Actual speed (CPU clock frequency) of 34S?

[Waon Shinyoe]

We know that 34s(or 30b) has the ARM7 core which clocks 32kHZ to 40MHz . But I wonder how fast it really is, since it's limited by the power supply? And according to my personal performance test on integration, 34s is apparently slower than a CASIO fx-991ES, which has an OKI ML610901 core. The built-in RC oscillation frequency is 500kHz. Maybe the comparison is unfair because of the routines/algorithm differences.

Information of the CASIO scientific calculator is from http://www.oki.com/jp/Home/JIS/Books/KEN...03_R06.pdf

[jebem]

Hi,
Interesting comparison...
Can You post the integration test you have run, and the average times to complete on the Casio FX-991ES and on the WP-34S?

[Steve Simpkin]

I suspect it is the integration algorithms that account for the time difference rather than clock speed. For raw performance, the WP34s is much faster.

For example, using the Summation function to perform a loops-of-addition test, the Casio fx-115ES Plus (similar to the fx-991ES) reaches a count of 8,530 in 60 seconds. The WP 34s reaches a count of 272,538 in the same time.
http://www.wiki4hp.com/doku.php?id=benchmarks:addloop

[Marcus von Cube]

(04-15-2014 02:27 PM)Waon Shinyoe Wrote:  We know that 34s(or 30b) has the ARM7 core which clocks 32kHZ to 40MHz . But I wonder how fast it really is, since it's limited by the power supply?

You've answered the question yourself. When I first started the hardware port of WP 34S I just left power consumption out completely: No problem because my prototype device uses a wall wart as its power supply.

Things are completely different with two coin cells as the sole source of power. The software is constantly switching between the various oscillators and speeds, depending on the task the calculator is performing and the time already spent on it. When you look at a steady display without a program running, the processor is essentially switched off. A wake up call (through the keyboard) speeds it up slightly, just enough not to appear too sluggish in accepting key strokes. When a function is executed, speed is raised to about 20 MHz for a fraction of a second. If things are taking longer, speed is increased to the full 37MHz until the task is completed or the voltage drops too much which causes a speed decrease to some 20 MHz again.

As you can see, knowing the maximum specs of the chip doesn't give you enough information to judge the actual time a calculation will take.

The code itself isn't optimized for speed either. Many of the advanced functions are written as key code sequences in order to save valuable flash space. Adding the 34 digit accuracy of most functions to the mix may explain why some other calculator with a less powerful CPU can outperform the WP 34S in certain benchmark cases.

Edited to correct the actual speeds in use.

[Tugdual]

(04-15-2014 08:31 PM)Marcus von Cube Wrote:  

(04-15-2014 02:27 PM)Waon Shinyoe Wrote:  We know that 34s(or 30b) has the ARM7 core which clocks 32kHZ to 40MHz . But I wonder how fast it really is, since it's limited by the power supply?

You've answered the question yourself. When I first started the hardware port of WP 34S I just left power consumption out completely: No problem because my prototype device uses a wall wart as its power supply.

Things are completely different with two coin cells as the sole source of power. The software is constantly switching between the various oscillators and speeds, depending on the task the calculator is performing and the time already spent on it. When you look at a steady display without a program running, the processor is essentially switched off. A wake up call (through the keyboard) speeds it up slightly, just enough not to appear too sluggish in accepting key strokes. When a function is executed, speed is raised to about 10 MHz for a fraction of a second. If things are taking longer, speed is increased to the full 32mHz until the task is completed or the voltage drops too much which causes a speed decrease to some 10 MHz again.

As you can see, knowing the maximum specs of the chip doesn't give you enough information about the actual time a calculation will take.

The code itself isn't optimized for speed either. Many of the advanced functions are written as key code sequences in order to save valuable flash space. Adding the 34 digit accuracy of most functions to the mix may explain why some other calculator with a less powerful CPU can outperform the WP 34S in certain benchmark cases.

Hi Marcus, quick question. What governs the frequency switch? Is that hardware or software?

[Marcus von Cube]

(04-15-2014 08:38 PM)Tugdual Wrote:  Hi Marcus, quick question. What governs the frequency switch? Is that hardware or software?

