For the previous couple of years, I’ve been utilizing AI to help within the design of my {hardware} and firmware tasks. The expertise has usually been good, though the result isn’t all the time helpful. So, I’m presenting a brief abstract of some of the duties I’ve tried and offering my unscientific grade to the result. The grades shall be averaged on the finish. Be aware: I shouldn’t have any paid AI subscriptions—I solely used free AI instruments, principally Microsoft Copilot and ChatGPT (though I’ve tried a number of others). These are only a few of my experiences utilizing on-line AI.

Do you’ve a memorable expertise fixing an engineering downside at work or in your spare time? Inform us your Story

Changing voltage to % cost

Grade: A

I needed to point out the cost remaining in a lithium polymer battery used to energy a design. This isn’t simple because the perform to transform voltage to % cost for a lithium polymer battery just isn’t a linear perform. I requested Copilot to make a desk of 20 voltages from 3.2 V to 4.2 V and their respective cost percentages. Then I requested it to create a C perform to do that conversion. It created this properly, together with linear interpolation.

Discovering the median with out sorting

Grade: D

Some time again, I wrote a Design Thought (DI) article on non-linear filters. Whereas doing this, I queried Copilot to create a C program that may discover the median of 5 numbers and do that with out sorting. (No sorting for a small variety of factors is helpful for rising pace.) It created a nice-looking program—good formatting and good feedback. It additionally compiled effective. The issue was that this system didn’t work—it discovered the incorrect worth for the median in some circumstances.

Initializing an ADC

Grade: C+

One other undertaking required me to put in writing code for the SAMD51 MCU to initialize the ADC for high-speed sampling. As I used to be attempting to get most pace from the ADC, it was a considerably complicated setup, particularly the clocking system. I attempted creating the code in each Copilot and ChatGPT a number of occasions.

Some code wouldn’t compile attributable to issues like unhealthy register names, and a few code would simply not work, giving no ADC readings. After some forwards and backwards, these points have been corrected. A number of of the feedback in code have been deceptive or simply plain incorrect because it utilized to clock frequencies. Because the code bought near a working perform, I took over the code and reworked elements of it to make it work.

Graphic design

Grade: C+

I used to be performing some LCD graphics design for a undertaking, and one half was a battery cost indicator. This image, for battery % of cost, was to be displayed on an LCD with an ILI9321 controller. (This normal determine seems to be like an AA battery with a inexperienced inside representing the % cost.)

I requested Copilot to put in writing C code for this utilizing the GFX graphics library. The size of the inexperienced fill labored properly, however the battery determine appeared nothing like a battery. It was a rectangle with two giant circles on each ends. I needed to rewrite parts of the code myself.

Grade: F

In the identical undertaking, I requested Copilot for a USB image written utilizing the GFX graphics library, as above. This didn’t appear like the trident-like, common USB image. I used to be basically three protruding from a central level at varied angles. It was unusable.

Enclosure design

Grade: D-

Subsequent, I attempted to have Copilot and ChatGPT design an enclosure that may work on a workbench, permitting the person to see the LCD and to simply join BNC cables. All I bought have been photos of rectangular bins. Regardless of how I requested for a extra distinctive form, it by no means went a lot past an oblong enclosure. Then, even the oblong field couldn’t be delivered as a usable 3D file “step”, “stl” file with out utilizing different applications.

Filter design

Grade: C-

I requested ChatGPT, “Are you able to design and show a circuit that takes in a sign, AC {couples} it to a acquire stage of 5, after which filters it at 120 kHz earlier than outputting it?” As a substitute of explaining the consequence, the picture in Determine 1 will communicate for itself.

Determine 1 ChatGPT’s output for a filter design that takes in a sign, AC {couples} it to a acquire stage of 5, after which filters it at 120 kHz earlier than outputting it.

It did embody a pleasant clarification of how parts have been chosen, however the schematic was principally unreadable. Devoted instruments equivalent to TI’s Webench filter design device, Analog Units’ Filter Wizard, and ST’s eDesign Suite are the correct instruments for filter circuit design and are literally simpler to make use of.

Grade: Ungraded

I attempted to create C code, in each Copilot and ChatGPT, for calculating coefficients for a digital Sallen-Key 2-pole high-pass, low-pass, band-pass, and band-stop filters. I attempted many occasions and couldn’t get a great working algorithm. The code was shut, however the filters didn’t perform appropriately. Finally, I discovered the code after an in depth Google search. It’s potential my testing might have been a part of the issue—uncertain.

Grade: B

Alongside the best way, I attempted a number of smaller queries, a lot of which have been very useful.

A lab pocket book

I’m certain among the points are my ability in creating the AI prompts. This actually made my makes an attempt take longer as I had so as to add extra element in follow-up prompts. I truly discovered this conversational model extra partaking than utilizing search engines like google. It’s not like a Google search, the place you possibly can’t usually do follow-ups to your question—it’s a must to re-enter your authentic question with a modification.

The AI techniques work way more like a dialog with a colleague. You’ll be able to inform it that the code it gave you didn’t compile, because it didn’t acknowledge a register identify. Or you possibly can ask it to present you quicker code, or change a resistor worth in a circuit, and recalculate the remaining parts.

One factor I discovered when writing this text is that each ChatGPT and Copilot hold a whole historical past of conversations we had. It’s form of like a lab pocket book, displaying your path to a sure design—very useful.

A C ranking

Wanting on the common grade, it is available in between a C and a C-. I’ll give it the advantage of the doubt and name it a C. The C ranking matches my intestine really feel additionally. The interplay is pretty straightforward—it seems like interacting with a coworker. The dialog goes on, making an attempt to fine-tune the ultimate reply. The interplay course of is significantly better than doing a Google search and getting a listing of issues to select from, with out a straightforward strategy to refine the search.

Does it save time? That’s onerous to guage as I’m nonetheless studying methods to create higher prompts. Generally I get a helpful reply immediately. On extra complicated queries, I’ve gotten pulled down a rabbit gap and wasted time whereas the answer diverged from what I used to be on the lookout for. There have been occasions when it had me attempting to finetune the consequence, and I turned to Google and bought a solution a lot quicker.

You’ll be able to simply be lulled into the sensation that you just’re conversing with a savant, however it could be extra like AI-splaining. Each reply exudes confidence, nevertheless it could possibly be the boldness of ignorance. Keep in mind that these solutions haven’t been checked or examined.

Will I proceed to make use of it? Actually… I’ll get higher at utilizing it, and the instruments will proceed to enhance. What I wish to see is an AI device targeted particularly on electrical engineering ({hardware}, firmware, and system design). This may increasingly focus its expertise on discovering or creating circuits, and with the ability to dig down deep into information sheets, and so on. It could even be good if it may take a look at its outcomes by way of simulation or by executing the code in a sequence of checks. Perhaps sooner or later.

All in all, it’s value utilizing and everybody ought to give it a attempt, simply test the reply carefully.

Damian Bonicatto is a consulting engineer with many years of expertise in embedded {hardware}, firmware, and system design. He holds over 30 patents.

Phoenix Bonicatto is a contract author.

Associated Content material

• Design platform simplifies electronics design with AI
• How generative AI places the magic again in {hardware} design
• Understanding generative AI and its functions
• What can AI do for analog design?
• Rising bit decision with oversampling
• How AI is altering the sport for high-performance SoC designs

The publish How assist are free AI instruments for digital design? appeared first on EDN.