New G2 HP Prime

[grsbanks]

That does sound plausible. Maybe you'd want to re-use your G2 memory management code in the previous Prime hardware versions? That could be a cool boost for what is already a great machine.

[Claudio L.]

(08-29-2018 06:53 AM)cyrille de brébisson Wrote:  Hello,

I was very surprised to see the speed difference between the G1 and G2.
CPU frequency of G2 is faster (about 30% from memory). And it has more cache and a better micro architecture... But nothing that warant a *2 or more extra speed (no super scalar for example).
It's RAM is better, yes, but in a tight loop benchmark, only the cache is used, so it does not mater...

I think the CPU architecture change alone can easily give you a 2x speedup. From Wikipedia, the old ARM920T runs 200 MIPS at 180 MHz, or 1.11 MIPS/MHz, while a Cortex A7 is at 1.9 MIPS/MHz, that's 70% speed increase at the same clock. Take into account the clock increase and you are at 2.2x.

EDIT: Sorry, the ARM920T is the 50g, the ARM926EJ shows 220 MIPS at 200 MHz, still the same 1.1 MIPS/MHz.

[toml_12953]

(08-29-2018 06:53 AM)cyrille de brébisson Wrote:  My best guess is that the extra speed come from a better OS.
Cyrille

But isn't the same revision firmware used by both?

[grsbanks]

(08-29-2018 01:39 PM)toml_12953 Wrote:  

(08-29-2018 06:53 AM)cyrille de brébisson Wrote:  My best guess is that the extra speed come from a better OS.
Cyrille

But isn't the same revision firmware used by both?

Some parts differ significantly. One is about twice the size of the other!

[Anders]

(08-29-2018 06:53 AM)cyrille de brébisson Wrote:  Hello,

I was very surprised to see the speed difference between the G1 and G2.
CPU frequency of G2 is faster (about 30% from memory). And it has more cache and a better micro architecture... But nothing that warant a *2 or more extra speed (no super scalar for example).
It's RAM is better, yes, but in a tight loop benchmark, only the cache is used, so it does not mater...

Extra speed in the CAS, when it comes to floating points (invert a 100*100 floating matrix for example), or in the advanced grapher/3D grapher are easely explained by the presence of the FPU (which is used by these 3 items)...

But extra speed in Home is not as Home does not use floats.

My best guess is that the extra speed come from a better OS.
First G2 has way way less threads running around. Saving a lot of CPU time.
But more importantly, the OS did not come with malloc and I had to recreate the whole memory allocation system, and I think that my code is way better than the BestaOS one.
Since both PPL and CAS programs are HUGELY dependent on mallocs, this probably makes up for a lot of the speed up...

Cyrille

I missed that the f/w's are different (even though the have the exact same version numbers).
Regardless it's very well done to get so much improvement from s/w and h/w! Shows it's just not a straight port.

[BruceH]

(08-18-2018 06:16 AM)mfleming Wrote:  This may be the wrong thread or even the wrong venue, but is there a chance that HP might encourage third-party (likely, commercial) development on the Prime hardware platform? I'm thinking something along the lines of the old HP-41 "Blanknut" platform offering. The idea of a customizable handheld with such obvious capability aimed at niche markets would excite quite a crowd of developers.

Hi Mark.

There is precisely zero chance of a modern blanknut because there is precisely zero market for such a device. All my clients use mobile/tablet apps for their field workers because of the built-in wifi/4G connectivity (saves having to go back to an office or depot every day) and remote wipe (erase the device when lost/stolen) capabilities, to name just two.

Shame though.

[CyberAngel]

(08-29-2018 06:53 AM)cyrille de brébisson Wrote:  Hello,

I was very surprised to see the speed difference between the G1 and G2.
CPU frequency of G2 is faster (about 30% from memory). And it has more cache and a better micro architecture... But nothing that warant a *2 or more extra speed (no super scalar for example).
It's RAM is better, yes, but in a tight loop benchmark, only the cache is used, so it does not mater...
<CLIP>
Cyrille

Oh, I wish we could have a REAL processor inside the Prime.
Maybe not quite SD 1000, but a ten times faster one with much better RAM
Some buy the more expensive iPhone
Why not a HP Prime FAST?!
twice the price - ten times the speed!!!
VPN
PS: I would love to have three index or 3D Matrices
Do you offer a C Development Framework & Toolset for developers?

[Anders]

(08-30-2018 12:40 AM)CyberAngel Wrote:  

(08-29-2018 06:53 AM)cyrille de brébisson Wrote:  Hello,

I was very surprised to see the speed difference between the G1 and G2.
CPU frequency of G2 is faster (about 30% from memory). And it has more cache and a better micro architecture... But nothing that warant a *2 or more extra speed (no super scalar for example).
It's RAM is better, yes, but in a tight loop benchmark, only the cache is used, so it does not mater...
<CLIP>
Cyrille

Oh, I wish we could have a REAL processor inside the Prime.
Maybe not quite SD 1000, but a ten times faster one with much better RAM
Some buy the more expensive iPhone
Why not a HP Prime FAST?!
twice the price - ten times the speed!!!
VPN
PS: I would love to have three index or 3D Matrices
Do you offer a C Development Framework & Toolset for developers?

