New SuperCalculator
#1
Hi,

Just for info', fun & dream !

His name is : Sunway TaihuLight

He got ; 10 649 600 core / power : 93014,6 Tflops ( all detail in the link )

What we can expected in term of bench with that type of Machine ????

======> http://www.top500.org/news/new-chinese-s...p500-list/

Cheers,

Tony
Reply
#2
I'm not sure what you want to here.
Hashalgorithms use fixed-point-operations (integer) IIRC, so the 93 PFLOPS don't say anything about possible hash cracking speed. But maybe I overlooked something in the article that leads to an assumption about the speed.
Reply
#3
According to this, each ShenWei SW26010 processor, and this has 40k-odd, is comparable to an upcoming Intel Xeon Phi Knight's Landing CPU. 
Sadly, I don't keep up with hardware very much, certainly not the HPC stuff, so that means very little to me. 

If we assume that GPUs compare directly to these processors - which they don't, read what stepMode wrote - and that FLOPS gives a good idea of how well hardware will run hashcat - which it doesn't - then we get, using the epixoip's Titan X benchmarks as an example, from MD5: 16900 MH/s
According to this, a Titan X does about 7TFLOPS (I'm aware that this is SP, DP is far less, and that it doesn't matter anyway since hashcat does integer maths), so that's about 2414 MH/s / TFLOPS. 

This thing does 93,000 TFLOPS, so that gives a grand total of 224,502,000 MH/s, or 225 TH/s, which, as some friends of mine like to say, is "not too shabby". About 13k Titan Xs. 

How good is this estimate? Well, you saw everything I said about my assumptions being worse than bullshit. It's probably correct, to within 4 or 5 orders of magnitude. 

Lets not even think about having to compile hashcat for this, or the bottlenecks that storage and everything else would introduce, and just fantasise about having this baby to run for a few days.
Reply
#4
You know, 225 TH/s is a far more realistic and educated guess than this motherfucker's "10 million cores that each crack at 10 billion passwords per second" (essentially saying it can do 100 PH/s): https://www.linkedin.com/pulse/which-com...n-buchanan
Reply
#5
Thank you guy's for your return....well done, is more clear for me now....appreciated the detail...maybe "rendez vous" in 5, 10 years so ???
Reply
#6
What I wanted to do was find an Intel CPU that has similar performance and has been benchmarked on hashcat; it's still not ideal but it would be a lot better than what I've done above. 

A different way of looking at it might be to use a CPU example from the bottom of https://hashcat.net/hashcat/. Info about Knight's Landing was a bit hard to find, with Wikipedia claiming it was anounced 2 or 3 years ago, and Intel saying some of them are still to be released; but they're given as having 64 (ish) cores, and somehow 4 times as many threads (hyperthreading but better?). If each thread can do as much as one from a Phenom II X6 T1090 @ 3.8 GHz from the link above (those benchmarks are outdated by the way, says it used hashcat 0.40) then:

total power = 40960 CPUs * equivalent Intel power per CPU
= 40960 * 12.9 MH/s * 64 cores * 4 threads each 
= 135,000,000 MH/s

Which is actually relatively close to what I had before! 

This is an overestimate, since these Xeon Phi's (and the Chinese things) don't run at 3.8 GHz, but rather at about 1.5 GHz - I assumed each SW CPU can run hashcat just as well as a single Xeon Phi, and that a Xeon Phi thread is as fast as a single Phenom thread; I realise this is probably a gross overestimate. 

It does however mean that my earlier estimate is about right, within about a single order of magnitude, which I'm happy with, and that that bloke from epixoip's link really is pulling numbers out of his arse. 

More numbers: 

- 135,000,000 MH/s while pulling 15.3 MW gives about 8.8 MH/s / W
   A GTX 1080 does about 25,000MH/s at 183 W (from epixoip's bench) giving a whopping 136 MH/s / W !
   We can conclude that the Chinese didn't build this thing for hash cracking efficiency. 

- It has 1.31 PB of "primary memory", that's about 1030 MH/s / GB; a 1080 does about 3000
  I have no idea if this a useful statistic (insofar as any of this is "useful" - just a bit of fun for me).

this says it cost about USD 270M total - Building, hardware, software and R&D
  this comes out very nicely to exactly 0.5 MH/s / $
  I've no idea how much a functional system using a 1080 costs, but if we take it to be USD 2000 then that's about   12.5 MH/s / $, or 25 times better. 

Conclusion? It's really really fast, but you'd be better off using lots and lots of 1080s if you want perf/dollar or power efficiency.
Reply