Writing about timing and timing constraints: Kind-of behind the scenes

This post was written by eli on March 16, 2023
Posted Under: FPGA

Oh my goodness

I’ll say this from the start: This is me babbling. You have been warned. Don’t expect anything coherent in this post. Not that I necessarily keep things very tidy in my other posts, but this one is clearly me typing at full speed, not trying to be organized or anything.

I’ll buy you a pizza if you figure out the real motivation behind this post. Well, actually I won’t. At most, I’ll answer with an emoji with a pizza. But you’re welcome to try your luck anyhow. If you bother to read this, that is. Which I’m not sure is the best way to spend your time.

So these are my thoughts after publishing a series of pages about FPGA timing and timing constraints (which begins with this page). It’s not only a relatively long series of pages, but it also involves the translation of the pages into Chinese, Japanese and Korean. I know none of these languages, so I used Google Translate. I’ve already written a post about the challenges and solutions with automatic translation of technical documents.

The translations are a huge experiment. It’s not clear at all if people who speak these languages will find them appealing. There’s clearly a readability problem with an automatic translation, no matter how hard I work to make the text digestible for an AI machine.

People who speak English and one of these languages have expressed little enthusiasm when looking at the results (to say the least). But then it boils down to the choices of people who only have that East Asian language. They will surely have a laugh or curse when reading encountering the hick-ups of automatic translations, but do they have a better alternative? Time will tell. Or more precisely, the logs will tell.

I’ve never written a behind-the-scenes post about other things that I’ve written. Maybe I do this because it’s the first time in ages that I can write English exactly the way I think, and not keeping restricting myself to write in a way that works well with the translation machinery. One of the main limitations with automatic translation to CJK languages is that the sentences must be short and concise. The ordering of words is very different in those languages, so if I use a word like “it” or “which” that refers back to something mentioned earlier in the sentence, the translation machinery is forced to decide what it referred to, and name it explicitly in the translation. And quite often, that goes wrong. When the translation is to another European language, or even Hebrew, it’s possible to retain the ambiguity in the translation, and leave it to the intelligence of the reader.

So if this post feels like a huge ramble, yes, that’s exactly what it is. Me feeling free at the keyboard. The liberation from thinking how each and every word may be mistranslated. I mean, even a work like “generation” in the sense of creating something can be translated into like generation gap. Or if I say “that’s a special case”, will that become something about a bag? I’m really tired of translating and back-translating and cutting down Chinese into single characters in order to ensure that when I write “segment” I don’t get something related to marketing.

Let’s talk about timing

I’ve been thinking about writing a guide to FPGA timing and timing constraints for quite a while. It’s one of the most important topics, if not the most important topic in the field, and somehow it seems like the explanations about this field are either very theoretic, or they’re so specific that it’s difficult to understand the whole point. It was clear to me that I needed to begin with the basics of the theory (and translate it, of course). I was however quite surprised myself how little theory there was to convey. The entire thing is about setup time, hold time and clock-to-output. That’s it. That’s the whole story.

It was quite clear where to take it from there: To discuss the most commonly used timing constraint, that is to say the period constraint. And make that international, once again.

But how do you talk about this simple constraint? I mean, the normal explanation is just “this is what you should write”. How do you explain what it really means?

So the approach that I went for throughout the series of pages was to explain it through the timing reports. It’s a bit of a tedious way, but the idea was that the correct way to understand timing constraints, and in fact one of the most essential capabilities for working with timing, is to read the timing reports. Hence I went for explaining the timing report.

Another alternative could have been to jump into explaining the meaning of the timing constraint as a clock object. This is a more technical approach, and it’s necessary to go through it at some stage. But I deferred that to later.

To PLL or not to PLL

And by the way, that’s a title that I really wanted somewhere in those pages, but it was clearly going to be lost in translation. So all you guys from East Asia out there, if you won’t read the translated pages, just know I sacrificed proper English in favor of you getting something decent to read. Kindly appreciate it.

Anyhow, it turned out that I couldn’t finish the entire topic of the period constraint in one page, so there’s a second page on the same topic, this time for the sake of explaining how a PLL changes the picture. And that brings up the whole story of clock domain crossing between related clocks. Plus I also filled in the part about hold time.

