bit about transistor cost (2023 Update)

"If they weren't so good, why would I buy so many?" Anyone remember that Volvo ad?

Gradually electronics design without having access to a silicon factory becomes useless, hopefully the process is slow enough so we don't see that in full. Sort of like nowadays you can somehow master an internal combustion engine if you have a lathe and a milling machine but you have no chance to make it comparable to those car makers make, not to speak about cost.

Some things have got good enough. Hammers, spoons, beds, LED lights, microwave ovens. Moore's Law can't go on forever, and is probably at or in same cases past its practical limit.

We don't need 3 nm chips to text and twitter. I can't imagine my cell phone needing to be better hardware.

I need a dumb, last-gen FPGA that is less fancy inside but fast pin-to-pin. The trend is in the opposite directions.

Maybe Moore's law is running on psychological momentum, fear of getting behind. I think I can see that happening.

Oh they have already bloated the software so the need for todays (and way back from today) hardware would be there. Just have faith, they'll manage it for 3 nm if say TSMC get there.

I get this obviously, I have similar needs (not yet your ps thing but for how long). But this is my point, so you can make what you want to make you will need access to a silicon factory... Clearly you can do better if you design your silicon instead of tweaking an fpga. However this is unlikely to become a viable alternative simply because of cost - and well, only the large ones will be allowed to design. Does not get much shittier than that (not just for the likes of us) but this is where the world is heading.

I don't give that much thought, I think it is gone since the clock frequencies for processors etc. stopped getting higher but I am neither sure not interested in what exacltly that "law" means.

Depends. There are very few applications that will support the sheer NRE cost of a full custom chip down at single-nanometer nodes.

Cheers

Phil Hobbs

I can imagine some serious head slapping when someone finds a fatal flaw in this mask set.

Welcome to the future :). Like I said before, I just hope I don't live long enough to be part of it.

Your imagination is very limited. I'd like my phone to be a LOT faster and use less power. Both of those things are what drives semiconductor advances. I'd like my laptop to run as fast as now or faster, but using less power so the battery lasts longer.

??? FPGAs, like most semiconductors, run faster as they shrink the feature sizes. There are companies that address the low end. Try Gowin and I believe Renesas, who bought Dialog is offering small FPGAs although I don't know how fast.

Like with many devices, if you aren't going to buy millions, you are invisible to the companies providing the product.

LOL! You are so funny sometimes. I guess the companies that are slower to adopt new technology actually make more profit. Then they get the money to play catch up and make less money again. Yeah, that's a thing.

I suppose that fits with your philosophy that everyone who doesn't think like you is an idiot no matter how successful they are.

I would like a *lot* more battery life - the speed is more than adequate for my needs. I'd trade slower when idle for longer life.

Likewise with PC's. I'm in the market for a new one right now but I'm not convinced that any of them offer single threaded performance that is

3x better than the ancient i7-3770 I have now. That has always been my upgrade heuristic (used to be every 3 years). Clock speeds have maxed out and now they are adding more cores (many of which are idle most of the time). Performance cores and efficient cores is the new selling point. It looks on paper like the i5-12600K might just pass this test.

Software will always grow to use the memory and speed available. CPU cycles are cheap and getting cheaper and humans are expensive.

IBM claim to have 2nm chip fab technology as of this year.

It may yet swing the other way when simulations are so good that the conversion to masks is essentially error free. Where it gets tricky is when the AI is designing new chips for us that no-one understands.

This years BBC Rieth lectures are about the rise of AI and the future by Stuart Russell of Berkley (starts this Wednesday).

It is still at least partially holding for number density of transistors if not for actual computing performance. We must be very close to the limits where quantum effects mess things up (but 3D stacks allow some alternative ways of gaining number density on a chip).

It was originally specified in terms of transistors per chip.

They can be edited, up to a point. Dunno how easy that is for a < 10-nm mask though.

Cheers

Phil Hobbs

Part of the problem, is that those chips go into disposables. Auto computers are wedded to one particular model and option set of auto, cannot be easily made generic replaceables. PC high tech components are churned into landfills on circa 5 year timescales. Cellphone processing is even less open; my service provider is now nagging me to do yet another upgrade of my pocketable. Tablets, same story.

