Informal Benchmark HW C vs HW D

[CyberAngel]

(10-14-2018 09:44 PM)akmon Wrote:  

(10-13-2018 08:50 AM)CyberAngel Wrote:  Time the Inversion ONLY (leave the display out):

Code:


#cas
InvMat1(n):=
BEGIN
LOCAL t,Mat1;
  Mat1:=randMat(n,n)*0.1;
  MSGBOX(Mat1); // debugging
  t:=TICKS; Mat1:=Mat1^-1;
RETURN (TICKS-t)/1000.+"_s";
END;
#end

Usage: InvMat1(50) [ Enter ]
see the random matrix, for the inverse timing press again [ Enter ]
or tap | OK |
Press directly again [ Enter ] tap | OK |
and one more time and the fourth time to get the average

Now increase to 100 and do it 4 times, too by pressing the [ Enter ]

Try four times 200
Does anyone have also the G2?

These are my results:

G2:

Invmat1(50) average: 0,0135s
Invmat1(100) average: 0,0795s
Invmat1(200) average: impossible to test. Calculator crashed, only restart with a clip. reported to calcbeta

G1:

Invmat1(50) average: 0,1665s
Invmat1(100) average: 1,113s
Invmat1(200) average: it shows a strange error message but it can continue doing the calculus: 8,472s.

With eight times the RAM
and perhaps a eight times the MatInv speed
it's a shame that the G2 crashes.

Can anyone try on their phone up to 800, please.

[StephenG1CMZ]

On Android it can reach up to 141 before crashing.

[akmon]

(10-15-2018 05:52 AM)parisse Wrote:  

(10-14-2018 09:00 PM)akmon Wrote:  I´ve just executed the python mendelbrot program, in G2 version, chronograph on hand.
So taking the time in the video, these are the results:

Prime G1: 140 seconds
Numworks: 85 seconds
Prime G2: 57 seconds

Try this one:

Code:


#cas
def fra1(X,Y,Nmax):
  w=2.7/X
  h=-1.87/Y
  Y=Y-1
  for y in range(ceiling(Y/2)+1):
    c = -2.1+i*(h*y+0.935)
    for x in range(X):
      z = 0
      for j in range(Nmax):
        if abs(z=z**2+c)>2:
          break
      PIXON_P(x,y,5100*j+512)
      PIXON_P(x,Y-y,5100*j+512)
      c = c+w;
  FREEZE
#end

fra1(320,220,10)[/code]

G1: 66 s
G2: 16 s
Numworks: ?

With this code, the program is quite faster, on both models.

[akmon]

(10-15-2018 02:01 PM)StephenG1CMZ Wrote:  On Android it can reach up to 141 before crashing.

On physical calculators
G1
InvMat1(800): makes a warmstart after 25 secs.

G2
InvMat1(800): crashes


About the example above:
In G1
InvMat1(141): average: 2,988s
In G2:
InvMat1(141): average: 0,21s

In G1
InvMat1(142): after a strange error message, average: 3,083s
In G2:
InvMat1(142): crashes

So it confirms that the crashes happens with n>141

hope this data will help to HP team.

[John Keith]

If any size greater than 141 causes a crash, there must be a limit of 20000 elements somewhere, since 141 is IP(SQRT(20000)).

Also an 800 x 800 matrix requires 64 times as much memory as a 100 x 100 matrix, not 8 times. I doubt that any hand-held device could invert an 800 x 800 matrix but I'm prepared to be pleasantly surprised.

[Albert Chan]

(10-15-2018 09:32 PM)akmon Wrote:  In G1
InvMat1(142): after a strange error message, average: 3,083s

I know the inverse were not displayed, but is it any good ?
I think it is better for a benchmark do double duty, timing and verifying.

Instead of display nothing (but timing), we could add this to InvMat1:

err := Mat1 - inv(inv(Mat1)) ;
print(dot(err, err)); // dot product, so just 1 (hopefully small) number

[parisse]

You can do that for matrix inverses (no display, time and check):
a:=ranm(150,150)*0.1:;
time(b:=inv(a));
maxnorm(a*b-identity(a))

[John Keith]

(10-16-2018 05:23 AM)parisse Wrote:  You can do that for matrix inverses (no display, time and check):
a:=ranm(150,150)*0.1:;
time(b:=inv(a));
maxnorm(a*b-identity(a))

On my Prime (hardware version C) around 3.65 seconds.

On the emulator says 0.013 seconds, which seems impossible (7 year old Win7 laptop).

Both running latest beta firmware.

[CyberAngel]

(10-21-2018 02:35 PM)John Keith Wrote:  

(10-16-2018 05:23 AM)parisse Wrote:  You can do that for matrix inverses (no display, time and check):
a:=ranm(150,150)*0.1:;
time(b:=inv(a));
maxnorm(a*b-identity(a))

On my Prime (hardware version C) around 3.65 seconds.

On the emulator says 0.013 seconds, which seems impossible (7 year old Win7 laptop).

Both running latest beta firmware.

On my Prime (hardware version A around 3.713 seconds.
( Other runs: 3.71 ; 3.723 ; 3.72 ; 3.721 )

On the emulator says 0.024 seconds, which seems impossible (7 year old Win10 laptop).
( Other runs: 0.012 ; 0.0199 ; 0.0187 ; 0.0191 )

Former running 2.0.0.13865, latter running 2.1.14091 beta firmware.
AND
Xiaomi Mi Max 2 running HP Prime Pro 13443 : time= 0.532 seconds.
MIUI 9.6.3.0 Global Stable, Android 7.1.1, patch level 2018-06-01
( Other runs: 0.461 ; 0.434 ; 0.434 ; 0.425 ) - the duplicate is correct!

VPN

[parisse]

I confirm that inverting a 150x150 float matrices takes about 0.01s on my old laptop (7 years old) with Xcas. The emulator timing is perhaps correct (I don't know how precise it is).

[akmon]

(10-15-2018 09:32 PM)akmon Wrote:  

(10-15-2018 02:01 PM)StephenG1CMZ Wrote:  On Android it can reach up to 141 before crashing.

On physical calculators
G1
InvMat1(800): makes a warmstart after 25 secs.

G2
InvMat1(800): crashes


About the example above:
In G1
InvMat1(141): average: 2,988s
In G2:
InvMat1(141): average: 0,21s

In G1
InvMat1(142): after a strange error message, average: 3,083s
In G2:
InvMat1(142): crashes

So it confirms that the crashes happens with n>141

hope this data will help to HP team.

After doing the same example with 14181 firmaware, InvMat(142) still crashes, but at least it restarts automatically, not needed to use a clip on its back. Curious, before restarting it warns you: "your unit will reboot in 3 seconds".