Forty Years of Moore's Law

kjh1 writes "CNET is running a great article on how the past 40 years of integrated chip design and growth has followed [Gordon] Moore's law. The article also discusses how long Moore's law may remain pertinent, as well as new technologies like carbon nanotube transistors, silicon nanowire transistors, molecular crossbars, phase change materials and spintronics. My favorite data point has to be this: in 1965, chips contained about 60 distinct devices; Intel's latest Itanium chip has 1.7 billion transistors!"

Comments

Keeping Count

[ackthpt]

Intel's latest Itanium chip has 1.7 billion transistors!"

That's Montecito dual core Itanium, w/24MB of cache (only about 120 million transistors actually per CPU with the balance largely that motherlode of cache) and you could probably fry a steak on.

"We can keep Moore's Law alive just by stuffing the cache!"
"Brilliant!"
"Brilliant!"
Suddenly they were crushed by a giant can of Guinness containing not even an electronic sausage...

Re:Keeping Count

[timeOday]

But this misinterpretation is the only reason anybody cares about the "law" in the first place. There's no reason to care about increasing transistor counts unless there's a payoff.

The problem with bigger & bigger cache is that it has diminishing returns. This is why Intel's "Extreme" chips are a waste of money.

The inability to do anything useful with all those transistors is why we're seeing the advent of multi-core chips, which are neat but fail to preserve the conventional single-threaded programming model. This places the burden of creating explicit parallelism on the programmer, and leads to more complicated code, which means it costs more to write and also contains more bugs.

Don't hold your breath...

[gaber1187]

So many people really doubt Moore's law will die anytime soon. Just because intel isn't jumping MHz every year, doesn't mean its ending... There are so many things left to do to squeeze out more performance in the same area or smaller. You can go to 3D stacks of transistors, higher K oxide dielectric, the list goes on and on. I agree with the article that says that we could see it go into the 2020s... the main problem that will hinder moore's law will be the economics of investing in new fabs, and waning demand of chips, not research and technology limitations. I see more money being pumped into memory chips and special purpose ARM style chips with a focus on low power. Eventually, people will just say, "Moore's law just doesn't matter anymore, the market has changed".

Re:Don't hold your breath...

[Anonymous Coward]

"Just because intel isn't jumping MHz every year, doesn't mean its ending..."

Maybe not, but there's certainly been a bit of a bump in progress recently; no notable new desktop CPUs, and certainly no increase in the complexity, component count or speed - unless you want to count cache - nothing in the last 18 months has fulfilled the criteria set out in Moore's Law. Having said that, this anomaly only applies to CPUs.

I would hazard a guess that the law still holds true in memory - major advances there in transistors per square inch - and almost certainly in graphics processors. I envisage more specialised chips appearding to take a lot of the core work from the CPU - World Physics Processors anyone?

With the current circumvention of limitations being to cobble two cores together on a chip, could this also be the route that GPU manufacturers take in a few years or so?

Do you have a source for the 120M transistors ?

[PaulBu]

I was looking for logic vs. cache break-down numbers for a while, obviously Intel is not keen on providing it on their own.

The way I see it, 24 MB = 1024*1024*8*24 * 6 transistors/SRAM cell = 1.2B transistors for cache, still leaving 500M for logic. Well, we can factor in address storage and cache access logic, but I'd still like to see some harder data than this.

Paul B.

Re:Do you have a source for the 120M transistors ?

[questionlp]

Keep in mind that the Montecito has 24MB of L3 cache, plus 2.5MB of L2 and 32K of L1 cache. You also need to include links between the two cores, the cores themselves, tags, bus interface and arbiter, plus redundant SRAM cells so that one or two defects doesn't render the die worthless.

I don't know how many additional SRAM cells Intel is planning in each of the cache levels, so the 1.2B transistors for cache can climb up to 1.4-1.6B.

Someone posted a number of 1.47B transistors for the L3 cache at Real World Tech. I'm not sure how credible or accurate that number is.

Another article on RWT shows approximate die floor plan and othat info at:
http://www.realworldtech.com/page.cfm?ArticleID=RW T100404214638&p=4

Re:law?

[kaosrain]

A theory is an integrated set of principles that organizes and predicts observations.

Re:Electronics Magazine!

[homerj79]

The paper he wrote for Electronics Magazine is here.

Re:When was the last time Moore's law was correct?

[fafalone]

Wow, how that display of ignorance got modded +5 Insightful on a site like Slashdot really makes you think.
First of all, Moore's Law implies that the number of transistors per integrated circuit will double every 18 months (which, is not really what he said, see Understanding Moore's Law).
Second of all, this has held true and is continuing to hold true.
Third of all, clock speed does not reflect transistor number or density, neither of which are the sole contributing factor to 'power' or 'performance'.


I don't know what's sadder; wondering if the parent was actually a joke, or wondering how it got +5 insightful. Damn.

Re:It definitely has less that 300 - 400 years.

