brand-new PRObE (http://nmc-probe.org/) computing center where RR would be available for use by academic researchers.
The hard drives will probably be sent to a military base to be melted down, I don't think they risk shredding them nowadays. The hardware that doesn't store data will probably be kept intact for quite some time.
Since Roadrunner takes up about 1/3 of the Metropolis datacenter, getting rid of it leaves some room for the 2015 "Trinity" machine (assuming LANL hosts it).
You're a bit right in that it's more complicated than they say in the article. The power of keeping it up and running is definitely costly, but there's more to it than that.
* It takes people to run it. They're expensive, and you don't want your people wasting a lot of time on the old super while your new one is being installed and accepted. I don't recall having seen any plans for a new super at LANL, but there's probably something new showing up soon.
* The architecture of roadrunner is outdated and obsolete. The NNSA isn't buying any more, and you can bet that with the new Blue Gene/Q at Lawrence Livermore (sequoia), all their new code is being written for different non-roadrunner architectures. There is considerable cost to porting new code to legacy hardware.
* Kind of in line with the last point, sequoia is >17x faster than roadrunner at only 3.5x the power. There's not a lot of point in paying such a premium in flops/watt by running on roadrunner since the new one is already online.
* 2.3MW is only what the computer itself uses. There's a ton of other supporting equipment needed to use it (chillers, disk, control nodes, etc.)
* This computer isn't being manufactured anymore, so maintenance is going to cost a fortune. When they bought it, they either paid for a hardware support contract or bought a ton of spares and did self-maintenance. If they did the former, IBM is either charging through the nose for replacements or refusing to support it outright. If they did the latter, they're probably running low on spares.
If it's cheaper to build (or rent) and run a new supercomputer, then it makes sense to dismantle the old one.
Retiring Road Runner may be significantly cheaper than adding additional megawatts of capacity, particularly when considering redundant power services and backup power supplies.
[1] Plug: That's also what I'm working on currently: https://github.com/muellermichel/Hybrid-Fortran
At this point the 120M investment is a sunk cost. And the investment was so bad that the operational income can not even cover just the operational expenses, so shutting it down is the wise thing to do.
"Oh, but there's a cost to support obsolete hardware!" Yeah, sure, but the person supporting everything was me, and I was a constant cost to keep around whether I supported crappy obsolete hardware or shiny new hardware.
Heck, some of the developers might prefer to use a Haswell+7990 in a few months over time on those things... if they can get one on their desk.
Meanwhile in the colo world, everything up to Core 2 Xeons were being tossed out in lots when Nehalem came out because it was just that much more efficient.
As far as I know most passwords hashes are also integer operations.
The cell processor in the ps3, despite being optimized for floating point ops, can do about 22.23 Mh/s. Assuming each of the 12,960 PowerXCell 8is in Roadrunner can do the same: 12960 * 23.23 ~= 288100 Mh/s which would earn you about 21.6 BTC/day at the current difficulty. That's about $2000/day at current prices.
The PowerXCells also have 8 SPUs vs 6 in the ps3, so they are potentially 25% faster.
This was the first petaflop machine. A petaflop is a million gigaflops, or ten million current cellphones. That's roughly 23 doublings in performance. If Moore's Law were to continue at one doubling per two years, that would be 46 years, or about 2059. Another ten or fifteen years might pass before low-end cellphones have it.
But 46 years is a long time. 46 years ago was 1967. There were no packet-switched computer networks. Nobody had ever seen a video arcade. Computers mostly interacted with people via punched cards, or at best a teletype. The Bloom filter had not yet been discovered. Timesharing and high-level languages were radical, risky concepts; the popular high-level languages of the time were all domain-specific, not general-purpose like C. There was no such thing as a mouse. There were graphical user interfaces and virtual reality. Humans had never walked on the moon.
46 years from now, there may not be cellphones, or humans. We certainly won't still be doubling the density of planar silicon chips every 18 months.
Yes, but they were 2 years away. We are now 41 years from the last moon mission.