Do-it-Yourself ASICs (2013) | Hacker News Reader

Do-it-Yourself ASICs (2013)

analog.com

43 pointsomgwtfbyobbq15 comments

I'm sure I'm missing something. This is saying, for analog, if you can't find the right IC, then build that part of the circuit from discrete components.

That's true, but seems obvious, and I don't know of a practical general alternative for low volumes.

I read it as a reminder that yes, that should be obvious, should be 'bread and butter', not an ancient outdated forgotten practice.

It's kind of funny; I came out of school ready to design circuits with discrete transistors and they laughed at me. "Sure, that can work, but it will definitely work if you use this opamp." Much like we caution people not to write assembly, because the compiler can do it better.

But sometimes it can't. Sometimes the opamp designers made choices I would not prefer for my circuits. To me it's funny how after all this time, I've only recently started to consider transistor-level design again. I suspect it's a deeper understanding of the tradeoffs than I had as a novice engineer. Or hubris, thinking I can do it better.

There absolutely isn't a practical alternative for low volumes, there's barely even one for medium volumes. There once were IC houses with semicustom parts - they had a variety of transistors, resistors and maybe a few (really huge bipolar) gates onboard, and you could design the metal layer, resulting in a single custom mask on an otherwise standard product. Actually, Astec still has one, Device Engineering offer one, but there's not a lot out there. This was always a niche business anyway.

So yeah, it does seem obvious that we should design from discretes, actually crawl into the boxes in our block diagrams and build the things in there from scratch. But it's funny how easy it can be to stop doing that, to just find a single chip/module that does it all. And it's been amazing to me how well the IC market has mapped out to what I have needed over the past several years. Every time I've tried to build something, I found I could just buy it. Eventually, you think "buy" before "build." For me, it's taken extra effort to consider "build" first anymore.

Just like the author found, it's fun, too. Just buying that ublox GPS chip is okay, but building the parts I need is cool.

But don't waste time reinventing the wheel. I needed a small audio amplifier the other day and looked for an IC to build it. My junk box was empty of audio amplifiers so I went to a distributor and found a nice IC for 65p but minimum order 10 off. With postage the price was about £10, and I still had to find the other components and build it.

I was just about to order it when I realised it might be cheaper on eBay. There a found an assembled board with the same amplifier for 56p with a minimum order of 5 pieces:- £2.80 post paid and ready to use - and four spare ready-made amps for future projects.

Beating the system is fun, too.

The eBay sourced chips are likely counterfeits that won't meet all the specs.

In my experience the random stuff I buy off of whatever $Auction_Site seems identical to what I buy from $Local_Vendor, but with no QA or packaging. The defect/failure rate is higher and it takes a lot more time to arrive, but the cost is significantly less.

It's possible that there are more counterfeits, but I don't think the stuff off of $Auction_Site is likely (>50% chance) to be a counterfeit just because it's from $Auction_Site.

If it's a chip that you can search for in a product description it has a high enough profile to be an SZ fake. See this example of a fake signal gen IC that almost works properly:

https://www.youtube.com/watch?v=02XtneCHnDA

Add in a price that's too good to be true and the odds go up.

I can see that. I've purchased several laptop batteries that likely used old cells pulled from another laptop or whatever. At the same time, I don't mind that the runtime is a 1/2 to 3/4 of a new battery when the price is 1/5 of a new battery.

Not quite - what I meant was build it from whatever you need in the simplest way that does what you need. That may be just discrete components, or it may be a handful of ICs, or a mixture. And don't overlook the possibilities of embedding a minimal microcontroller to do any logic - it may be the smallest, cheapest, simplest solution.

And the other half of the message was to build it as a well-defined entity which can be proved by itself before it is used in its final application.

Design is not all. There is a reason IC design houses are vanishing, it is too expensive to design custom chips. Only large companies can afford the tapeouts, testing, Packaging, and large volume production.

Any idea what parts are too expensive, and why? One would think that over time tapeout would have more and more processes automated, as with some testing. Packaging could be made standardize as much as possible. Thanks.

Competitive performance requires access to latest technology. Those are super expensive.

Purely analog circuits, including those with digital functions as an ancillary feature, are never on the latest process node. Analog functions don’t scale down well, high drive currents or voltage support require minimum feature sizes. The notable increase per node in terms of frequency support would be wasted even on parts operating up to the gigahertz range. Even the most aggressive analog circuits, like current RF-CMOS designs, are only released when the node is a few generations old. The process control required to get them functional within usable tolerances for analog circuits like current mirrors and differential pairs is greater than what digital designs demand

My experience is that ASICs are used for a variety of scenarios, such as consolidating components used in circuit design. Current limiters for LEDs or signal processing are two examples. They are also useful for replacing processes that were previously done in software or on FPGAs in such a way that they perform with more efficiency. None of these scenarios necessitate performance, and the latest node size. Hell in the open source community an open architecture without proprietary device drivers(blob free) is often quite desirable.

Off topic: the cookie dialog in this site is a textbook example of dark patterns.

On topic: the article basically says you don't need your own asic since you can but whatever you need as discrete ICs from them??

Reply on news.ycombinator.com