I’m not really one to brag publicly about expensive toys, but a few weeks ago I managed to get one that’s really something special. It is a Curta Type II, a mechanical digital1 calculator manufactured in Liechtenstein between the 50s and 70s, before solid-state calculators killed them and the likes of slide-rules.
I have wanted one since I was a kid, and I managed to win an eBay auction for one.
It’s a funny looking device, somewhere between a peppermill and a scifi grenade. Mine has serial number 544065, for those keeping score, and comes in a cute little bakelite pod (which has left hand thread?!).
I wanna talk about this thing because unlike something like a slide rule, it shares many features with modern computers. It has operations, flags, and registers. Its core primitive is an adder, but many other operations can be built on top of it: it is very much a platform for complex calculations.
I’m the sort of person who read Hacker’s Delight for fun, so I really like simple numerical algorithms. This article is a survey of the operation of a Curta calculator and algorithms you can implement on it, from the perspective of a professional assembly programmer.
Many of the algorithms I’m going to describe here exist online, but I’ve found them to be a bit difficult to wrap my head around, so this article is also intended as a reference card for myself.
Let’s dive in!
There are two Curta models, Type I and Type II, which primarily differ in the sizes of their registers. I have a Type II, so I will focus on the layout of that one.
The Curta is not a stored program computer like the one you’re reading this article on. An operator needs to manually execute operations. It is as if we had taken a CPU and pared it down to two of its most basic components: a register file and an arithmetic logic unit (ALU).
The Curta’s register file consists of three digital registers, each of which contains a decimal integer (i.e., each digit is from
9, rather than
1 like on a binary computer):
sr, the setting register, is located on the side of the device. The value in
srcan be set manually by the operator using a set of knobs on the side of the device. The machine will never write to it, only read from it. It has 11 digits.
rr, the results register, is located at the top of the device along the black part of the dial. It is readable and writable by the machine, but not directly modifiable by the operator. It has 15 digits.
cr, the counting register, is located next to
rralong the silver part of the dial. Like
rr, it is only machine-modifiable. It has 8 digits.
There are also two settings on the device that aren’t really registers, but, since they are changed as part of operation, they are a lot like the control registers of a modern computer.
The carriage (there isn’t an abbreviation for this one, so I’ll call it
ca) is the upper knurled ring on the machine. It can be set to a value from
72. To set it, the operator lifts the ring up (against spring tension), twists it, and lets it spring back into the detent for the chosen value. This is a one-hand motion.
There is a small triangle in the middle of the top of the device that points at which of the digits in
cr will get incremented.
rl, the reversing lever, is a small switch near the back of the device that can be in the up or down position. This is like a flag register: up is cleared, down is set.
We have all this memory, but the meat of a machine is what it can do. I will provide an instruction set for the Curta to aid in giving rigorous descriptions of operations you can perform with it.
The core operation of the Curta is “add-with-shift-and-increment”. This is a mouthful. At the very top of the machine is the handle, which is analogous to a clock signal pin. Every clockwise turn of this handle executes one of these operations. Internally, this is implemented using a variation on the Leibniz gear, a common feature of mechanical calculators.
This operation is not that complicated, it just does a lot of stuff. It takes the value of
sr, left-shifts it (in decimal) by the value in
ca, and adds it to
rr. Also, it increments
1 shifted by
ca. In other words:
Recall that this is a decimal machine, so
<< is the same as multiplication by a power of 10, not a power of 2.
Addition can overflow, and it wraps around as expected: adding one to
999_999_999_999_999_999 already in
rr will fill it with zeroes.
Pulling the handle up reveals a red ring, indicating the machine is in subtraction mode. This flips the signs of both the
The Curta cannot handle negative numbers, so it will instead display the ten’s complement3 of a negative result. For example, subtracting
0 will produce all-nines.
