Comment Re:Microsoft Walgreens(tm) (Score 1) 40
I agree with everything in your post except for that. My search engine is startpage.com, which acts as a proxy between me and Google so that it has no way of knowing who made which query.
I'm going to make some rough approximations here.
There are difficulties in dissipating power in high speed processors. Assume that the power that can be dissipated is proportional to the area of the chip. Relative to a single active layer chip, the power that can be dissipated per layer is 1/(number_of_layers * thermal_conduction_to_coolant). Thermal conduction to coolant is dominated by copper in the heatsink and SiO2 in the chip. Copper is at least 200 times more thermally conductive than SiO2. Assume that the maximum acceptable temperature rise is 50 Kelvin across a 1 cubic centimeter copper cube; that corresponds to 200 Watts. Assume that diminishing returns occurs when the thermal drop across SiO2 equals the drop across the copper. Since they add, if we keep the limit at 50 K the limiting power is 100 Watts. The implied thickness of SiO2 is (1 cm)/200 = 50 microns. How many layers can be squeezed into 50 microns?
A brief internet search seems to yield a minimum layer thickness of 100 nm (0.1 micron) for gate logic -- (1 active layer plus many interconnect layers.) Thus 500 active layers can be squeezed into 50 microns. What happens then?
Power dissipation in CMOS logic, ignoring leakage, is proportional to freq * V^2. Let our single layer CPU performance be 1 unit, limited by 1 cm copper and running at 1.2 volts (There's very little SiO2 for the heat to pass through.) At first glance, our 500 layer CPU with same voltage limited by 1 cm copper plus 50 micron SiO2 is 1 * (500 layers) * (1/500 heat per layer) * (1/2 thermal conductivity) = 1/2 unit. Layering loses. However, that is not the whole truth. Layering allows many more transistors, thus more clever circuitry, which might be enough to improve the performance some. 3D means shorter interconnects, shorter interconnects means less capacitance, less capacitance means less power dissipation. (The other major contributor to capacitance is the FET's gate.) I can only guess how much lower heat (more speed) that allows. Maybe 1.5X? speed is then 3/4 unit. That (1/500 heat per layer) is (1/500 speed) and with CMOS reduced speed allows reduced voltage.
Over a limited range, CMOS speed is proportional to voltage. By lowering voltage, heating is reduced. Thus reducing voltage means speed does not have to be reduced to 1/500 of the single layer CPU. With a supply voltage of 1.2 x 1/10 = 0.12, speed reduced to 1/10, power per layer is reduced to 1/1000 compared to the single layer CPU. 500 layers operating at 1/10 the speed is a 50x performance improvement.
Alas, we can't do that. Huge CMOS CPUs can't be made to operate at 0.12 V, and I don't know if it will ever be possible. I'll guess and say that somewhere in the range of 0.3 V and 0.6 V will some day be practical. If it's 0.6 V, speed could be 1/250, times 500 layers = 2 units. If it's 0.3 V, speed could be 1/62.5, times 500 layers = 8 units.
The above is too optimistic, because of difficulties in controlling threshold voltage and leakage, and the difficulties in massive parallelism and massive multi-threading.
I'd like to repeat the calculations for 10 layers and 50 layers. I'd like to check my work. I've already spent about 2 hours on this reply, so I'm giving up. Have fun.
In my estimate, the speed improvement in AI chips is going to see the same slowdown we've already seen in CPUs: single threaded performance is almost at a standstill and multi-threaded performance is increasing much less rapidly than it used to. If this slowdown occurs, there will be less pressure to replace existing AI machines with faster AI machines. This means a longer life cycle for existing machines.
Whether new facilities continue to be built will depend upon the degree to which AI is useful, and whether AI's usefulness requires more hardware. Nobody really knows.
Show me the how you can create a system where the price totals of all possible combinations of inventory selections result in only (3 or 4) mod 5.
The historical success rate of students from various schools and their GPAs is available to the colleges. Colleges that act wisely can adjust the GPAs reported by various grade schools by the school's historical reputation.
Do colleges still require a pre-acceptance interview? That should weed out many dullards and ignoramuses.
Some grade schools, even public schools, teach and test in a manner consistent with SAT preparation. No special paid SAT prep classes required.
Intelligence helps a person be financially successful. Successful people can provide at least two advantages for their children: better genes for intelligence and money for SAT prep classes. Without lots of data and good statistical analysis, the relative influence of those two (and other) factors cannot be stated with certainty. (Other factors include tendencies toward good study habits, good manners, and good nutrition.)
It's also worth noting that there's some similarity between SAT tests and tests in college, so SAT results correlate with college success. Thus SAT tests are partially predictive of college success.
So back then, prices were incremented by more than today's quarter.
People need to consider: Rounding to a nickle isn't going to be greater than 2 cents more inaccurate than rounding to pennies. Let's say you live in a backwater state, and still only make $7.25 per hour. Each transaction could potentially cost you at most 10 seconds of extra wages. However, transactions randomly round up and down, so the average error gets reduced by the square root of the number of transactions you make. Statistically speaking, you'll gain or lose only a couple of seconds of your time per purchase. Probably less time than it took to fumble for all those pennies.
But it sucks to be poor. Without pennies, someone who makes $50k per year will gain or lose only milliseconds worth of salary per transaction on average.
"But the stores will set prices so that it always rounds up!!!!1!" -- That only works for one item at most. Savvy shoppers would strategically buy combinations of items that always round down.
In a few years, all of these GPUs will be available on eBay for a few bucks each.
Then I'll finally be able to snag a whole bunch of them and build a Beowulf cluster to run SETI@home faster than anybody else.
If we were to get vehicles at near China's prices its hard to argue that demand for evs wouldn't improve.
Not necessarily.....most of the folks that want and EV, have one.....there just is NOT the demand for them here in the US that you have in other parts of the world.
A lot of this is due to the recharging infrastructure not being in place unless you live at the extreme west and maybe the east coast too.
I live in the New Orleans area....and from the maps and charging station finders I've seen we Still have precious few public charging stations anywhere in this area....
This is typical for most of the US.
With that comes range anxiety, and there's a TON of people, about 1/3 of the nation's populace that can't charge at home due to being in apartment complexes with large parking nots and no chargers or renting homes without chargers out side or no off street parking.
Unless you own your home and can charge at home, it's just a PITA to deal with and EV over here for a significant % of the populace.
I don't want one.....wouldn't work for me.
Never mind that their FSD is more capable than any current system on the market today.
What is "FSD"? New term to me....
Because someone still has to take time to read the slop. Over and over.
That work sounds like a great candidate to offload onto AI!
OMG, I'm so stealing that
There's no sense in being precise when you don't even know what you're talking about. -- John von Neumann