Memory Cells

We’ve spent four chapters now building machinery in order to construct memory blocks – large arrays of our Snap* machine – capable of being selectively written-to and addressed-from. This “selectively” criterion generally goes by the name addressable, which means that each Snap* machine has an “address”, which we know beforehand and can use to route signals to and from it.

As usual, this notion of addressability has deeper and more interesting implications in the wider theoretical sense, but we are not interested in that today. Abstractions are only valuable in the fact that they allow us to abstract over something, and we do not yet have an example over which we can abstract.

But we do, in fact, have something else which we’d like to abstract over, and that is this memory block we’ve so desired over these four long chapters. A good way to go about designing these things is to completely ignore the internals, and reason about how you’d like it to work if someone else had built it and didn’t provide any kind of documentation. One lens for abstraction is in terms of “usability” – can we provide a simple mental model for describing how this thing works, and then manage to never violate that model? If so, we have provided a useful abstraction. If not, the thing we have built deserves to be thrown into the trash and done again.

And so the question is asked: how would we like this machine to operate? Another phrasing of that question is “what are its semantics?” which is equivalently “what is our mental model of this thing?”

Let’s start with what we know so far. We’d like to be able to write to a Snap*, which suggests we’ll need to get input to the Snap*, which means at the very least we’ll need signals S:1 and V:n, albeit possibly manipulated somehow to make the interface “cleaner”. But we also need to fit several of these Snap*s into one logical machine, and we’ll need a way to address them. Fortunately we were forward thinking and have already built Demux, which, if you think about it, provides exactly the mechanism necessary for addressing machines.

So what does Demux take as input? A payload, A:n (although this :n need not be the same as the multiwire V:n on Snap* – remember, polymorphic wires with the same letter need only be consistent within a single diagram. When we merge these things together, we’ll rename one of those :ns to a different letter so that we don’t confuse ourselves later). Demux also takes an input D:d, which is responsible for describing a “destination” output multiwire to move the payload to.

Presumably this means we’d like a Demux machine in front of a bunch of Snap* machines, and that the payload to Demux should be the value V we’re sending to our Snap*s. As such, we’ll need to line up the output multiwires from Demux with the value multiwire to Snap*, and we realize that yes, indeed the :n annotation on the A is in fact the same as the :n annotation on the V. Remember – a wire has the same value all along it, and multiwires are no different. In order to connect an output multiwire to an input multiwire, they must have the same annotation, or they wouldn’t physically (or logically) “fit” together.

Recall that the other thing we want from our memory block is to be able to selectively read from an addressable Snap* machine. If you remember, Snap*’s output is always the last thing it snapshotted, and so if we have several of these things together, we’ll have nothing but a cacophony of values, only one of which we actually want. But only sometimes.

Why only sometimes? It seems more parsimonious that if we’re not always writing to a Snap* (because it requires you to raise the Snapshot wire), then we probably don’t want to always be reading from it either. This way, consumers of the data can tightly control exactly what they’re getting – they can raise a wire, receive the data they wanted, and then go on their jolly way. Making the data read on demand means that it’s tricky for someone else listening on the same output line to receive information meant for someone else.

It’s like the difference between a telephone and a walkie-talkie; a telephone notifies someone that you’d like some data, and the connection only stays open as long as you require. A walkie-talkie on the other hand is always listening, and having multiple people share the same channel results in “chatter” – you’ll be distracted any time anyone is talking, even if it’s not towards you.

With this semantic constraint, it’s now obvious that we need a way to tell our memory block machine that we’re listening, and it should send data to us. Because we’re listening for data, don’t want to put any data on the payload line, and so any writes that happen simultaneously with our read will accidentally write over the data we’re trying to read. As a result, we probably don’t want to be able to read and write at the same time; the only tool we know for dealing with that is to make it the same line, where a 0 value represents one of the possibilities, and 1 represents the other. Unfortunately, such a decision means that we need another wire to tell the machine to actually do something. A “go for it!” kind of signal, if you will.

On a cursory glance, we seem to have described exactly what we want out of our memory block, so let’s start synthesizing all of the things we decided. Our machine needs to take some inputs:

  • a payload A:n
  • a destination D:d
  • a read-write option RW:1, where 0 describes a “read” and a 1 describes a “write”
  • and an active “go” wire G:1

whose semantics are as follows:

RW G Output Side Effect
_ 0 Nothing Nothing
0 1 Snap* at D Nothing
1 1 Nothing Set Snap* at D to A

What tools to we have available to us to construct such a beast? Well we know that anything anded against a 0 is also 0, so anding against the G wire allows us to turn the entire machine on or off. In particular, the things we’d want to and against would be the Snapshot wire in the Snap*, and the output of the Snap* (so that we don’t read any data if G is low).

Demux gives us the capability to address a particular machine based on some input D:d, which allows us to move not only the payload A to the correct Snap*, but also the snapshot value S. The last trick is to “un-demux” the results from our many Snap*s onto a single output multiwire. This, of course, we have already looked at in the form of Mux2. We won’t be covering the construction of the generalization of Mux2 to a polymorphic Mux, but it follows exactly the same procedure we did to construct Demux from Demux2.

With all of the pieces and reasoning in place, let’s go about building this sucker.

Egads, what a monster. But remember, don’t panic! This Mem thing looks tricky at first, but it’s still exactly the thing we described earlier. As an exercise, convince yourself that Mem has the same semantics as the function table we drew earlier.

We’ve spent a long time (five chapters now!) building this Mem machine; it feels like time for a break. In the next chapter, we’ll upgrade our understanding of what an annotation is, and find that they are much more powerful than we’d ever imagined. We’ll then start looking at how we can build a new way of describing these machine diagrams, since they’re getting surprisingly large and difficult (for you to read, and for me to draw!).


  1. Go through the exercise of generalizing Mux2 to a polymorphic Mux.
  2. Work out the semantic function table of Mem, and ensure that it agrees with the one given previously.