Nice to see an article address what Moore actually said (transistor count doubles every 24 months) and not what people often thinks Moore said (performance doubles every 24 months).
It started out as every 12 months in the original 1965 article which lasted for about a decade before being updated to 18 months, and relatively recently people started talking 24 months.
In the early 1980s there was a time it seemed to hang in the air because all the clock rates in a micro were synchronized to the video scan. Steve Jobs had no idea the Apple ][ was going to last as long as it did so they came out with the the failed /// and the 128k Max that was really just an awesome demo.
It wasn’t until the mid 1980s that PC compatibles reached a place where you could buy a new computer that was exponentially better than the computer you bought 2 years ago but that was gone around 2005. Until then the smaller transistors were really faster and consumed less power but since then you mostly have more of them and moderate improvements in power consumption.
In the beginning, everything good correlated with more transistors: decreased pitch led to lower power per transistor, faster clock frequency (implying faster RAM), more RAM, and lower cost per transistor. In the decades since pretty much everything but "number of transistors" has fallen off. Pitch doesn't mean pitch, clock speeds are a function of cooling, cost is going up.
One thing that has correlated well with Moore's Law still is exponential cost to open a new fab, and fewer companies chasing the latest node. As Moore's Law slows, and possibly eventually ends, the end is going to be caused by economics, not just physics.
Yeah, to some extent- how much I'm not sure- Moore's Law became a self-fulfilling prophecy. Because everyone knew that transistor count was about to double, everyone- including the Wall Street investors- agreed that you needed to spend all that money to build that new fab, because if you didn't you can be damn sure your competition would. So the semiconductor industry "was allowed" by investors to continue making enormous investments in the next generation of fabs. If there was some similar law for, say, rockets, could we have had that sort of progression in space travel instead?
No, because the economics are different. To put some things in context:
- Cost of the Saturn V development program: $6.417bn then / $35.4bn in 2020 dollars
- Cost of TSMC's Arizona fab: $12bn now
- The Saturn V program launched a handful of humans into space for $0.185bn then / $1.23bn now per launch.
TSMC's fab may cost a lot, but it produces a very large number of chips and, most importantly, it's reusable. The most density- and performance-sensitive applications use the newest fabs first, but the fabs are not thrown out immediately after a newer process is developed. Instead, the older fabs start making cheaper chips that don't need leading-edge performance. All of that up-front cost is amortized over a decade of continuous use.
If investors had poured money into rockets instead of chip fabs, we wouldn't have had everyone going to space. The investors would have just lost their shirt. Space has inherent costs we can't engineer around (e.g. fuel) and getting there is primarily a scientific endeavor rather than a consumer good. Keep in mind, there's nowhere to go[0] in space once you get there, except back to Earth.
Meanwhile, chips are something everyone needs now, and owning a fab is still a profitable venture even when the R&D costs go up - at least, for as long as the manufacturing technology works and people are willing to pay a premium for better chips. It's not purely an R&D arms race.
[0] I am leaving the possibility of terraforming Mars out of this, as that has enormous problems on its own.
Cost per device might have flattened at the leading edge (where fabbing processes are increasingly customized, so skyrocketing costs are to be expected) but the trailing edge could be seeing more improvement, AIUI.
>and not what people often thinks Moore said (performance doubles every 24 months).
Performance double every 24 months isn't the modern take anymore. In the past 3 years all major PR has twisted the word "Moore's Law" to just meant transistor improvement.
It's not even that, and it never really has been. Anyone talking about performance was misrepresenting it. The prediction was that the density with which you're at the point of minimum-cost-per-transistor would double every year. It was also originally a prediction going out 10 years to 1975.
The more transistors you put on a chip, the less expensive each transistor is, but the higher the risk of defects, so there's always an optimum amount as your fabrication processes improve.
> The complexity for minimum component costs has increased at a rate of roughly a factor of two per year (see graph on next page). Certainly over the short term this rate can be expected to continue, if not to increase. Over the longer term, the rate of increase is a bit more uncertain, although there is no reason to believe it will not remain nearly constant for at least 10 years. That means by 1975, the number of components per integrated circuit for minimum cost
will be 65,000.
> I believe that such a large circuit can be built on a single wafer.
Personally, I am not particularly excited for the next round of Moore's Law on CPUs. What else are we going to be able to accomplish if CPUs get twice as fast?
GPUs, however, are exciting. ChatGPT is cool now. Real time voice recognition is cool. But can they run on your phone? Seems like people will probably make improvements on the algorithmic side to these things, but we could really use improvements on the GPU side as well. When GPUs are twice as powerful as they are today, I think we'll notice.
That won't happen anytime soon. GPU performance growth has outstripped CPU growth for well over a decade and that gap is likely to only widen in future.
Right. At one point floating point was a co-processor. Then it became technically feasible and supported by business cases so it was included in all chips.
It's only a matter of time before tensor cores are ubiquitous
What excites me about about ChatGPT is the fact that you can take a lot of data and a huge model and make it do something cool. Right now, "making it do something cool" costs tens of millions of dollars. If that cost can be brought down to the 10s of thousands of dollars, I think we'd start to see really mind blowing applications.
> What else are we going to be able to accomplish if CPUs get twice as fast?
I think we are leaving the era of "twice as fast" and entering the era of "half the power" meaning half the heat and twice the battery life. Wearables will continue to gain in popularity as they become more convenient. AR will be made possible by powerful, low watt SOCs.
I'm really excited for the future of NVMe, storage nowadays is hundreds of times faster than it was at the turn of the century and I think that'll open up a lot of opportunities
Cooling is probably less of a problem than many people think. Modern cpus are (or rather would be) very efficient, if you don't push them to the very end of their frequency-power bathtub curve, but the latest generations of desktop cpus are basically stock overclocked. They can achieve 80% of their performance at 40% the power. That basically means you could double performance by 2.5x-ing silicon for the same power/heat.