RAMBUS-memory motherboards got recycled lots sooner than most owners liked.

Face it, we buy many because we can't repurpose the old ones. The future of high-tech LSI CPUs is to go into NUC or MacMini disposables, or server-class boxes that die of energy effiiciency issues, and even the socketed RAM is the wrong generation for their replacements. Power cords, RETMA racks, we reuse; no shortages there.

Maybe, someday, auto networks will accept generic plug-ins, and an auto chassis can expect a lifespan that goes longer than the lifespan of its irreplaceable data-handling bits. And, maybe I can cluster old PCs, with ethernet, and using NAS disks...

afaik there is a lot of masks to a chip and you can change only some of them

That happened around 1990. The electron beam tester I was working on back then was the next generation of a unit which famously trimmed three months off the development time of the Motorola 68k processor chip set.

The project wasn't canned because out machine didn't work - we did get it working quite well enough to demonstrate that it was an order of magnitude faster than it's predecessor - but because simulation had got good enough that most mask sets produced chips that worked.

The older, slower, machines were quite fast enough to check out that the simulation software was predicting what actually happened on the chip and that killed our market.

Well, if you've got 13 metal layers, various ion implant steps, and so on and so forth, plus 8-exposure litho for the fine stuff, that can add up, for sure. ;)

Cheers

Phil Hobbs

Actually I think I had seen that of IBM... and had forgotten. I am not sure software just grows to saturate the hardware (which it of course does, we all do it). Sometimes I think they bloat it on purpose in order to have the market for the next generation of silicon (not that this is a bad thing, better silicon is welcome, just the secrecy about it is not). Or may be it is not "on purpose" in some sinister conspiracy-like way, may be it just regulates itself like this. But the software is

*way* too bloated for me to just disregard the conspiracy idea...

So it may still be working then? These figures keep on getting more and more insane, no wonder only a few factories on the planet can do it.

It is a shame such an advanced machinery has been lost (or did it survive for some niche applications?)

I bought an i5 machine and it was a real dog. I said something to the effect that they ran out of ways to add transistors to improve the speed of CPUs a few years ago and someone listed a number of architectural improvements they've added for a 20-30% boost.

It has been quite some time since you could expect significant speed improvements by adding transistors or faster clock speeds. I think it was the Pentium 4 where the clock rate peaked at about 3 GHz by adding pipeline stages for shorter gate delays. But the cost of pipeline stalls pretty much mitigated that advantage. I believe people could overclock the Pentium 3 to run faster than the 4.

It was never about performance, it was just the number of transistors doubling every 18 to 24 months.

We keep hearing that the limit is just ahead and they find ways of working it. I'm flabbergasted they have reached single digit nm, How big are silicon atoms? The Corona virus is 70 or more nm. We could build a whole bunch of transistors on one virus. I seem to recall a Ball Semiconductor who wanted to print ICs on balls. I don't recall the advantages. They ended up providing some services from the processing technologies they developed.

Yeah, it was just an observation back when the number was still in the thousands. It is interesting that the advancement wasn't a lot faster in the earlier stages, but that may have had to do with finances since the semiconductor market was so much smaller then. Less R&D money available. I think the physics has been developed along with the miniaturization push. They couldn't go faster because they didn't have the science to design smaller transistors at any given point. They needed the to build smaller transistors to study before they could put them into production. I took one semiconductor course and that's what the guy said basically. They first had heuristics which let them build devices and the understanding came as they worked with them.

tirsdag den 7. december 2021 kl. 23.32.09 UTC+1 skrev snipped-for-privacy@gmail.com:

Aside from smart phones and maybe some IOT stuff in the future I think you're right. The penalty is a bit of power and speed.

Of course it's possible some amazing new market will come along of left field that will generate demand, but it's hard to imagine something brand new on the global scale of smart phones (~1.5bn units/year) that is also power-sensitive.

I think the mask costs are more of the order of $1m or $2m, not including design, obviously.