You can detect when underflow or overflow occurs when the resulting value is unexpectedly larger or smaller than the prior value in
rr, respectively. (This trick is necessary on architectures that lack a carry flags register, like RISC-V.)
rl will reverse the sign of the operation done on
cr during a turn of the handle. In addition mode, it will cause
cr to be subtracted from, while in subtraction mode, it will cause it to be added to. Some complex algorithms make use of this.
Finally, the clearing lever can be used to clear (to zero)
rr, independently. It is a small ring-shaped lever that, while the carriage is raised, can be wiped past digits to clear them. Registers cannot be partially cleared.
Let’s give names to all the instructions the operator needs to follow, so we can write some assembly:
mr, or Machine Ready!, means to clear/zero every register. All Curta instructions use the term “Machine Ready” to indicate the beginning of a calculation session.
pturnis the core addition operation, a “plus turn”.
mturnis its subtraction twin, a “minus turn”.
set <flag>requests the operator set one of
clr <flag>is the opposite of
zero <reg>request a clear of one of
crusing the clearing lever.
add <reg>, <imm>requests manual addition of an immediate to
ca. This is limited by what mental math we can ask of the operator.
copy <reg>, srrequests a copy of the value in
wrnp <reg>, <symbol>indicates we need to write down a value in any register to a handy notepad (hence
pad), marked with
rdnp <reg>, <symbol>asks the operator to
read a value recorded with
if <cond>, <label>asks the operator to check a condition (in terms of
sr) and, if true, proceed to the instruction at the given
label:. Here’s some examples of conditions we’ll use:
rr == 42, i.e.,
rrequals some constant value.
rroverflowed/underflowed due to the most recent
cr == 9, i.e.
cr’s second digit (zero-indexed, not like the physical device!) equals
cr[0..ca] < sr[0..ca], i.e.,
cr, considering only the digits up to the setting of
ca, is less than those same digits in
goto <label>is like
ifwithout a condition.
donemeans we’re done and the result can be read off of
Note that there is a lot of mental math in some of the conditions. Algorithms on the Curta are aimed to minimize what work the operator needs to do to compute a result, but remember that it is only an ALU: all of the control flow logic needs to be provided by the human operator.
None of this is real code, and it is specifically for the benefit of readers.
So, addition and subtraction are easy, because there are hardware instructions for those. There is, however, no direct way to do multiplication or division. Let’s take a look at some of our options.
Given that a Curta is kinda expensive, you can try out an online simulator if you want to follow along. This one is pretty simple and runs in your browser.
The easiest way to do multiplication is by repeated addition;
cr helps us check our work.
Given a value like
8364, we can multiply it by
5 like so:
Here, we input the larger factor into
sr, and then keep turning until
cr contains the other factor. The result is
Of course, this does not work well for complex products, such as squaring
41820. You could sit there and turn the handle forty thousand times if you wanted to, or you might decided that you should get a better hobby, since modern silicon can do this in nanoseconds.
We can speed this up exponentially by making use of the distributive property and the fact that
turn can incorporate multiplication by a power of
Each nice round number here can be achieved in
cr by use of
ca. Our algorithm will look a bit like this:
There are two loops. The inner loop runs as many turns as is necessary to get the next prefix of the factor into
cr, then incrementing
ca to do the next digit, and on and on until
cr contains the entire other factor, at which point we can read off the result.
The actual trace of operations (omitting control flow), and the resulting contents of the registers
sr/rr/mr/ca at each step, looks something like this:
The result can be read off from
1748912400. In the trace, you can see
cr get built up digit by digit, making this operation rather efficient.
We can do even better, if we use subtraction. For example, note that
18 = 20 - 2; we can build up
cr by doing only 4 turns rather than nine, according to this formula. Here’s the general algorithm for
n * m:
Although more complicated, if we execute it step by step, we’ll see we get to our answer in fewer turns:
In exchange for a little overhead, the number of turns drops from 15 to 10. This is the fastest general algorithm, but some techniques from Hacker’s Delight can likely be applied here to make it faster for some products.