The device has several oscillators which can be programmed to almost any frequency between a few Hz and 40MHz (or even more which makes operation unreliable). The software is responsible for setting all this up.

BTW, please do not quote the complete (and meanwhile outdated because I've edited the text) post but only relevant portions of the text.

[Paul Dale]

(04-15-2014 07:43 PM)Steve Simpkin Wrote:  For example, using the Summation function to perform a loops-of-addition test, ...

This test would show the 34S in a poor light. The algorithm behind the summation function does quite some extra work in order to squeeze some extra digits out of the result.

Write your own naïve summation routine and you'll find that it is faster than the built in one!


- Pauli

[Steve Simpkin]

(04-15-2014 09:13 PM)Paul Dale Wrote:  This test would show the 34S in a poor light. The algorithm behind the summation function does quite some extra work in order to squeeze some extra digits out of the result.

Write your own native summation routine and you'll find that it is faster than the built in one!

Understood. More precision = more work = more processing time.

And as anyone who owns one knows...
Nothing about the 34s would rank as showing it "in a poor light"

[nsg]

(04-15-2014 08:46 PM)Marcus von Cube Wrote:  The device has several oscillators which can be programmed to almost any frequency between a few Hz and 40MHz

Can it be programmed to run at full speed when powered from Harald's usb board?
Available voltage could be an indicator of that.

Also, when run at full speed (when powered from a pigtail on your test bench), is there a difference in calculating speed of some slower functions, like inverse distributions?

[walter b]

(04-16-2014 12:31 AM)nsg Wrote:  Can it be programmed to run at full speed when powered from Harald's usb board?
Available voltage could be an indicator of that.

I doubt the sensibility. Even Harald's board will supply a voltage within the 'legal' range, next to the voltage of a fresh battery. And do you want your fresh batteries brought down by a processor running full speed without need until the voltage drops below a certain limit? Certainly not. (I'm not talking about the settings FAST and SLOW here.)

d:-/

[Marcus von Cube]

(04-16-2014 12:31 AM)nsg Wrote:  Can it be programmed to run at full speed when powered from Harald's usb board?
Available voltage could be an indicator of that.

Be assured that as soon as you start some serious computation or are running a user program, it will run at full speed. The tiny delay between half speed and full speed shouldn't be noticeable at all.

(04-16-2014 12:31 AM)nsg Wrote:  Also, when run at full speed (when powered from a pigtail on your test bench), is there a difference in calculating speed of some slower functions, like inverse distributions?

I don't think so. It's the iterative nature of the algorithms what makes them slow. If speed is an issue you can run the iOS emulator on your iPhone or iPad for those big tasks.

[Steve Simpkin]

(04-15-2014 09:13 PM)Paul Dale Wrote:  

(04-15-2014 07:43 PM)Steve Simpkin Wrote:  For example, using the Summation function to perform a loops-of-addition test, the Casio fx-115ES Plus (similar to the fx-991ES) reaches a count of 8,350 in 60 seconds. The WP 34s reaches a count of 272,538 in the same time.
http://www.wiki4hp.com/doku.php?id=benchmarks:addloop

This test would show the 34S in a poor light. The algorithm behind the summation function does quite some extra work in order to squeeze some extra digits out of the result.
Write your own native summation routine and you'll find that it is faster than the built in one!
- Pauli

Just for reference, the summation function on the WP 34S (V2.2 1990) reaches a count of 36,900 is 60 seconds with summation of the following program:
001 LBL 'O'
002 1
003 RTN

This is quite a bit faster than the Casio fx-115ES Plus which reached a count of 8,350 in 60 seconds or the TI-36X Pro which only reached a count of 5,130 in the same time.

[nsg]

(04-16-2014 09:12 AM)Marcus von Cube Wrote:  Be assured that as soon as you start some serious computation or are running a user program, it will run at full speed. The tiny delay between half speed and full speed shouldn't be noticeable at all.

Thanks for clarification. For some reason I (incorrectly) interpreted previous discussion as if it started at full speed and then slowed down if it does not get result immediately.