|
2018-08-22, 08:39
|
#1 |
Jun 2003
546410 Posts |
mfaktc performance on Volta - Is this for real?
So I noticed today that the mfaktc performance numbers for Volta has been revised WAY up @ http://www.mersenne.ca/mfaktc.php?sort=gpw. IIRC, it used to be something like mid to high 1xxx range for GHz-days/day and now they are at twice the values. Consequently, power efficiency is also off the chart.
Comments? |
|
|
|
2018-08-22, 12:48
|
#2 |
Sep 2003
1010000111102 Posts |
In this post from June, a 28.83 GHz-days calculation completes in 10m 23.287s, which is 3996 GHz-days/day.
And a 20.57 GHz-days calculation completes in 7m 28.293s, which is 3964 GHz-days/day. So that does seem consistent with the higher number currently shown at the link you posted. |
|
|
|
|
2018-08-22, 13:29
|
#3 | |
|
Jun 2003
23×683 Posts |
Quote:
|
|
|
|
|
|
2018-08-22, 14:41
|
#4 |
|
Einyen
Dec 2003
Denmark
345210 Posts |
I did a few trial factoring on the EC2 p3.2xlarge with Tesla V100.
In the 100M range from 75 to 76 bits was completing in 23 min, and it was running consistently at 4700+ Ghz-days/day. If I did 76-77 bits or 77-78 bits it dropped to like 4300-4400 Ghz-days/day. For the 100M the fasting settings in the .ini file was: GPUSieveSize=128 GPUSieveProcessSize=8 GPUSievePrimes=90000 Last fiddled with by ATH on 2018-08-22 at 14:46 |
|
|
|
|
2018-08-22, 19:19
|
#5 |
|
Sep 2003
259010 Posts |
Were the Voltas favored by Ethereum miners, and if so, are there any signs yet of fire-sale pricing for them?
|
|
|
|
|
2018-08-22, 22:12
|
#6 | |
"/X\(‘-‘)/X\"
Jan 2013
https://pedan.tech/
C7016 Posts |
Quote:
|
|
|
|
|
|
2018-08-23, 11:04
|
#7 |
"Victor de Hollander"
Aug 2011
the Netherlands
32×131 Posts |
Tesla V100, aren't those 10 000 dollar server GPUs?
|
|
|
|
|
2018-08-23, 12:09
|
#8 | |
|
Sep 2003
A1E16 Posts |
Quote:
You can rent them for about $0.95 / hour on AWS. Which is not particularly cost-effective, compared to $0.019 / hour for a Skylake core. But you can play with them for a few hours and sigh. |
|
|
|
|
|
2018-08-23, 14:21
|
#9 |
|
Einyen
Dec 2003
Denmark
22×863 Posts |
The Titan V at $3000 has close to the same performance, 3400+ Ghz-days/day according to the chart, but still a lot of money, and they cost close to $3900 in Denmark.
According to this: https://en.wikipedia.org/wiki/Volta_(microarchitecture) it has 13,8 TFLOPS SP and 6,9 TFLOPS DP compared to 14 and 7 for Tesla V100 PCI-E: https://en.wikipedia.org/wiki/Nvidia_Tesla Last fiddled with by ATH on 2018-08-23 at 14:24 |
|
|
|
|
2018-08-25, 20:48
|
#10 |
"Oliver"
Mar 2005
Germany
5·223 Posts |
Yes, those number are real. And power efficiency is even better because those cards are nowhere near their TDP while running mfaktc...
There is no magic included and CUDA 9.2 isn't the secret. CUDA 9.2 is the first version of CUDA which can compile mfaktc for Volta, earlier versions had compiler bugs. Volta has a 32bit wide integer multiplier just as Fermi and Kepler have. Maxwell and Pascal only have a 16bit wide integer multiplier in hardware and thus need 4 16bit wide multiplications to perform a single 32bit wide multiplication. Oliver |
|
|
|
|
2018-08-26, 01:20
|
#11 |
|
Bemusing Prompter
"Danny"
Dec 2002
California
47108 Posts |
That is really cool. I wonder how the new GeForce 20 series performs.
|
|
|
|
 |
Similar Threads
|
||||
| Thread | Thread Starter | Forum | Replies | Last Post |
| GTX 1180 Mars Volta consumer card specs leaked | tServo | GPU Computing | 20 | 2018-06-24 08:04 |
| Nvidia announces Volta based Titan V | tServo | GPU Computing | 8 | 2017-12-10 16:04 |
| Nvidia announces Volta GPU | tServo | GPU Computing | 2 | 2017-05-13 07:58 |
| Is this for real??? | kurtulmehtap | Math | 2 | 2014-09-29 14:16 |
| Is this guy for real? | CRGreathouse | Math | 21 | 2010-08-23 18:05 |