Or generally, a support for tensor calculus would be nice ... maybe even up to type (2,2) and type (1,3)-tensors.

[CyberAngel]

(08-30-2018 02:08 AM)Anders Wrote:  

(08-30-2018 12:40 AM)CyberAngel Wrote:  <CLIP>
PS: I would love to have three index or 3D Matrices

Or generally, a support for tensor calculus would be nice ... maybe even up to (2,3) type tensors.

Yes, indeed!
I think that Bernard Parisse could possibly do it
along with the needed index manipulations commands (among others)
I could easily wait a year as long as I know it will be done,
VPN
PS: I still want a wide & fast ARM CPU with 10*speed potential
HP Prime Flash

[grsbanks]

Does anyone know if Klaas is OK? I ordered a G2 Prime this weekend, was told that it would be sent on Monday and that I would get an e-mail when that happened. It's now Thursday morning in the UK, I've not had that e-mail and a private message I sent Klaas 36 hours ago is as yet unread!

[pier4r]

(08-30-2018 12:40 AM)CyberAngel Wrote:  Oh, I wish we could have a REAL processor inside the Prime.
Maybe not quite SD 1000, but a ten times faster one with much better RAM
Some buy the more expensive iPhone
Why not a HP Prime FAST?!
twice the price - ten times the speed!!!
VPN
PS: I would love to have three index or 3D Matrices
Do you offer a C Development Framework & Toolset for developers?

You make it too simple. With words everything is simple. What is a REAL processor? The prime (or the 50g) have real processors, if the system is locked (or the skills are missing) one cannot really use it.
The prime processor is really fast, if the algorithm you use is good. With poor algorithms not even supercomputers can help.

Then again over time one can read a lot of "I will use C and I will do everything with it and also Coffee" but then very few people do anything with C or similar low level languages. See the amount of discussion about C with HPGCC for example. Only few (gold) project are out there. See Egan (didactic) or Claudio (newRPL, hpgcc), the majority of people don't use them or don't even know them.

Therefore the effort to put a better processor or produce a better SDK is not worth the result, as the results are practically absent aside from people asking for features that they won't use. But the latter is easy.

[CyberAngel]

(08-30-2018 02:55 AM)CyberAngel Wrote:  

(08-30-2018 02:08 AM)Anders Wrote:  Or generally, a support for tensor calculus would be nice ... maybe even up to (2,3) type tensors.

Yes, indeed!
I think that Bernard Parisse could possibly do it
along with the needed index manipulations commands (among others)
I could easily wait a year as long as I know it will be done,
VPN
PS: I still want a wide & fast ARM CPU with 10*speed potential
HP Prime Flash

To the Real People out there:
Do you want a slower or faster calculator?
(instead of a wisecrack answer)

[Albert Chan]

(08-30-2018 01:48 PM)CyberAngel Wrote:  To the Real People out there:
Do you want a slower or faster calculator?
(instead of a wisecrack answer)

I used to chase the fastest computing devices ... Not now.

I rather have a slower calculator (same features), but cheaper price.
For intensive computing problems, I use my computer.

[xerxes]

Very impressive and unexpected speedup of the G2. It would be nice, if somebody can run the n-queens problem
on the G2 to see the speed factor for another type of benchmark:

Code:

 EXPORT NQUEENS()
 BEGIN
   R:=8;
   L1:=MAKELIST(0,X,1,R,1);
   S:=0;
   X:=0;
   REPEAT
     X:=X+1;
     L1(X):=R;
     REPEAT
       S:=S+1;
       Y:=X;
       WHILE Y>1 DO
         Y:=Y-1;
         T:=L1(X)-L1(Y);
         IF T==0 OR X-Y==ABS(T) THEN
           Y:=0;
           L1(X):=L1(X)-1;
           WHILE L1(X)==0 DO
             X:=X-1;
             L1(X):=L1(X)-1;
           END;
         END;
       END;
     UNTIL Y==1 END;
   UNTIL X==R END;
   S;
 END;

The speed on the G1 was tested with 0.346 seconds. Maybe the G2 can overtake newRPL on the 50G with 0.106 seconds.
I don't know, if there is any speed advantage declaring the variables as local.

