Comment by cjk2 | Hacker News Reader

Comment by cjk2 | Hacker News Reader

Parent Full thread

The thing is though, stuff rarely works like that when it comes to reality. It shows more of a lack of understanding of electronics than is required to build something end to end.

There is a lot of futzing that goes on after the idealisms are found to be less ideal and contracts between components are never quite what they seem (consider logic interfacing/impedance etc). And when it comes to analogue things, it gets pretty wild. At that point you have to break the abstraction and know exactly what you are doing. This is a typical failure mode of the Arduino ecosystem; the abstraction only goes so far and then you're shit out of luck.

On top of that concerns like decoupling and power routing/distribution are never straightforward. So you have an ESP8266 and you connect it to a sensor and an actuator. When your sensor fires, the actuator causes voltage sag, which causes the sensor to flip back to the existing state. etc etc.

Another good example is the ubiquitous LM358. The designer needs to know about the electrical interface and constraints of both the input and the output. The input does not have compliance across the full supply line and the output can only push and pull current on certain load curves leading to weird things like distortion. The abstraction of the black box of the IC is broken then.

It's a very complicated subject which is very hard to abstract away. For ref I was an RF design engineer for several years and designed PCBs before that. I moved into software because it paid more and was much much easier (hint)

I think a major issue with how datasheets are translated into existing EDA tools is none of the major constraints are transferred, and you usually need to manually place de-coupling capacitors or other major important components which is pretty tedious. Some tools are a bit better at this, but they might not have parameterized variants or other things that limit how well they can do it.

I 100% agree that at some point somebody graduates from Arduino and has to understand that traces are antennas. But the gap is very wide, if every EDA tool gave great static analysis errors, people could incrementally learn many EE concepts that are best learned in practice. A classic error is connecting an LED directly to digital output w/o a resistor for example, why doesn't that throw in every EDA tool? Also: assembly is fairly cheap now- part of the reason people use Arduino is because directly placing an MCU would make a circuit more difficult to assemble, and Arduino stuff can be stacked etc. I think we should have a lot more people's first circuits go right to PCB/assembly, but to do that there has to be good static analysis.

For analog designs, SPICE output and static analysis is a big goal. For complex routing, I do think people will often export into other EDA tools or need a GUI router (maybe this is important enough that simple GUI routing should be built in- I have some ideas on how to do this while keeping it code) Unless there is a really really good autolayout engine that can take user-specified hints the reality is real boards will need careful layout and re-layout.

> Another good example is the ubiquitous LM358. The designer needs to know about the electrical interface and constraints of both the input and the output. The input does not have compliance across the full supply line and the output can only push and pull current on certain load curves leading to weird things like distortion. The abstraction of the black box of the IC is broken then

I think this can be represented pretty well in code, one of the planned outputs is SPICE models and 3d. Sometimes it's not practical for an IC to have a SPICE model, but I think you could emulate a group of components with a black-box SPICE model if it's set up properly. One of the things I'm looking forward to is "complex linting" where a dev could get an error if they e.g. don't emulate a power path via SPICE.

If you're open to it (likewise to frognumber above!!!) I'd really love to talk through some of the solutions/alternatives I have planned and see if they land. My email is seve at tscircuit dot com!!

"....why doesn't that throw in every EDA tool?"

This would require repetitive SPICE simulations, or basic rule checking at the very least. Nobody does full SPICE simulations at the board level however basic input/output port checking (usually in the ERC check) does get performed. Even with RF designs, you carve out the piece you need to examine or design and simulate that. For the chips I've worked on, the full chip would get a SPICE simulation that would take days/weeks but this was for more R+D oriented mixed signal designs. I guess what I'm saying is the simulation of a circuit is best performed as a deliberate, iterative step in the circuit design process.

When it comes to layout however, you do get hints from the DRC checking tool (Design Rule Constraints) that will tell you if your trace is drawn incorrectly based on the DRC constraints and nowadays sometimes from an EM simulation that can be run in the background.

Completely automated design especially for analog will most likely never be a thing for the other reasons you list. However, I already can use "known good" circuits and modularize them for reuse which does speed things up. This is critical in the ASIC world due to the large hierarchies in the design. Modular reuse is also a growing tool in the PCB world. Cadence now has a very nice module/reuse tool that can even detect and create modules to prevent you from having to redraw the layout for a sub-circuit multiple times if its not instantiated as a module already. I always like when more people want to get involved in HW, but what the OP is showing largely exists in the form of TCL and SKILL scripts in current EDA SW packages.

I think a major question is WHICH current EDA SW packages.

Cadance OnCloud Platform for PCB Design is $1500/month. Altium Designer is $11k.

Most of the universe is in the world of KiCAD, Eagle, GEDA, and similar.

You're wrong.

The reason I gave a specific use-case is because in this use-case, it does work like that. I did not give a use-case which had RF signals, large currents, or much of anything else. I2C "fast mode" is 400 kbit/s.

The reason things like STEMMA work is because everything is slow enough that parasitics don't matter, and power is low enough that voltage doesn't sag. The lumped element model is nearly perfect there.

Even many domains not like that -- let's say audio -- have a lot of places where abstraction break, but where those failures are well-contained. An audio amplifier might have 50MHz gain-bandwidth, and the layout around it needs to be pretty tight, but once that's a block, things matter a lot less. That's why you can have all sorts of analogue synthesizers where people plug blocks together. A qualified engineer needs to design those blocks, but connecting them, abstractions hold pretty well.

No one will design an oscilloscope or a software-defined radio front-end using React for Circuits. However, as much as those occupy a lot of engineer time, they're actually in the very tiny minority of boards people want built. Most things are pretty simple and dumb.

Heck, even many high-performance designs have simple stuff around them. If you're making a radio receiver, the RF might be voodoo, but you know what's not voodoo?

- The controller for the buttons and the seven-segment displays

- The battery charger / monitor circuit (at least as a block)

- The USB-to-RF232 converter for development (at least as a block)

- The power, temperature, and other monitors

- The neato audio spectrum bar display

- The bass/treble filter (at least as a block)

- A lot of the test fixtures needed to test the RF equipment (not all; a lot of the other test fixtures are deep voodoo; it depends on what you're testing)

- Etc.

Reply on news.ycombinator.com