Advantages of Parallel Hz

[krw]

They haven't, because it's wrong.

For cmos, P = f C V^2, which tells us that for N
processors running at f' = f/N the power will be
P = N f' C V^2 = N f/N C V^2 = f C V^2, which is
no improvement at all, even if N = 1 billion.

I'm not sure how this plays out when you throw in leakage (it's quite
significant). If 'f' can be kept to a minimum one can do all sorts
of things to reduce leakage. It doesn't solve the parallel
programming issue though.

 

[MooseFET]

They haven't, because it's wrong.

For cmos, P = f C V^2, which tells us that for N
processors running at f' = f/N the power will be
P = N f' C V^2 = N f/N C V^2 = f C V^2, which is
no improvement at all, even if N = 1 billion.

Actually it is P = f C V^2 + MoreStuff

With f=0, each gate loses a little power due to leakage. This means
that 1 processor doing 1 million instructions persecond will draw less
power than 1 million 1Hz processors.

 

[krw]

Actually it is P = f C V^2 + MoreStuff

With f=0, each gate loses a little power due to leakage. This means
that 1 processor doing 1 million instructions persecond will draw less
power than 1 million 1Hz processors.

Not necessarily. At 1Hz, one can lower the voltage pretty low and
build gates pretty thick. Both will minimize leakage.

 

[MooseFET]

Not necessarily. At 1Hz, one can lower the voltage pretty low and
build gates pretty thick. Both will minimize leakage.

Good point.

I think you run into a lower limit. Vcc must be more than Vth and the
OP wants to make a billion processors in a reasonable amount of
space. I'm fairly sure that something would set a lower limit on
leakage before you get down to 1Hz.

 

[krw]

Good point.

I think you run into a lower limit.

That limit is quite low, as evidenced by a digital watch.

Vcc must be more than Vth and the
OP wants to make a billion processors in a reasonable amount of
space.

Not necessarily. Sub-threshold operation is possible and Vth can be
very low. There is a lot of work in this area because leakage is
such a big problem.

I'm fairly sure that something would set a lower limit on
leakage before you get down to 1Hz.

Probably the wires from the battery. ;-)

However, 1Hz x 1Billion is an absurdity to show an example. How
about a MHz by a thousand? If the parallelism problem can be cracked
there are all sorts of games that can be played with all those free
transistors.

 

[MooseFET]

That limit is quite low, as evidenced by a digital watch.


Not necessarily. Sub-threshold operation is possible and Vth can be
very low. There is a lot of work in this area because leakage is
such a big problem.

I thought that when you made a low Vth, you always got a higher
leakage. It seems that this must be true if the GM is to remain
finite at the extreme of Vth=0.

Probably the wires from the battery. ;-)

It is likely the leakage in the supply capacitors would be more than
in the wires. Also if the supply has Schottky diodes in it, the
leakage in them will be high. If we are considering subvolt Vdd then
the recifier would likely be a MOSFET.

However, 1Hz x 1Billion is an absurdity to show an example.

I see it more as an extreme case to make the argument clearer, but yes
it is absurd.

How
about a MHz by a thousand? If the parallelism problem can be cracked
there are all sorts of games that can be played with all those free
transistors.

Parallelism has been cracked for special case problems. The very
large CPU time users seem also to be the places where the problem is
parallel in nature. Modeling the flow of fluids, explosions and the
propigation of waves through nonuniform mediums are the things that
come to mind quickly.

 

[Radium]

However, 1Hz x 1Billion is an absurdity to show an example. How
about a MHz by a thousand? If the parallelism problem can be cracked
there are all sorts of games that can be played with all those free
transistors.

Well, "parallel Hz" is meant for problems that are serial.

"Parallel Hz" actually doesn't have anything to do with whether the
task is parallelizable or not.

The parallelism you are describing has to do with the bits being
parallel [such as in a parallel printer].

"Parallel Hz" is a different story.

 

[Donald]

However, 1Hz x 1Billion is an absurdity to show an example. How
about a MHz by a thousand? If the parallelism problem can be cracked
there are all sorts of games that can be played with all those free
transistors.

Well, "parallel Hz" is meant for problems that are serial.

"Parallel Hz" actually doesn't have anything to do with whether the
task is parallelizable or not.

The parallelism you are describing has to do with the bits being
parallel [such as in a parallel printer].

"Parallel Hz" is a different story.

Still peddling your dream, eh.

 

[krw]

I thought that when you made a low Vth, you always got a higher
leakage. It seems that this must be true if the GM is to remain
finite at the extreme of Vth=0.

True. Physics can't be cheated (that easily). ;-) You can operate
in the sub-threshold region though.

It is likely the leakage in the supply capacitors would be more than
in the wires. Also if the supply has Schottky diodes in it, the
leakage in them will be high. If we are considering subvolt Vdd then
the recifier would likely be a MOSFET.