[grsbanks]

Would you not rather run a benchmark that finds ALL the solutions to the n-queens problem rather than stopping at the first one? My Rev. C Prime takes 10.3 seconds to run this:

Code:

sol, nodes, rowChange;
nrow(board,sz,row);

EXPORT nqueen2(sz)
BEGIN
  LOCAL startTime, duration, board, res;
  sol := 0; nodes := 0; rowChange := 0;
  // is sz a real or an integer?
  IF TYPE(sz) = 1 THEN
    sz := B→R(sz);
  ELSE
    IF TYPE(sz) <> 0 THEN
      RETURN {};
    END;
  END;
  // Make sure we're in range
  sz := IP(ABS(sz));
  IF sz=0 OR sz>15 THEN
    RETURN {};
  END;
  // Off we go. Create the board and start working.
  board := MAKELIST(0,X,1,sz);
  startTime := TICKS();
  nrow(board,sz,1);
  duration := (TICKS() - startTime)/1000;
  RETURN {duration,sol,nodes,rowChange};
END;

nrow(board,sz,row)
BEGIN
  LOCAL c, r, safe;
  FOR c FROM 1 TO sz DO
    nodes := nodes + 1;
    safe := 1;
    FOR r FROM 1 TO row-1 DO
      IF board(r)=c OR ABS(board(r)-c)=row-r THEN
        safe := 0;
        BREAK;
      END;
    END;
    IF (safe=1) THEN
      IF row=sz THEN
        sol := sol + 1;
      ELSE
        rowChange := rowChange + 1;
        board(row) := c;
        nrow(board,sz,row+1);
      END;
    END;
  END;
END;

[Anders]

(08-30-2018 02:48 PM)grsbanks Wrote:  Would you not rather run a benchmark that finds ALL the solutions to the n-queens problem rather than stopping at the first one? My Rev. C Prime takes 10.3 seconds to run this:

Code:

sol, nodes, rowChange;
nrow(board,sz,row);

EXPORT nqueen2(sz)
BEGIN
  LOCAL startTime, duration, board, res;
  sol := 0; nodes := 0; rowChange := 0;
  // is sz a real or an integer?
  IF TYPE(sz) = 1 THEN
    sz := B→R(sz);
  ELSE
    IF TYPE(sz) <> 0 THEN
      RETURN {};
    END;
  END;
  // Make sure we're in range
  sz := IP(ABS(sz));
  IF sz=0 OR sz>15 THEN
    RETURN {};
  END;
  // Off we go. Create the board and start working.
  board := MAKELIST(0,X,1,sz);
  startTime := TICKS();
  nrow(board,sz,1);
  duration := (TICKS() - startTime)/1000;
  RETURN {duration,sol,nodes,rowChange};
END;

nrow(board,sz,row)
BEGIN
  LOCAL c, r, safe;
  FOR c FROM 1 TO sz DO
    nodes := nodes + 1;
    safe := 1;
    FOR r FROM 1 TO row-1 DO
      IF board(r)=c OR ABS(board(r)-c)=row-r THEN
        safe := 0;
        BREAK;
      END;
    END;
    IF (safe=1) THEN
      IF row=sz THEN
        sol := sol + 1;
      ELSE
        rowChange := rowChange + 1;
        board(row) := c;
        nrow(board,sz,row+1);
      END;
    END;
  END;
END;

ok ran it a few times with board size = 8

Got 3.124_s with a range of +/-0.2_s
That would give a about a 3x speed up. Pretty good

   

[grsbanks]

Sorry, I should have made it clear that the 10.3 seconds was running it with a board size of 8 and getting all 92 solutions.

[xerxes]

(08-30-2018 02:48 PM)grsbanks Wrote:  Would you not rather run a benchmark that finds ALL the solutions to the n-queens problem rather than stopping at the first one?

Using the first solution only, was more practical in this case, due to the very slow speed of older calculators.
The MK-52 for example needs 5 hours for the first solution. For very fast calculators or assembly language tests,
I've used an additional outer loop for more accurate timing. It's a compromise to be able to compare very slow and
very fast results for the same algorithm, although I'm not really satisfied with the test code of some devices.
It's not easy to test so many calculators and languages absolutely accurate.

[Anders]

(08-30-2018 02:43 PM)xerxes Wrote:  Very impressive and unexpected speedup of the G2. It would be nice, if somebody can run the n-queens problem
on the G2 to see the speed factor for another type of benchmark:

Code:

 EXPORT NQUEENS()
 BEGIN
   R:=8;
   L1:=MAKELIST(0,X,1,R,1);
   S:=0;
   X:=0;
   REPEAT
     X:=X+1;
     L1(X):=R;
     REPEAT
       S:=S+1;
       Y:=X;
       WHILE Y>1 DO
         Y:=Y-1;
         T:=L1(X)-L1(Y);
         IF T==0 OR X-Y==ABS(T) THEN
           Y:=0;
           L1(X):=L1(X)-1;
           WHILE L1(X)==0 DO
             X:=X-1;
             L1(X):=L1(X)-1;
           END;
         END;
       END;
     UNTIL Y==1 END;
   UNTIL X==R END;
   S;
 END;

The speed on the G1 was tested with 0.346 seconds. Maybe the G2 can overtake newRPL on the 50G with 0.106 seconds.
I don't know, if there is any speed advantage declaring the variables as local.

ok, I ran it a few times and put TICKS() just before first REPEAT loop and after the last UNIL to measure the time it took.

I got 0.158_s +/- -.04_s (not 1.158_s)

[xerxes]

Thank you for testing, but are you sure with 1.158 sec? The G1 needs 0.346 sec only.