Hold time is a neglected topic regarding timing constraints. Probably because the timing constraints rarely fail on that, and still: The tools deliberately add delays to the paths in order to meet hold timing, and by doing so, they put an obstacle before themselves in the quest to attain the setup timing. It’s quite amazing to see how a simple routed net suddenly gets a delay of several ns, just because the hold timing required so. In particular when crossing clock domains between related, but unaligned clocks.

And that’s another expression I couldn’t use on those pages: Crossing clock domains. It always says “clock domain crossing”, and it’s protected against translation. Otherwise, I get crucifixion in the translation. And the protection against translation works only on nouns, not verbs. Did I just divert from the topic? Maybe I need a closure?

Timing closure

When is the right time to talk about it? Timing closure is often the reason we even start think about timing (ghosts in the FPGA is another common reason). So it’s definitely an important topic, at least to the impatient reader. But shouldn’t I put it in the end? After all, it’s the ultimate surgical operation that is done once you understand it all.

So no. I went with the people on this one. It’s the ultimate topic, yes, but I also need to realize that some people will read one page or two and then give up. So better get to the important thing ASAP. Actually, I also gave away a checklist kind of page (with trans-la-tions). Maybe that’s the biggest contradiction I made in those pages. I said all the time that timing is like a divine science, and don’t be lazy about it. Be deep and profound, learn it as an art, and then I come up with a checklist. Not only that, I repeat that crime at the end of the page series. And I do that in three additional languages, too.

To my defense (or is it?), the repeated crime had a practical and silly reason. I had that page already written. More or less. So it was a bit of low-hanging fruit. And my rule (in this blog and otherwise) has always been not to throw away food. If you have some random pieces of cheese in your fridge, make a quiche. If there is mold on the cheese, you just make a different kind of quiche.

So that last page in the series is not 100% coherent with those preceding it, but it kinda summarizes the point. Besides, I already mentioned that the whole series of pages is one big experiment…?


That’s a word in Jiddish, literally taken from the Hebrew word תכלית. Actually, it’s the pronunciation of that word in the mouth of European Jews. The meaning of the word is “purpose”, the reason for doing something, but more in the direction of the intention, what result you want to achieve etc.

So after talking about timing closure for two pages, I finally got to the pudding: Actually talk about how to write timing constraint. But not so fast! There’s still one tedious phase to go through, namely learn how to select the logic that we want to be specific about.

This whole thing with Tcl commands, which are used in SDC format (that is, by Vivado, Quartus and several other tools) gives a certain level of compatibility. But as it turns out, there’s Synopsys’ set of capabilities, which seems to have been adopted in full by Vivado, and then there’s Quartus’ take on the same topic, which is same-same but completely different. So the simple commands work the same, but when it comes to more complicated stuff, Quartus diverges, and not for the better. In particular, Quartus lacks the “-filter” option that allows more complicated expressions for defining whether to include a logic element (actually, the object representing it) in a timing constraint. So one is limited to playing with wildcards, and that, well, sucks.

So in short, this was the point I had to admit that Vivado is superior, with a margin. I also began to suspect that Vivado is Synopsys in disguise, but that’s just a wild speculation.

It’s only after that, that I could begin to really talk about commands like set_max_delay, set_min_delay and set_false_path. That’s what I call tachles. I finally got to the point, after rambling tons.

I think that the most important point about set_false_path is that it has maximal priority. One would usually consider priority to be an advanced topic.

Ah, and here comes another messy, out-of-nowhere comment about this series of pages. The idea wasn’t to write a full and comprehensive guide that includes everything that there is possible to know. That’s what the official docs are for. The idea was to cover the things that I personally use and feel that I must be aware of, and hence the convey the minimal toolbox that one needs to carry.

Return from subroutine: I didn’t want to cover an advanced topic like priorities between timing constraints, because who really needs that? It’s quite clear that those specific timing constraints are stronger than the period constraint, and except for that, nobody should mess up with more complicated combinations of contradicting constraints.

But set_false_path is scary: It wins everything else, and it disables the timing check of the relevant paths silently. It’s literally a killer. So make a small mistake with that constraint, and your design is cooked. So I spent quite a few words on warnings about that.

Eight o’clock rock

When I first started working with FPGAs, my mentors were from the ASIC industry. For some reason, which I’m still to find out, ASIC guys like to everything with one clock, if possible. In the FPGA world, it’s more a like clock being given away like they were air. So the whole issue of timing constraints needs to be addressed in the context of crossing clock domains.

