SkylakeX teasers (aka prime95 29.5)

[preda]

Originally Posted by R. Gerbicz

Ofcourse I've spoken about prp with my error checking

[unrelated to the thread]
Robert, I would personally be interested in a small write-up about how you came up with the error check. A small "history" of the idea of the check.

[R. Gerbicz]

Originally Posted by preda

[unrelated to the thread]
Robert, I would personally be interested in a small write-up about how you came up with the error check.

off
First of all, without the Jacobi check (and their topic) I wouldn't have discovered my check. And it had some handicap, because the topic was in the GPU section, and I rarely read those topics (this has changed). My first impression when saw the Jacobi check that there could be something more on this area, because of the simplicity of the test.

For some days searched for another Jacobi check and other possible error check in the LL sequence. When for me these seemed dead-end, I have switched to work with Prp test, here the Proth numbers seemed easier, because there the exponent is N-1=k*2^n, or if you delay the k-th powmod (or even you can start with it), then you need to find the error check in a^(2^n) mod N. Actually it was easy to extend this to the general N=(k*2^n+c)/d form.

But return to the easier version, how to find the check:
let f(x)=a^(2^x) mod N, and in the following equations we'll omit writing the mod N to shorten the formulas.
from this we can see that
f(s+t)=f(s)^(2^t) for any s,t.

if we want to see a "check" at x=s, then what else could we do:
f(s)=f(s-t)^(2^t)
but this checks "only" the last t iterations at duplicated cost.
And mainly this is essentially the same thing that we redo the last t iterations.
But we can "continue" this, checking the previous t iterations in a block:
f(s-t)=f(s-2*t)^(2^t)
f(s-2*t)=f(s-3*t)^(2^t)
...
When we see such equations then the standard technique is that we multiple (or say add) all these equations:
f(s)*f(s-t)*f(s-2*t)*...=a*(f(s-t)*f(s-2*t)*f(s-3*t)*...)^(2^t)
if s is divisible by t.

Checking at every s, where s is divisible by t is still costly, so just delay it, the rest was pretty trivial.

additional short/easy examples
In general when we multiple/add equations the only goal is to solve a recursion, like in:
f(n)=f(n-1)+n^2, here you add(!) these equations.

Or with this we can also solve a system of (linear) equation in a nice way, like in:
y+z=a
z+x=b
x+y=c
this is coming from geometry (if a,b,c are the sides of a triangle), see: http://www.cut-the-knot.org/triangle/InExCircles.shtml

[petrw1]

…Sorry had to trim the picture to get under 1MB.

It was part of the standard Sad Face Blue Screen referring to windows stop codes

29.4.7 is running fine

[S485122]

Originally Posted by petrw1

…Sorry had to trim the picture to get under 1MB.
...

Even more efficient would be to just cite the error "WHEA_Uncorrectable_Error" of whatever. Less than 25 bytes !

Jacob

[chalsall]

Originally Posted by petrw1

It was part of the standard Sad Face Blue Screen referring to windows stop codes. 29.4.7 is running fine

This is on your brand new, seriously expensive and seriously bleeding edge machine. Correct?

Neither Prime95 nor mfaktc should crash your machine. Unless there is something fundamentally wrong...

Have you looked at your temperatures, your voltages, and your power draw?

Have you tried running your new kit under a Linux environment (to try to eliminate a driver issue)?

Have you tried running only Prime95, and then only mfaktc?

I'm sorry if I'm telling you how to chew gum. But when "making friends with new kit" it is important to change only one variable at a time when the kit is being obstinate.

[petrw1]

Originally Posted by chalsall

This is on your brand new, seriously expensive and seriously bleeding edge machine. Correct?
============= YES

Neither Prime95 nor mfaktc should crash your machine. Unless there is something fundamentally wrong...
============= AGREED....I did try the second version that was supposed to have fixed a memory corruption error (this would/could crash a machine)

Have you looked at your temperatures, your voltages, and your power draw?
============= NOT YET

Have you tried running your new kit under a Linux environment (to try to eliminate a driver issue)?
============= NOT THAT BRAVE....and SWMBO uses it also and would not approve LINUX

Have you tried running only Prime95, and then only mfaktc?
============= I did run Prime95 only first with the SkyLake version.....then backed off to 29.4 which runs fine.

I'm sorry if I'm telling you how to chew gum. But when "making friends with new kit" it is important to change only one variable at a time when the kit is being obstinate.
============ A appreciate all the help

P.S. Still some growing pains....they installed 32GB (8x4) but windows only sees 24GB.

Also mfaktc on the GPU "messes up" the display then "blue screens" the CPU ... after running about 1 minute.

[chalsall]

Originally Posted by petrw1

Also mfaktc on the GPU "messes up" the display then "blue screens" the CPU ... after running about 1 minute.

Just putting this out there...

On all of my machines which have GPUs, I have the motherboard drive the displays and the GPU(s) just run CUDA code (but, then, I'm not a "gamer").

It's possible there is a conflict in the code trying to drive the display and that trying to run compute.

P.S. I understand all too well the SHMBO dimension of the problem space. Personally I've found that handing off something to be used which doesn't work perfectly is rarely rewarding....

[petrw1]

Originally Posted by chalsall

Just putting this out there...

On all of my machines which have GPUs, I have the motherboard drive the displays and the GPU(s) just run CUDA code (but, then, I'm not a "gamer").

It's possible there is a conflict in the code trying to drive the display and that trying to run compute.

P.S. I understand all too well the SHMBO dimension of the problem space. Personally I've found that handing off something to be used which doesn't work perfectly is rarely rewarding....

Does this require a video card on the MB? Or do they all have this capability? And if so would we NOT lose a lot of screen speed or quality?

AND.....
I had to buy a new monitor because the ONLY video parts on the back of the case were HDMI ports from the GPU

[chalsall]

Originally Posted by petrw1

Does this require a video card on the MB? Or do they all have this capability? And if so would we NOT lose a lot of screen speed or quality?

Yes, no (but most do, look for video ports on the MB), and yes.

Originally Posted by petrw1

I had to buy a new monitor because the ONLY video parts on the back of the case were HDMI ports from the GPU

Are you sure about that?

Most Intel based MBs have HDMI and legacy VGA video ports, driven by the GPU contained within the processor (and is crap at OpenCL). Further, with the correct cable, you can drive a VGA monitor from a HDMI port.

I'm not suggesting this is the optimal solution, but you might consider trying driving your display(s) from a device separate from the GPU you want to run compute on, and see if the system is stable.

[Mysticial]

Originally Posted by chalsall

Yes, no (but most do, look for video ports on the MB), and yes.



Are you sure about that?

Most Intel based MBs have HDMI and legacy VGA video ports, driven by the GPU contained within the processor (and is crap at OpenCL). Further, with the correct cable, you can drive a VGA monitor from a HDMI port.

I'm not suggesting this is the optimal solution, but you might consider trying driving your display(s) from a device separate from the GPU you want to run compute on, and see if the system is stable.

Skylake X doesn't have onboard video.

[chalsall]

Originally Posted by Mysticial

Skylake X doesn't have onboard video.

OK. Thanks. I didn't know that.

I guess going "headless" for a diagnostic "deep dive" is going to be more difficult than expected....