Slashdot Mirror
Move Over Moore's Law, Make Way For Huang's Law (ieee.org)
Tekla Perry writes: Are graphics processors a law unto themselves? Nvidia's Jensen Huang says a 25-times speedup over five years is evidence that they are. He calls this the 'supercharged law,' and says it's time to start counting advances on multiple fronts, including architecture, interconnects, memory technology, and algorithms, not just circuits on a chip.
I think the attractive aspect of Moore's law was that it was simple and everyone got the general gist. Some people like to argue about the details but they mostly don't have anything else to do with their time.
I don't think we need a Huang's law. If you asks what Huang's Law is, everyone will just say it's like Moore's law except applied to GPU's.
Same applies to all those other people who want to name things after themselves.
>> t's time to start counting advances on multiple fronts
Hobbiests have for a while, ever since last-generation AMD stumbled and Intel slowed down the processor speed increases for a while. Now that Ryzen is out, graphics chips are ruling desktops, and no one cares about Intel in the mobile space, we're finally seeing progress get back toward Moore's Law's long-term trend line.
Moore's Law: the density of devices (transistors) that can be packed into a microchip doubles roughly every 18 months.
Or phrased more colloquially, if partially inaccurately:
Moore's Law: every 18 months, the speed of hardware doubles.
Gates' Law: every 18 months, the speed of software halves.
Let's not call any and all predictions laws. They're not laws. They are functions fit on a short stretch of data that have no predictive power in the future. Not even experts can predict the future, a "law" has no chance here.
The appropriate term now is GPGPU.
Everyone gets a law now
Yes, that's the well known "Anonymous Coward's Law".
While Moore's Law is a doubling in density every 18 months, this 25 times speedup in 5 years is achieved via a doubling every 13 months.
Moore's Law is an economic one, not a strictly a technological one. Although, keeping it going depends on semiconductor processes getting finer.
The costs per nanoacre more-or-less follow a predictable curve relative to how bleeding-edge a process is required to fabricate a chip. If you need a really old process, availability will be low, so demand will push the costs up. If you're using the latest, availability is low and yields will initially be low, so the costs are way up. Everything in-between is pretty cheap because demand tends to go towards newer stuff, and the fabrication plant for those middle-aged processes has already started depreciating-out.
If you're trying to be an industry leader, you're targeting the newest processes, so you have the highest expenses. As a result, you want to keep your chip small because costs scale with semiconductor area (N chips per wafer, X dollars per wafer, etc.).
However, there's a lower bound on that. If you've got a chip with over a thousand pins, like a high-end microprocessor, the chip has to be physically large enough to have that many bumps to wire out to pins. Also, there's the case of heat dissipation to consider. A teeny-tiny part that draws a lot of current may shatter or desolder itself if there's not enough surface area to mate with the thermal solution.
So, you want to be small, but there's a lower bound to that. That implies wasted space on the die, which you're going to pay for, anyway. What to do with it?
Add a core! Make the pipelines deeper! More cache! Add some multimedia accelerators. Add an FPGA! Add some other dedicated-function unit! Then present it as a bullet-point for selling more of your part versus your competitor.
Ergo, Moore's Law.
Pining for the days when The Glorious MEEPT!!! graced SlapDash with his wisdom.