Question XOR and other bitwise operations in new CPUS

R

r_scooter

Junior Member

Aug 8, 2023

11

1

41

• Nov 29, 2024

• #1

Does anyone know the power requirements of different CPU instruction sets well enough to help me decide how to test the voltage that is pumped out of my power supply during different programming operations?

I have a multimeter but a software-based app would be a nice-to-have.

 

Last edited: Nov 29, 2024

fkoehler

Senior member

Feb 29, 2008

214

175

116

• Nov 29, 2024

• #2

There are known power/time attacks that can be done, however IIRC its commonly done a lot closer to the proc, probably from the Gnd/Power pins.
Even there, power/timing attacks are going to be flooded with everything else requiring power/signals to discern some ALU traffic.
I'd bet money a simple MM isn't going to provide anything of value, even the best Keysight scope without a lot of intrusive engineering, bare metal programming, and massive data acquisition and effort to find the signal to the noise.

 

Z

zir_blazer

Golden Member

Jun 6, 2013

1,206

502

136

• Nov 30, 2024

• #3

I believe that your idea is... very badly thought. What on earth are you trying to achieve?
To remove noise in results, you most likely would want to have a functioning Processor executing NO code at all and still have a way to feed it instructions so that you can measure your thing. Obviously, this is most likely impossible. Closest thing would be to use Linux with isolated CPUs as Kernel Parameter so that the CPU Scheduler doesn't uses some Cores unless being instructed to do so, then manually run your test application to load those Cores. You could Software measure power consumption with available tools. But to observe Voltage from the Power Supply? What for?

 

fkoehler

Senior member

Feb 29, 2008

214

175

116

• Dec 3, 2024

• #4

You didn't bother to reply, so...

If someone wanted to do this, I'd start by not trying to re-engerd stuff and try to find public source architecture documents from the manf's themselves.
A lot of this stuff is going to come down to process/lithography, pico-jule per bit, etc, etc, and is often bragged about at conferences like Hot Chips, etc.
Power per operation are a thing in supercomputer research and development where they are always looking to max out math performance at scale.
If you're just wondering if using SSE vs AVX is going to use more power, or be faster, much easier to just park all but one core and run a couple of programs bare metal is possible over a long enough time to measure watts used by the proc.
Again, not easy or probably possible if the first piece of test equipment that comes to mind is a DMM.

 

D

Doug S

Diamond Member

Feb 8, 2020

3,005

5,168

136

• Dec 3, 2024

• #5

Assuming you have a good, trusted way of measuring power on that scale (probably the on chip power monitoring stuff is as good as you can get, unless you have really good reason to believe you can do better) then you want to boot into as minimal an operating system as you can. Ideally you'd run something out of EFI, but failing that booting into single user mode on Linux is probably the best you can manage. Then run a program that consists of doing nothing but the instruction(s) you care about in an infinite loop and measure the power used while that is running, versus what is used while doing nothing and versus what is used while doing an infinite busy loop of NOOP.