Slashdot Mirror
$30M Stampede 2 Supercomputer To Provide 18 Petaflops of Power To Researchers Nationwide (techcrunch.com)
An anonymous reader writes: Funded by grants from the National Science Foundation and built at the University of Texas at Austin, the Stampede 2 supercomputer looks to contend with the global supercomputer Top 5. With 18 petaflops of processing power, it aims to help any researcher with a problem requiring intense number crunching. For example, atomic and atmospheric science simulations would take years to work-out on a desktop PC but only days on a supercomputer. Texas Advanced Computing Center director Dan Stanzione said in a UT press release, "Stampede has been used for everything from determining earthquake risks to help set building codes for homes and commercial buildings, to computing the largest mathematical proof ever constructed." The Stampede 2 is about twice as powerful as the original Stampede, which was activated in March of 2013. Instead of the 22nm fabrication tech in the original Stampede, the Stampede 2 will feature 14nm Xeon Phi chips codenamed "Knights Landing" forming 72 cores compared the original system's 61 cores. With double the RAM, storage and data bandwidth, the Stampede 2 can shift up to 100 gigabits per second, and its DDR4 RAM can perform fast enough to work as a third-level cache as well as fulfill ordinary memory roles. In addition, it will feature 3D Xpoint non-volatile memory. It will be at least a year before the Stampede 2 is powered up since it just received funding.
"It will be at least a year before the Stampede 2 is powered up since it just received funding" That's some real hard news there Slashdot. Thanks.
So this new computer is twice as powerful as the original Stampede, which cost $27 million and came online in 2013. It sounds like Stampede 2 will come online in 2018 at a cost of $30 million. So over 5 years the power only doubled for the same cost - if computing power per cost unit doubles every 18 months like Moore predicted, why isn't Stampede 2 like 16 times more powerful than its predecessor?
Enigma
This page claims:
The Apollo Guidance Computer weighed 70 pounds, consumed 55 watts of power and occupied only 0.97 cubic feet inside the spacecraft. These first digital flight computers were limited to only 36,000 words of fixed memory and 2,000 words of RAM, and operated at a 12-microsecond clock speed.
That was 16 bit words. 12 microseconds is about 83K-ops/sec. Unclear how many cycles each operation took, but let's give it the benefit of the doubt and say the throughput was instruction per clock. (Probably optimistic).
Those 18-petaflops are probably 32 bit, if not 64 or 80. Let's say each is worth two of the AGC 16-bit ops.
So, 18 petaflops would be similar to 4.3 x 10^11 apollo guidance computers. You can probably swing that by 100X in either direction depending on the assumptions you make, but let's roll with it.
Each AGC weighed 32 kg. 4x10^11 of them would weigh in at ~ 1.4x10^13 kg. For comparison, this is roughly 35 times the combined mass of the entire human population.
But how fast can it mine Bitcoin?
It is not an appropriate architecture for bitcoin mining. Much of the design is focused on moving massive amounts of data in, out, and between the cores. You don't need any of that for bitcoin, since each hash is independent and uses very little data. Bitcoin mining can be done way more cost effectively on ASICs.
Since TFA is about a Super Computer that has yet to be built, please allow me to introduce two links to another Super Computer that has yet to be built
http://www.nextplatform.com/20...
http://www.nextplatform.com/20...
I realize that DDR4 is pretty fast; and given how relatively cheap it is it is certainly good for what you pay; but is it honestly anywhere close to being comparable to L3 cache? I would think that, even if you spared absolutely no expense and went with the fastest RAM money could buy, signal propagation delays would, at contemporary clock speeds, make system RAM tens of cycles slower to access than anything on die.
How much computing power is wasted/used up by crappily designed programs that anticipate having lots of RAM access?
Only 100 gigabits can be moved per second? You can't move one bit per clock cycle or even every four?
Still waiting on Serviscope_minor to wake up to fucking reality and realize that Jessica Price isn't going to fuck him.
If half of the AI hype was true the Beowulf cluster would imagine YOU!!!
Confucius say, "Find worm in apple - bad. Find half a worm - worse."