If Moore's law is broken : what will be the consequences ?
Since the last two decades the level of performance of CPU have increased in a quasi linear way (the Extended Moore's law*).
Many chips constructors seems to encounter problems to be able to follow this law right now. Perhaps they will find solutions like multicore .., but imagine that the increase of performance slow down : what will be the consequences ?
- for the customer
- for the chips designers, and the whole computer industry (software and hardware).
Many chips constructors seems to encounter problems to be able to follow this law right now. Perhaps they will find solutions like multicore .., but imagine that the increase of performance slow down : what will be the consequences ?
- for the customer
- for the chips designers, and the whole computer industry (software and hardware).
Comments
It may mean a long term drop in the computer industry, not necessarily the application areas (e.g. medical diagnosis software, simulation, traffic control, etc.)
Note: in the late 80's, early 90's, the computer industry stalled. (and then, along came the Internet!)
Dr. L
2) Dianer : Right, it will be bad news for the hardware industry. People would not feel the need anymore to changer their computer every three years or less, especially if what Scott pointed it out is coming.
He cites Moore's law but chooses to pick a different start date to show different curve behaviour.
Moore's Law when originally proposed in 1965 as a doubling of transistors/area every 12 months.
Moore later revised it in 1975 to a doubling every 24 months.
Subsequent interviews with Moore suggest he compromised in the 80s on 18 months.
Now it's returning closer to a doubling every 12 months.
Kurzweil doesn't start in the Transistor and IC era, however, he chooses to compare computing power over time in terms of processing capability in bang for the buck terms.
Kurzweil argues that Babbage's Analytical engine should count as the first computer (if functioning in 1900), and progresses through calculating machines to Hollerith tabulators to the National Ellis 3000 built in 1928, in the "mechanical computing degices" category.
Zuse 2 and 3, and the Enigma fit between 1939 and 1941 in the "electromechanical (relay based) computers" category.
Colossus in 1943, ENIAC, EDVAC and machines up to 1955's !BM 704 belong in the "Vacuum-Tube" category.
1958's Datamatic 1000, DEC PDP 1 through 8 and the IBM 360 Model 75 can be classes as "Discrete Transistor" computers.
And from 1968's DEC PDP-10 to 1975's Altair 8800, 77's Cray 1 and Apple II, 1982's IBM PC, 1984's Macintosh, and everything since fit into the "integrated circuit computers" category.
We first note that the exponential growth of computing did not start with Moore's Law on Integrated Circuits. In the accompanying figure, "The Exponential Growth of Computing, 1900-1998", I plotted forty-nine notable computing machines spanning the twentieth century on an exponential chart, in which the vertical axis represents powers of ten in computer speed per unit cost (as measured in the number of "calculations per second" that can be purchased for $1000). Each point on the graph represents on of the machines. The first five machines used mechanical technology, followed by three electromechanical (relay based) computers, followed by eleven vacuum-tube machines, followed by twelve machines using discrete transistors. Only the last eighteen computers used integrated circuits.
I then fit a curve to the points called a fourth-order polynomial, which allows for up to four bends. In other words, I did not try to fit a straight line to the points, just the closest fourth-order curve. Yet a straight line is close to what I got. A straight line on an exponential graph means exponential growth. A careful examination of the trend shows that the curve is actually bending slightly upward, indicating a small exponential growth in the rate of exponential grown.
In terms of the limits we're likely to run up against, now we're talking physics.
Tomorrow it may be quantum physics.
Gates and transistor layers are rapidly approaching the limits where with only a few atoms of depth, the discrete design elements can't physically get smaller without starting to lose coherence or struggling to maintain electromagnetic or atomic properties.
Then it's nano time if we want to improve. Dynamic adaptations on the fly.
There's a cool video from one of the Intel Developer Forums in 1997 with Moore himself discussing this issue and linking up with an electron microscopy lab to narrate a video zooming into a chip and describing features angstroms thick. Not sure if it's still online, but it was cool to hear Moore himself walking you through this.
Transcript of Moore's Presentation "An update on Moore's Law" is here
Not the video i wanted, but other Intel clips on Extending Moore's Law are here
Seems more semantics than physics or wafer fab, but what is nanometer etching except precision hairsplitting.
Life goes on.
Now I guess they should have banned me rather than just shut off posting priviledges, because kickaha and Amorph definitely aren't going to like being called to task when they thought they had it all ignored *cough* *cough* I mean under control. Just a couple o' tools.
Don't worry, as soon as my work resetting my posts is done I'll disappear forever.
Now I guess they should have banned me rather than just shut off posting priviledges, because kickaha and Amorph definitely aren't going to like being called to task when they thought they had it all ignored *cough* *cough* I mean under control. Just a couple o' tools.
Don't worry, as soon as my work resetting my posts is done I'll disappear forever.
Originally posted by Placebo
I think that clockspeed will flatline for a while, until chip manufacturers say, "Fuck this, let's keep the density the same, and make the wafer a bit larger."
It would help if you knew WTF you were talking about.
Wafers = those huge discs you see at the chip foundry with hundreds of potential chip cores on each.
Increasing the size of a wafer doesn't really do anything other than potentially increase output and decrease costs.
Making chips bigger doesn't make them clock to higher frequencies. There's a reason why faster chips are SMALLER. More transistors doesn't translate to higher clockspeeds. More transistors usually does mean more heat and high power requirements...
Originally posted by Eugene
It would help if you knew WTF you were talking about.
Wafers = those huge discs you see at the chip foundry with hundreds of potential chip cores on each.
Increasing the size of a wafer doesn't really do anything other than potentially increase output and decrease costs.
Making chips bigger doesn't make them clock to higher frequencies. There's a reason why faster chips are SMALLER. More transistors doesn't translate to higher clockspeeds. More transistors usually does mean more heat and high power requirements...
I had the understanding that the main gain of moving to 90nm was reduced size.
Originally posted by Placebo
I had the understanding that the main gain of moving to 90nm was reduced size.
The move to 90nm is more than that...
What is primarily responsible for a processor's ability to scale across the move from 130nm to 90nm is a reduced transistor gate length. This should also reduce power consumption.
2.5 GHz G5s didn't just pop out of thin air. It took the jump from 130nm PPC970s to 90nm PPC970s to make it possible because shrinking the process generally allows you to up clockspeeds. That's how it's been done forever, and that's why your comment about increasing wafer sizes is completely...irrelevant.
Granted, object orientation itself is not the fault, but it does breed laziness.
Of course, that means that Matlab has to send over their hot sales reps in order to get me to buy their poorly coded software, which isn't entirely a bad thing. If you are ever in the position to buy a bunch of copies of Matlab, demand a live demonstration. It's fun and worthwhile for everyone.
Problems will be represented in different ways, and software will begin to be designed for parallel architectures.
On the hardware front, who knows... PCs will many processors, I suppose.
Well, that's what I reckon... m.
Originally posted by Powerdoc
If Moore's law is broken : what will be the consequences ?
It is a grand felony, with 21 year minimum prison terms for all chip designers.
Originally posted by a_greer
maybe we could see a focus shift to faster storeage(maybe solid stare HDDs), faster I/O busses, better media(something with much larger capacity than thumb drives but without the drawbacks of dvd-r, that is it would work like a HDD), and maybe an increaced focus on ram, there is no reason that any desktop over 1000$ today should not include 1 gig of ram, if manufacturers would stop being so sitingy, this could be done for the same price of 256 megs now.
Yes.
The problems for manufacturers, is that the turn over rate of computers will slow down. The number of unity per year in the world will decrease, and the market will slow down.
In order to survive and prosper, chip designers must do progress at any costs.