And here’s another word that I avoided all the time: Addressed. Because who knows what that will be translated into, when it comes to Chinese, Japanese and Korean.

The alternative to a lot of clocks, is multi-cycle paths, of course: This is what the ASIC guys do in order to stick to one clock for the entire design. And in order to take advantage of the relatively relaxed timing of logic that depends on a clock enable, there are timing constraints. Which is a risky business. I won’t repeat the explanation here, but let’s say that it’s generally not recommended.

Who you wanna talk to?

One of the things that I actually planned and carried out as planned (well, not really as planned) was to put the discussion about I/O constraints last. This went against the principle of important things first, but since I/O constraints are technically just a private case of constraints on internal paths (that is, paths that start and end inside the FPGA), it made a lot of sense to put this part last.

Now read this last sentence again, and realize that there’s nothing even close to that length in the pages I wrote on timing constraint. No chance a sentence like that would have been translated correctly.

So where were I? Oh yes, timing constraint for I/O. But soon enough I realized that I can’t talk about timing constraints for I/O before discussing system synchronous vs. source synchronous vs. asynchronous I/O. But it’s not that simple. One can’t just throw these terms in the air and hope for good. So all of the sudden I found myself with a small בלת”ם (Read “baltam”, meaning “unplanned” in Hebrew), which wasn’t so small.

That’s because source-synchronous I/O means both inputs and outputs, and these are implemented in different ways. So there was no choice (or was there?) to sit down and explain how to do that. How can you discuss the possibilities for timing constraints of a source-synchronous input without explaining how the I/O is implemented. There are quite a few possibilities, after all.

Source-synchronous outputs are actually a much simpler story, because it just means to generate all signals with registers.

On top of all this mess, there’s another topic which is important both for the timing of I/O and for ensuring that the design behaves in a consistent way: IOB registers.

But oops. IOB is a Xilinx word. Oddly enough, it doesn’t seem like there’s a standard term for the idea of having a flip-flop close to the I/O pin. It’s such a common thing, and yet there’s no common term. I think. So I used the term I’m used to, IOB. It’s quite apparent that I’m a Xilinx guy. I’ve worked a bit with Altera (should I call them AMD and Intel FPGA by now? It takes time to get used to that). I nevertheless tried to be vendor-neutral in those pages. They were supposed to be about all FPGAs, not only the two that were leading back in 202x, just before being acquired by two giants, leaving an unknown future ahead.

But that was quite impossible, because timing constraints are usually written in SDC, and SDC was invented by Synopsys, and Vivado are the closest that I know of to the full and accurate interpretation of the constraints, as well as the Tcl environment. That was yet another long sentence that I would never write for translation. But I actually expect that all tools will end up supporting Synopsys’ full vocabulary, so maybe a few years from now, these posts will really be vendor-neutral, and the only thing people won’t understand is what Vivado and Xilinx is, because some other FPGA vendor that we’ve never heard of has taken control of the market, while “AMD FPGA” put all they had on datacenter acceleration and ended up closing the FPGA department because, well, because.

Did I diverge again? Definitely. All I wanted to say is that I also wrote a page about IOB registers, and needless to say, that page is available in three additional languages, just like the rest.


I really wonder if anyone will read through this post from top to bottom. If you haven’t, and just scrolled down to the bottom to see if this goes anywhere, I can’t blame you. Truth to be said, this post went nowhere, and neither was it intended to do so.

Any guesses about the reason I wrote this anyhow. I still promise to send a pizza emoji in response to correct answers. Even though all answers are incorrect by the nature of this foolishness.

Reader Comments

I did read the post from top to bottom.
One thing I absolutely dislike during timing closure is using aggregated measures such as TNS (Total Negative Slack). If you plot a histogram of a typical TNS – slack bin on the x-axis, and count of paths on the y-axis – it’d look like a long-tail distribution. All those paths with smaller slacks are totally unrelated to the ones with WNS (Worst Negative Slack). WNS paths are actually those that
prevent Place & Route to close timing, and the rest of the paths participating in TNS are just a distraction, making it more difficult to correctly interpret the results.

Written By Evgeni Stavinov on March 17th, 2023 @ 05:29

Interestingly enough, I didn’t even relate to that concept in the said series of pages. The reason is simple: I never paid attention to TNS, exactly because of the reasons you mentioned. :)

Written By eli on March 17th, 2023 @ 05:35

Add a Comment

required, use real name
required, will not be published
optional, your blog address