4 ms·

Is there any type of calculations in which analog computer can outperform modern digital computer?

by vasili111 10y ago

quakeguy 10y ago
Infinite amount.
- edejong 10y ago
  Noisy comments here...
tlb 10y ago
A modern GPU (10 TFlops) can simulate a million nodes at 1 MHz, far more than any analog computer. The article was written in 2005, when that might still have been true.
nickpsecurity 10y ago
Analog computers perform a few, specific operations at full-speed that take quite a few components to do in digital with them only running at a specific, clock rate. Combining them with digital computers that feed them problems and measure the answers can be quite powerful. See especially the math-coprocessor at first link. The neural simulation at second is illustrative where one wafer (300+ chips) of analog NN's took 294,912 cores to simulate. That's full-custom, high-end digital stuff rather than the average stuff. Analog also gets used a lot in mixed-signal ASIC's where specific things are cheaper or lower power per unit in analog so they mix digital and analog. https://www.schneier.com/blog/archives/2015/07/friday_squid_bl_488.html#c6701962 https://www.schneier.com/blog/archives/2015/07/friday_squid_... https://news.ycombinator.com/item?id=11731697 https://news.ycombinator.com/item?id=11731697 EDIT to add: The paper below is a short history of them that gives examples whose benefits are consistent with other stuff I linked to. Some really cool stuff there, too. :) http://oro.open.ac.uk/5795/1/bletchley_paper.pdf http://oro.open.ac.uk/5795/1/bletchley_paper.pdf
- Animats 10y ago
  The question is the level of simulation. It's common to use a large number of computers to run a VLSI simulation before fabbing a chip. NVidia has been known to use 50 racks of 1U servers to run a gate-level simulation of a CPU. The machines are running C code generated from VHDL. You could do the job of the GPU with far less CPU power; a gate level simulation is executing multiple instructions for each bit transition in a gate.
  - nickpsecurity 10y ago
    The level of detail is part of it. The inherent speed of analog on a large range of numbers is the other part. I recall early efforts to model analog circuits like digital gates overwhelmed the computer. One circuit with 9 transistors took over 50,000 equations or something to check. It just did that much. This is only true when they're doing the things they directly implement like filters or differential equations. They run faster at fraction of size and power.
Quanticles 10y ago
Sure - http://mythic-ai.com http://mythic-ai.com
- Something1234 10y ago
  I call BS. I can't believe the claims that they are making about this chip. 50 times better battery life, and desktop level gpu in a button. I'm sure about that.
akamaka 10y ago
Yes, it's possible to do incredibly fast signal processing calculations using interfering laser beams. http://www.nextbigfuture.com/2014/08/startup-with-optical-solver.html?m=1 http://www.nextbigfuture.com/2014/08/startup-with-optical-so... In practice, it's not easy to compete with the economies of scale of conventional chips, though.

Is there any type of calculations in which analog computer can outperform modern digital computer?

quakeguy 10y ago

Infinite amount.

edejong 10y ago
Noisy comments here...

tlb 10y ago

A modern GPU (10 TFlops) can simulate a million nodes at 1 MHz, far more than any analog computer. The article was written in 2005, when that might still have been true.

nickpsecurity 10y ago

Analog computers perform a few, specific operations at full-speed that take quite a few components to do in digital with them only running at a specific, clock rate. Combining them with digital computers that feed them problems and measure the answers can be quite powerful. See especially the math-coprocessor at first link. The neural simulation at second is illustrative where one wafer (300+ chips) of analog NN's took 294,912 cores to simulate. That's full-custom, high-end digital stuff rather than the average stuff. Analog also gets used a lot in mixed-signal ASIC's where specific things are cheaper or lower power per unit in analog so they mix digital and analog. https://www.schneier.com/blog/archives/2015/07/friday_squid_bl_488.html#c6701962 https://www.schneier.com/blog/archives/2015/07/friday_squid_... https://news.ycombinator.com/item?id=11731697 https://news.ycombinator.com/item?id=11731697 EDIT to add: The paper below is a short history of them that gives examples whose benefits are consistent with other stuff I linked to. Some really cool stuff there, too. :) http://oro.open.ac.uk/5795/1/bletchley_paper.pdf http://oro.open.ac.uk/5795/1/bletchley_paper.pdf

Animats 10y ago
The question is the level of simulation. It's common to use a large number of computers to run a VLSI simulation before fabbing a chip. NVidia has been known to use 50 racks of 1U servers to run a gate-level simulation of a CPU. The machines are running C code generated from VHDL. You could do the job of the GPU with far less CPU power; a gate level simulation is executing multiple instructions for each bit transition in a gate.
- nickpsecurity 10y ago
  The level of detail is part of it. The inherent speed of analog on a large range of numbers is the other part. I recall early efforts to model analog circuits like digital gates overwhelmed the computer. One circuit with 9 transistors took over 50,000 equations or something to check. It just did that much. This is only true when they're doing the things they directly implement like filters or differential equations. They run faster at fraction of size and power.

Quanticles 10y ago

Sure - http://mythic-ai.com http://mythic-ai.com

Something1234 10y ago
I call BS. I can't believe the claims that they are making about this chip. 50 times better battery life, and desktop level gpu in a button. I'm sure about that.

akamaka 10y ago

Yes, it's possible to do incredibly fast signal processing calculations using interfering laser beams. http://www.nextbigfuture.com/2014/08/startup-with-optical-solver.html?m=1 http://www.nextbigfuture.com/2014/08/startup-with-optical-so... In practice, it's not easy to compete with the economies of scale of conventional chips, though.