[conspire] CPU power draw and heat

Nick Moffitt nick at zork.net
Sun Dec 9 16:02:17 PST 2018


On 09Dec2018 02:52PM (-0800), Rick Moen wrote:
> I'm actually pretty impressed that the CPU with the defunct fan never burned
> itself out.  I'll classify this as another indicator of just what an insane
> jump in power requirements arrived with the P4 and Xeon era, relative to the
> PIII.  I've dealt with RMAed P4 / Xeon pizza-box servers where the
> motherboard around one of the CPUs actually _charred_ when a CPU or its fan
> failed.

There is a YouTube series of interviews with some pretty big names in computing history (Brian Kernighan, David Brailsford, etc) known as "Computerphile".  They seem to have filmed entirely with hand-held camcorders, and spend far too much time zooming in and out, but the subject matter delivered by experts can be captivating.

One of the ones I often send people is this interview with Prof. Steve Furber, who worked at Acorn Computers on the development of the original ARM chip:

	https://youtu.be/1jOJl8gRPyQ?t=492

After the BBC Micro model B came out, they'd just about stretched the hardware to its limits.  The data bus on the BBC Micros was so oversubscribed that the most subtle changes in side effects of support chips would cause the things to cease to function.  

The folks at Acorn reasoned that the limiting factor for performance wasn't actually CPU speed or abilities, but (wait for it) memory performance.  They had just read some cool new papers from UC Berkeley about this RISC model of CPU design, and figured they could make it work.

The thing that drives ARM's popularity these days is largely the low power consumption, but they hadn't even considered it too hard at the time.  The only thing that mattered to them on that front was the packaging: if the chip ran below 1W, they could use an inexpensive plastic housing.  But if the chip ran over 1W, they'd need to use a ceramic housing because of the heat from all the switching activity (which scaled directly with transistor count, back then).

When the prototypes had been put through their paces, Furber describes hooking up an ammeter in series with the CPU's power supply to get a reading.

  It read 0.  
  
  No draw whatsoever on the power supply.

Of course he realised he'd fat-fingered one of the power connections, and the PSU circuit to the new ARM CPU hadn't been closed properly, so that explained the 0W reading.  But how did the chip manage to run through its tests without being plugged in?

The answer to that was simple: the signals coming into the CPU from the various support chips were enough to power the core.  After all, the measured CPU power draw was about 0.1W, so the memory bus and the other signal lines had enough margin to cover that.  But that was a surprising result for that generation of processor.

The MOS Technologies 6502 that powered the Apple ][, Commodore 64, BBC Micro, and a number of other popular micros of the late 70s and early 80s had 3,510 transistors on its die.  The first ARM that ran in the Acorn Archimedes had about 25,000 and it ran about 25 times the speed of the 6502 (again, solely because the memory bus could be clocked up by that much more).

These days even the "low power" ARM chips found in mobile phones have a few billion transistors, and we only get the low power draw because so much of that logic is shut off most of the time.  It also helps that phones these days are almost entirely batteries, by volume.

I ran K6-based servers for most of the early 2000s, largely for the benefit of a lower power draw.  The K6, according to Wikipedia, had 8,800,000 transistors, while the P4 family ramped up from 42 million to nearly 200 million in only a few years.  

Of course as I write this I am installing an i386 relase of ubuntu onto a salvaged Lenovo Ideapad S10-3, which has an Atom processor.  Not a bad chip, considering everything that Intel was going through at that time.  This system is largely intended as a development workstation to run the mu editor for my daughter's BBC micro:bits and maybe a bit of "PyGame Zero" work.

	https://codewith.mu/
	https://microbit.org/
	https://pygame-zero.readthedocs.io/

The micro:bit has an ARM Cortex M0, which is probably the closest spiritual cousin to that original ARM chip in the Archimedes.  The ARM folks claim it has "fewer than 12,000 gates" which is probably something on the order of 75,000 transistors.  I have some AA battery packs I charge these cards from, but you can run the thing off of a coin cell battery if you like.  

I'll tell you, though, they definitely run fanless!





More information about the conspire mailing list