Comment by ChuckMcM | Hacker News Reader

Comment by ChuckMcM | Hacker News Reader

Parent Full thread

This captures it. The cost to add the transistors to put an ARM CPU in silicon is < the cost of the metals in the USB and HDMI connector you use to talk to it.

I gave a talk on the Internet of Things where I tried to communicate this point clearly, for more and more applications the marginal cost to add a computer is nearly 0, the marginal cost to add a network is about $0.35, so a lot of things that wouldn't have had networks or computers in them before suddenly do.

Back when Sun Micro was trying to get everyone in the Java group to write up patents on anything they could think of, James Gosling in what was a great comment on the process wrote up one for Java inside a light switch. He reasoned it was the most ridiculous patent ever since a light switch from Home Depot was $1.50 and there was no amount of "coolness" you could add with Java that would merit building a light switch with its own processor. What he didn't count on was that the cost of the computer that could run Java would drop below the cost to make a mechanical switch.

> The cost to add the transistors to put an ARM CPU in silicon is < the cost of the metals in the USB and HDMI connector you use to talk to it.

This statement amazes me. It may be true, but it is very counterintuitive to my experience.

A micro USB connector is about $0.35/unit (for quantites of 1500). An HDMI connector is about $0.50/unit (similar quantites). These prices are from Digikey. So, in larger quantites from larger distributors, you could certainly get cheaper prices. The cheapest ARM processor (cortex M0) is about $0.78/unit on Digikey. I understand your point is about licensing the ARM design and integrating it into your silicon, but most low cost devices I've seen have just been PCBs that integrate these off the shelf components. I would have to imagine the cost of hiring engineers to do the VLSI design/integration, the cost of licensing the ARM CPU, and then the cost of fabricating/testing your silicon would have to exceed integration on a PCB. So, I would then assume that the processor, while certainly cheap, is still a very substantial portion of the cost of the device. But, I have no data on the cost of ARM licenses to back that up. You're definitely right in asserting the trend is cheaper and cheaper processors, but I don't think we've arrived at the "processors are so cheap they're basically free" world quite yet.

The difference is that what you and I can buy vs what can be done in the world of building semi-custom ICs for your consumer gear.

The ARM7/TDMI core is about 100K transistors, that is nominally a square 316 transistors x 316 transistors, which with a 22 nm process is about 3 microns square. The marginal cost to add 3 square microns to a chip is very nearly 0, as an example the 'test feature'[1] on a chip the company I worked at in 2000 was 18 square microns and "wasted space" in the final chip. I say its 'nearly' zero because while the cost to produce the chip doesn't change measurably, the yield curve does and the 'cost' is the chips that fail due to this extra core not functioning.

So if you make a consumer electronics gizmo in quantity and it has a semi-custom chip on it, adding a computer to that chip these days won't make your semi-custom chip that much more expensive and by having a programability aspect you can add features without re-spinning the chip.

As for the cost, TSMC offers "add an arm core" to your ASIC as a design service. I've not been part of a wafer start negotiation for over a decade but it would not surprise me in the least if they offer to throw that in for free these days to sweeten the deal.

[1] The "test feature" is a part of the chip that the fab uses to verify the wafer processing worked correctly, it generally can be probed to with a simple voltage or current pulse to quickly screen out die which didn't get baked correctly.

That doesn't sound right: 22 nm * 316 = 69 microns, giving 69^2 == ~4800 square microns; a transistor is larger than the feature size, and don't wires take the majority of the area? (Still, that's well under a square millimeter.)

So we haven't quite reached the day Eric Drexler hinted at with "so-called microcomputers still visible to the naked eye".

Nice catch, in my haste I was dividing and should have been multiplying. Typical transistor size is 4x feature size. So 316 * 2 * .022 ~ 14 microns and squared its ~ 200 square microns.

That said, we are still talking about an incrementally small addition to a chip.

Oops, I misplaced a decimal myself. That's amazingly small (still without wiring overhead, of course).

Do you have a link to the video of your talk?

Doesn't look like it, it was for an IEEE symposium in San Jose.

Reply on news.ycombinator.com