Sounds exactly like the giant core count server chips; well they're still a little overclocked but not nearly as much.
A major issue is the vast quantity of state machine dependent programs, like videogame servers, more simply coded data processing generally.
Games and other simulations generally are often single threaded since they're engineered as iterative processes rather than parallel expressions due to possible conflicts in state propagation.
I've wondered whether little micro-filament heat pipes sandwiched right into the chips might be possible, all plumbed into a heat exchanger on the side of the die that can be directly cooled, facilitating the extraction of heat from deep within the stacked chips.
Direct water cooling huh. With channels right in the silicon I imagine water purity is just that much more important.
God help you if the water gets gunked up with disintegrating gasket material[0] and blocks some of those channels off, or even worse, conductive stuff.
Someone needs to invent a diamond layer between stacks! High thermal conductivity with virtually no electrical conductivity, which is unusual as thermal and electrical conductivity usually go hand in hand. Diamond would be perfect to transfer heat out of densely stacked chips.
A lot of it is just two layers - the chip and the interconnect below. If you have a chip that generates a lot of heat, you can stack another one that generates less heat on top of the hot one and use it as a heat spreader (a lot of mobile chips are like that). Also memories have only a small part of them being used at any given time, so they are a good fit for tall stacks (current flash memory does that too)
On-die CPU caches typically use SRAM. Modern DRAM chips are already die-stacked in several layers. But maybe DRAM is more heat sensitive than SRAM, so more difficult to stack atop a stronger heat source?
The real problem is that SRAM get 20% density increase where transistors get 75%. Compound over a few generations and your chip area is all cache. Some foundry should focus just on that problem. Then recombine in package will the way to go.
Yes, that’s how Moore’s law has worked for 60 years: incremental advancements. The past 15 years have actually involved far more radical step changes than the 45 years prior.
This seems false to me. For example, electromechanical to relays, to tubes,transistors, integrated circuits, highly parallel ICs are all paradigm shifts rather than incremental, and yielded multiple orders of magnitude rather than incremental improvements in price/performance.
The next really significant boost is likely something like optical transistors, memristors, analog computing, bio-electrical computing, or anything that is highly divergent from the current base concept.
I mean, if the organization of the current basic idea of silicon and the software can align with a compute-in-memory paradigm, I guess that would be more than incremental in a way. But the lady seems to be talking about doing that in still kind of a half-assed way as far as that part.
Moore’s Law (formulated 57 years ago) by definition applies to dense ICs. It’s true, there were a bunch of advancements before the invention of ICs, but those were 60-80 years ago.
The Apollo Guidance Computer was a 16 bit computer operating at 2 MHz in 1966, costing $200,000 per unit. It was made out of many individual ICs, based on fab limitations at the time. A single IC today with the same functionality costs under $2 https://www.digikey.com/en/products/detail/renesas-electroni...
That's 5 orders of magnitude in cost over 56 years, with the only "paradigm shift" being "we can now fit the whole functionality on 1 IC".
That's honestly what we need right now. There is a plethora of new optimizations available (big.LITTLE, chiplets, 3D stacking, et. al) and neither Intel, AMD or Apple are using them all. There are also ISA optimizations still on the table for both x86 and aarch64, so the real race over the next 5 years will be seeing which company is able to compile these optimizations in a meaningful way.
Yeah. After reading the article, I have to agree. But I'm no expert in this area. It seems like it could certainly help. If I had to guess, things like ECRAM and reconfigurable analog computing chips will have more of an effect than this (assuming they can be made to work).
I'm amazed that we got as far as we have. It takes a well composed assembly of hundreds of machines, many costing USD 300,000,000 to operate as a "Fab".
The coming credit crunch as Boomers retire will make this type of investment much more difficult in the future.
Imagine a country of a hundred 18 year olds. They all get jobs on the same day, and start putting savings into the bank. The bank takes those deposits and loans them out to various productive (profit producing) enterprises.
Time passes. The hundred all turn 65 on the same day, and retire. The savings rate whipsaws, since nobody is depositing money anymore, and is instead withdrawing it. The bank can no longer make loans. Credit crunch.
US lending banks have had a reserve requirement of 0% since 2020. We are already in the days of cashless credit creation. No drop in savings will effect lending so long as this policy is in place.
Baby boomers life savings are still mostly in retirement funds being used to finance various ventures to get a reasonably good ROI. As the generation ages, the more money will be taken out of those accounts with fewer people to put their new earned money in the fund. Thus causing a credit crunch.
Demographic problems are a bit of an everything problem though.
> Investing in typical stocks doesn't directly fund anything, you buy the stock from some other private party. No funds go to the company itself.
Of course it does, buying a company's stock raises it's value, making it easier for a company to raise money by issuing more stocks instead of financing itself through taking on debt.
> If you believe you can predict stock market returns from demographics you should form a hedge fund.
Believing that the market is impacted by demographics does not really give you an alpha to use when investing, slower growth in the next ~30 years is just a trend.
" Of course it does, buying a company's stock raises it's value, making it easier for a company to raise money by issuing more stocks. "
Many companies are buying back stock (i.e. doing the exact opposite of this) Since the topic is intel, please explain when Intel last issued new stock and why they couldn't simply end dividends instead. (Giving money to their investors is the exact opposite of raising money).
"Believing that the market is impacted by demographics does not really give you an alpha to use when investing"
If you think stocks will be impacted by demographics you should be able to tell us what the correct price to earning ratio is of the S&P giving demographic a, b, or c and we should be able to trade based on that information when the S&P is mispriced.
If you can't do that, your ideas are not provable or disprovable and are not particularly compelling.
