[extropy-chat] diffraction limit

[Brent Neal]

 (5/31/04 17:13) Eugen Leitl <eugen at leitl.org> wrote:

>> but rather an industry consensus on the progress they expect to make. 
>
>No, Moore's law is an empirical observation about integration density 
>progress over time, which has been linear in semi-log plot since 1959.

It was an empirical observation when Gordon Moore made it. Since then, its a marketing tool and an industry goal. Nothing more.


>
>> There is nothing sacred about it, nor is there any particular reason 
>> why we -have- to continue to follow it at its 18-24 month doubling rate.
>
>Yes, there are very compelling reasons why Moore's law will hold until at
>least 2012, or so -- economical demand driving feature shrink in
>semiconductor photolithography. Major deviations from that will indicate that
>something has gone seriously wrong with our economy, or ourselves. Physics
>forecast is plain sailing, under clean blue skies.

I never said that the underlying science was the problem. My argument was solely based on the economics of the semiconductor industry. The costs of either a new fab or a fab refit are prohibitively high, at a time when the products being made are being squeezed into lower price points due to commoditization of the technology. Right now, there is virtually -no- demand for high end chips in any significant quantity, when compared to circa 2000.

>
>I do expect a discontinuity before molecular electronics can pick up the
>torch -- but there might not be one.

Or optoelectronics, or spintronics, or, or, or....

>
>
>Big costs are building a new plant, refitting an existing plant to 300 mm are
>less as Intel has shown. This is completely nonapplicable to organic devices
>and self-assembly nanoelectronics, where the fab will become cheaper, and scale
>to desktop fabbing.

The cost of refit is not significantly less. There is a very good reason no one is fabbing 400 mm wafers. The technology is there, wafers can easily be grown that size. No one is willing to refit. There is no return on that investment, despite the efficiencies you get from being able to put more devices on the wafer.

I disagree about the applicability of the costs to organic devices.. You're talking like a theoretician, assuming that the engineering problems are trivial.  I certainly think that desktop fabbing is a compelling "holy grail," but it isn't going to start there. As I mentioned in a previous email, I think it is quite naive to assume that the first generation of fabs for that technology will be any cheaper. The only thing that will save you is having the cost growth be smaller. I even buy the argument that the associated growth constant might be negative, representing technology that gets cheaper as it matures. But you have to get there from somewhere, and despite the fervent wishes of researchers everywhere, it will not spring fully-formed from the head of Zeus. A lot of engineers will spend a lot of time and money making it happen, solving the problems that said researchers neglected to consider. :)

>
>OLED displays and printable electronics for smart RFID tags is already very
>marketable, so I wouldn't start worrying yet. My next hot pick would be
>nonvolatile organic memory, then molecular FPGAs (reconfigurable hardware in
>general).

I'm pretty excited about the next generation of CMOS-based FPGA technology myself. There is a research group in California somewhere who has figured out how to get single-cycle reconfigurability for an FPGA. One of their PowerPoint slides showed one of these FGPA's running in tandem with a general processing unit, and talked about the efficiencies that could be gained from being able to dynamically reconfigure the FPGA bank in order to optimize for the code that is about to come streaming down the pipeline. Cool stuff. I'd be surprised if no one is trying to commericalize it as we speak. Once that engineering problem is solved, then using molecular FPGAs to reduce the density of the gates on the chip (and increase max throughput) is icing on the cake.

> 
>> There's also the fact that the current top end of CPUs have 
>> begun to exceed the needs of the average user. Without a 
>
>Moore's law doesn't unfortunately translate into system speed very well, and
>current systems are always too slow for the power user -- that'd be power
>gamer.

Sure it does. You can plot clock frequency on the same semilog plot as well, though the resulting curve is more noisy. And even the power gamers are not finding a compelling reason to upgrade their boxes. You can't tell 120 fps in Counterstrike from 100 fps visually. At that point you're just comparing penis length. And the power gamers with budget to buy a large penile surrogate aren't a large enough market to sustain Intel's R&D efforts. As we've seen time and time again, the sweet-spot on that curve is the corporate market.

>
>Lower power consumption isn't a bell & whistle, it's the only way to drive a
>current process forward at a power density now rapidly approaching nuclear
>reactor cores. Ditto power supplies, the demand spikes rise to fast to supply
>the CPU. 
>

Sure it is. Your process isn't going forward - you can't, because you're dissipating 100W or more, and your cooling tech is pretty much maxed. So you underclock the chips, add some clever trickery with heat pipes, ignore the fact that you've sacrificed about 10% of your peak performance, and call them "mobile" processors.  Throw in a 802.11 interface on the south bridge, and you've just re-spun your current technological obstacle as a win. Good for you. But if you were IBM, and understood that GFLOP/watt is something that people really do care about, you'dve done a better job engineering your chips for power consumption and heat dissipation in the first place.

I remember when I was working on a cluster-based supercomputer. We had to reserve half the budget for the computer to pay for the new cooling system and the new electrical line to be run to the building that would house it. By comparison, the IBM SP3 that I was using for my research had pretty much equivalent performance with much lower operating costs. The current shipping PowerPC chips at 2 GHz offer comparable performance FLOP-wise to the Pentium 4s at a quarter of the power dissipation. Even the AMD chips offer equivalent performance FLOP-wise at about 70-80% of the power dissipation, since they've integrated a lot of cool stuff they licensed from IBM.

>> their M series chips - since laptops are where the growth 
>> in personal computing is currently. (Of course, if they'd 
>
>The growth in pesonal computing is embeddeds. Mobile applications are
>low-power, which is incompatible with high-performance in current 
>technology.

Are you agreeing with me, or trying to tell me that the cell phone market has more money than the laptop market right now? If its the latter, then I will point out that the processors in most cell phones are using tech that's one to two generations behind what's currently in the top-end on the desktop or in a laptop. That makes it hard to argue that the industry is going to continue to push harder to integrate more transistors per unit area. 

B

-- 
Brent Neal
Geek of all Trades
http://brentn.freeshell.org

"Specialization is for insects" -- Robert A. Heinlein