[Tfug] "Downgrading" ("underclocking?") processors

[Bexley Hall]

Hi Zack,

On 2/18/2014 9:06 PM, Zack Williams wrote:
> On Tue, Feb 18, 2014 at 8:03 PM, Bexley Hall<bexley401 at yahoo.com>  wrote:
>> Would you use this technique on an *enterprise* server?
>
> Nope.  On a HTPC or similar, sure.

For "multimedia", you typically don't care about transient anomalies
(as long as they truly are transient and don't gain persistence by,
for example, corrupting the program store).  Those are (largely)
"attended" operations -- an annoyed user can reset the device,
power it off, return it to the vendor, etc. to compensate for any
"misbehavior".

[I wonder how many "current generation" TV's could stay "up" for
months at a time?]

OTOH, users are not as forgiving/accommodating of all appliances!
Would you be keen on a garage door that opened itself sporadically
(e.g., while you were away from home)?  Or, an HVAC system that
decided to apply heat during the coolling season ("Hmmm, it's STILL
too hot in here!  Better switch the furnace to HIGH!")?  Or, an
emergency response system that autodialed 311 in an emergency?

[Hence my choice of "enterprise server" for an example -- presumably
the stakes are higher with an enterprise application]

> Most CPU's these days have various power modes and can control their
> supply voltage.   To save on power, they'll try to spend most of the
> time asleep or in similar low-power states with portions of the chip
> turned off.

There's nothing new, here.  Actually, it was the PC world that
started *locking* operating conditions to particular frequencies,
etc. (some of which an attempt to thwart commercial and hobbyist
overclockers!).

I designed a slot machine ~20 years ago that dynamically controlled
it's clock frequency (inserting a VCA in the oscillator and driving
that from a DAC output).  I use the same technique, now, to control
the clock frequency of my "network speakers" without having to add
"precision" (e.g., TCXO's) oscillators to each device.

[You want to ensure you are delivering audio "samples" to the DAC's
at a consistent rate derived from some *external* timebase which
can appear to vary with respect to your local timebase (because
of environmental differences between "there" and "here").  This isn't
possible with the programmable dividers present in most MCU/SoC's.
Even if you could reprogram the divisors at the sampling rate, you
introduce a lot of "jerk" into the system -- phase noise, etc.]

> They also will have varying speeds depending on the
> power/heat dissipation Intel, for example, can sell a chip that
> supports 7W, 10W and 13W power modes, with the vendor determining
> which they want to use:

I'm not worrying about "designed from scratch".  There, I can look
at my power and thermal budgets and "optimize to fit".  (though
most of my hardware is ARM-based and an order of magnitude less
power -- think hundreds of milliwatts.  Would you put a 40W PC
in a box to generate sounds for your doorbell?  :> )

Rather, what I am looking at is ways I can guide others to selecting
suitable COTS kit that *they* can "downgrade" easily.  You probably
*don't* want a noisey, hot, power hungry PC sitting in the back of your
closet running 24/7/365 *hoping* the fan never gums up, a pair of slacks
never falls across the air intake grill, etc. -- let alone the cost
of running it).

If, OTOH, you can replace the processor (without causing the rest of
the design to fail) to eliminate the fan AND conserve power, then you're
ahead of that game.

> What you might look at are the Lower power SKU's - for example, Intel
> will sell a desktop -T suffixed chip:

> or a laptop -U suffixed chip:

> Or a xeon -L suffixed chips:

> which have a much lower wattage than the other chips in the series.

Yes, but this is largely hit or miss when it comes to a "random"
PC.  I.e., I would have to tell folks "find a PC that uses one of
these power hogs"; then, downgrade it using these options...

It seems that most PC's (desktops/laptops/servers) try to maximize
performance (at the expense of power).  Even laptops *will* run
hot when taxed (instead of risking being considered "slow" in
the marketplace).  The remedy is almost always a fan (or, burn
even *more* power in a Peltier cooler, etc.)

> Given a suitably massive cooler, passively dissipating a 30-45W load
> shouldn't be difficult.

Far better not to *have* the 40W to dissipate!  :>  E.g., I try to
avoid having all that heat in one place by distributing the computation
physically.  So, dissipate a few watts in 100 locations instead of
hundreds of watts in *one* location.

But, there are some services that I can't provide without lots of
MIPS.  COTS hardware is the logical choice, there -- with the
power/noise caveat stipulated above.

(E.g., the SBC's that I use in my prototype are in the 20-40W
range FOR THE ENTIRE BOARD.  Even putting 10 of them together
lets me dissipate all that heat over a pretty large volume
without having to resort to active cooling.  If I could remove
the unnecessary hardware -- e.g., video, audio, etc. -- I could
probably trim that even more.)