As a quick note, computing the cube of a number without taking extra notes is easy, so long as the number is already written down somewhere you can already see it. After computing
n^2 by any of the methods above, we can do
This sequence can be repeated over and over to produce higher powers, and is only limited by the size of
Division is way more interesting, because it can be inexact, and thus produces a remainder in addition to the quotient. There are a few different algorithms, but the simplest one is division by repeated subtraction. Some literature calls this “division by breaking down”.
For small numbers, this is quite simple, such as
21 / 4:
This works by first getting the dividend into
rr and resetting the rest of the machine. Then, with
rl set, we subtract the divisor from
rr until we get overflow, at which point we add to undo the overflow. The quotient will appear in
cr: we set
rl, so each subtraction increments
cr, giving us a count of
mturns executed. The remainder appears in
In this case, we get down to
1 before the next
mturn underflows; the result of that underflow is to
99...97, the ten’s complement of -3. We then undo the last operation by
cr: this is our quotient.
rr is the remainder.
The same tricks from earlier work here, using
ca to make less work, effectively implementing decimal long division of
Let’s execute this on
3141592653 / 137, with an instruction trace as before.
For a quotient this big, you’ll need to work through all eight
cr digits, which is a ton of work. At the end, we get a quotient of
22931333 and reminder
Unfortunately, we can’t as easily “cheat” with subtraction as we did with multiplication, because we don’t know the value that needs to appear in
Computing square roots by approximation is one of the premiere operations on the Curta. There’s a number of approaches. Newton’s method is the classic, but requires a prior approximation, access to lookup tables, or a lot of multiplication.
A slower, but much more mechanical approach is to use Töpler’s method. This consists of observing that the sum of the first
n odd numbers is the square of
n. Thus, we can use an approach similar to that for division, only that we now subtract off consecutive odd numbers. Let’s take the square root of
9 as our result, but that’s pretty awful precision. We can improve precision by multiplying
92 by a large, even power of ten, and then dividing the result by that power of ten’s square root (half the zeroes).
Unfortunately, this runs into the same problem as naive multiplication: we have to turn the handle a lot. Turning this algorithm into something that can be done exponentially faster is a bit fussier.
One approach (which I found on
It is based on the so-called “digit-by-digit” algorithm, dating back to at least the time of Napier. Wikipedia provides a good explanation of why this method works. However, I have not been able to write down a proof that this specific version works, since it incorporates borrowing to compute intermediate terms with successive odd numbers in a fairly subtle way. I would really appreciate a proof, if anyone knows of one!
The algorithm is thus, for a radicand
Let’s compute some digits of
sqrt(2). Here’s the instruction trace.
Over time, the digits
14121356 will appear in
cr. This is the square root (although we do need to place the decimal point; the number of digits before it will be half of what we started with, rounded up).
There’s a quite a few other algorithms out there, but most of them boil down to clever use of lookup tables and combinations of the above techniques. For example, the so-called “rule of 3” is simply performing a multiplication to get a product into
rr, and then using it as the dividend to produce a quotient of the form
a * b / c in
I hope that these simple numeric algorithms, presented in a style resembling assembly, helps illustrate that programming at such a low level is not hard, but merely requires learning a different bag of tricks. ◼
Although this seems like an oxymoron, it is accurate! The Curta contains no electrical or electronic components, and its registers contain discrete symbols, not continuous values. It is not an analog computer! ↩
The Curta is a one-indexed machine, insofar as the values engraved on
8. However, as we all know, zero-indexing is far more convenient. Any place where I say “set
n”, I mean the
n + 1th detent.
Doing this avoids a lot of otherwise unnecessary
-1s in the prose. ↩
The ten’s complement of a number
xis analogous to the two’s complement (i.e., the value of
-xwhen viewed as an unsigned integer on a binary machine). It is equal to
MAX_VALUE - x + 1, where
MAX_VALUEis the largest value that
xcould be. For example, this is
999_999_999_999_999_999(fifteen nines) for