[kesuki]

If the number started at 60 40 years means ~27 doubling of 60 so today's processores should have 8 billion tansistors 200 doublings of 8 billion is about 1.32*10^74 According to answers.com earth is composed of roughly 10^50 atoms and the Observable universe is estimated at 10^80 to 10^85 which is 335-356 years from now, not 300-400 Also, composing a transistor out of a single atom it pretty tough. plus you have to have gates etc. And if the whole observable universe is the processor, where is the rest of the system? ;) obviously you could make a system on a chip, but even then valuable atoms are being used and taking away from moore's law. plus the atoms of the device used to fabricate the observable universe into a giant processor... on the plus side, with that many transistors, you can probabbly encode the entire history of the universe into a mathmatically lossless codec that can achieve fit the entire sum of knowledge into a single byte of data. Some people believe this already happened, and the resulting processing caused the universe to collapse into a singularity and expolode into a new universe.

In addition...

[Dire Bonobo]

In addition, Peak Oil catastrophes can be headed off with biodiesel (Peakists say we'd need more cropland to produce a useful amount than we have; they're demonstrably wrong), with solar (Peakists say we'd need to use more cropland for solar energy, which is again demonstrably wrong---ever heard of a nice, sunny desert?), and so forth.

Peak Oil folks take one valid idea (oil is finite, and running out will be painful), but then devolve into irrational fear-mongering about it. If thermal depolymerization can net the US four billion barrels of oil from agricultural waste we currently throw away, running out of ground oil ain't going to be causing a new Stone Age.

You know what bothers me even more?

[Moraelin]

That it's mostly useless in real-world terms anyway.

Sure, taking Moore's law literally, computers are 1 million times faster than 30 years ago. Arguably that should translate into _more_ than 1 million times more work per second, because compilers have evolved too, and expensive optimization techniques have become more affordable. (A compiler optimization technique that would have taken a week on a 70's mainframe, now takes seconds.) We also have better tools.

But are we doing 1 million times more with them? Nope.

Every time we get better tools, the accounting dept just get the idea "w00t! Now we can _really_ hire untrained monkeys to use them." In fact, the better tools and computers you get, the worse code you get.

It's not just code _performance_ that went south, any clue about security or good design went south too. Actually analyzing what could go wrong got at some point replaced by magic talismans like "we use Java so we can't possibly have a security problem" or "we use HTTPS, so our site is by definition secure." Too bad that one only has to edit an URL to bypass all those magic talismans.

And then there's the BDA (Buzzword Driven Architecture) effect.

The whole computer industry is one big scam where marketting is in control, and the biggest outright liar and con wins the contract. So every single dud or unfinished (or outright _stupid_) idea is marketted as _the_ second coming of christ, cure for all enterprise problem, cure for cancer, etc. And there's one born every minute who actually believes that drivel... yet again.

So programs are written with the sole purpose of having as many buzzwords in them as possible. Everything _must_ involve a SOAP call, to an EJB, which uses XSLT instead of just processing the damn data, etc.

True story: I've actually benchmarked one such crap buzzword-driven framework we were forced to use here. It took 1.1 seconds for a call to an empty method, on a 2.26 GHz P4 computer. No, not milliseconds. 1.1 _seconds_. A cool 2.5 billion CPU cycles just for a function call to an empty function.

We've actually exceeded Moore's law. A computer in '70 may have been 1 million times slower, but we're taking a _billion_ times more computer cycles to do the same. Yep, the modern version actually runs _slower_.

Being an ex-assembly programmer, that realization hurt. I'm talking physical pain.

So to end this long rant, IMHO I'm not sure that Moore's law will become that irrelevant any time soon. You could increase the CPU speed another 100 times, and someone will just find the monkeys to write 1000 times slower code for it.

Re:Moore's Law is probably being exceeded at...

[Anonymous Coward]

I post this every six months or so when a moore's law article comes up... (anonymizing to avoid karma whoring)

. 8086. . . . .: 0.03 million transistors (1978)
. 80286 . . . .: 0.13 million transistors (1982)
. 80386DX . . .: 0.27 million transistors (1985)
. 486 . . . . .: 1.2 million transistors
. Pentium . . .: 3 . million transistors
. Pentium Pro .: 5.5 million transistors
. Pentium 2 . .: 7.5 million transistors
* Nvidia TNT2 .: 9 . million transistors
. Alpha 21164 .: 9.3 million transistors (1994)
. Alpha 21264 .: 15.2 million transistors (1998)
. PPC G3. . . .: 22 . million transistors
* Geforce 256 .: 23 . million transistors
. Pentium 3 . .: 28 . million transistors
. PPC G4. . . .: 33 . million transistors
. Pentium 4 . .: 42 . million transistors
. PPC G5. . . .: 52 . million transistors
. P4 Northwood : 55 . million transistors
* GeForce 3 . .: 57 . million transistors
* GeForce 4 . .: 63 . million transistors
* Radeon 9700 .: 110. million transistors
* GeForce FX. .: 125. million transistors
. P4 Prescott .: 125. million transistors
* Radeon X800 .: 160. million transistors
. P4 EE . . . .: 178. million transistors
* GeForce 6800 : 220. million transistors

properly formatted at http://nothings.org/trans.txt

I wish I'd bothered to keep citations for all these numbers, but I didn't realize when I started this how long it was going to go.