I see it more as an extreme case to make the argument clearer, but yes
it is absurd.

Diffeent words...

Parallelism has been cracked for special case problems. The very
large CPU time users seem also to be the places where the problem is
parallel in nature. Modeling the flow of fluids, explosions and the
propigation of waves through nonuniform mediums are the things that
come to mind quickly.

No parallelism hasn't been "cracked". Some problems are
"embarrassingly parallel".

 

[krw]

Well, "parallel Hz" is meant for problems that are serial.

Well, then you're in for a world of hurt. The two can't mix. How do
you propose to parallelize:

A=B+C
D=A+E

"Parallel Hz" actually doesn't have anything to do with whether the
task is parallelizable or not.

Then it's a meaningless word (no shock here).

The parallelism you are describing has to do with the bits being
parallel [such as in a parallel printer].

No, it has to do with tasks that can be done in parallel. That is,
they have no interdependence.

"Parallel Hz" is a different story.

No question. It looks like a ghost story.

 

[MooseFET]

True. Physics can't be cheated (that easily). ;-) You can operate
in the sub-threshold region though.




Diffeent words...



No parallelism hasn't been "cracked". Some problems are
"embarrassingly parallel".

They have to hand off the values to the neighboring processors as they
go. The process involves computing and then passing values around.

 

[MooseFET]

Well, "parallel Hz" is meant for problems that are serial.

"Parallel Hz" actually doesn't have anything to do with whether the
task is parallelizable or not.

As has been pointed out, yes it does have to do with the task being
parallel or not. Consider the sort of program where every step depends
on the result of the previous step. If the processor take one second
to do a step, the next step must wait that one second. Now consider
one where every step stands alone. All of the steps can be started as
soon as there is hardware free to do it so the result comes out in
much less time.

The parallelism you are describing has to do with the bits being
parallel [such as in a parallel printer].

No, the parallelism we are describing is a property of the problem not
the hardware. Consider these two tasks:

(1)
Replace all negative numbers with zero.

(2)
Replace the first negative number with zero.

The first sounds like more work but the second would actually take
longer. To do the second, you must test values in turn to find the
first negative value. In the first, you only need to look at a single
value in each CPU section so if you have enough sections, the whole
operation would happen in one stroke.

 

[krw]

They have to hand off the values to the neighboring processors as they
go. The process involves computing and then passing values around.

That is a naturally parallel problem.

 

[Fleetie]

[ Kilobytes of unnecessary quoted text deleted]

That is a naturally parallel problem.

FUCKING LEARN TO SNIP. PRAT.

 

[krw]

[ Kilobytes of unnecessary quoted text deleted]

That is a naturally parallel problem.

FUCKING LEARN TO SNIP. PRAT.

I do when it interests me, dilbert. Id didn't, and I didn't. Live
with it, or not.

 

[Guy Macon]

Radium wrote:
[SNIP]

Could you please give a straight answer to just one queustion?

Recently I suggested that you start working with real logic
chips and gave you links to a good protoboard. Pretty much
everyone here agrees that this would be a good learning
experience for you.

Why have you not done so?

 

[Bob Myers]

Well, "parallel Hz" is meant for problems that are serial.

More correctly, an imaginary "solution" for imagined
"problems."

Bob M.

 

[Radium]

My dream PC is as hardware, real-time, and digital as possible. In
addition, it uses the least amount of buffering required [hopefully
none] and experiences the least amount of latency possible [again,
hopefully none].

This may be your dream but it may also be a nightmare. You also said
you wanted low power and no fan. If you want a lot of speed, you
really want a good cooling system and a whole lot of power. If you
want to make a faster system with low power, you want to make use of
things like lookup tables and hashes.

Couldn't the problem of excessive heat and large use of power be
solved [or at least mitigated] by using lower voltages while still
running things in real-time [and with the least amount of storage,
software, buffering, and latency possible] and not using fans?

 

[MooseFET]

My dream PC is as hardware, real-time, and digital as possible. In
addition, it uses the least amount of buffering required [hopefully
none] and experiences the least amount of latency possible [again,
hopefully none].

This may be your dream but it may also be a nightmare. You also said
you wanted low power and no fan. If you want a lot of speed, you
really want a good cooling system and a whole lot of power. If you
want to make a faster system with low power, you want to make use of
things like lookup tables and hashes.

Couldn't the problem of excessive heat and large use of power be
solved [or at least mitigated] by using lower voltages while still
running things in real-time [and with the least amount of storage,
software, buffering, and latency possible] and not using fans?

No. It largely can't be avoided. If you have to do your sine function
from first principles and you want speed, you need a huge number of
operations per second.

 

[[email protected]]

My dream PC is as hardware, real-time, and digital as possible. In
addition, it uses the least amount of buffering required [hopefully
none] and experiences the least amount of latency possible [again,
hopefully none].

This may be your dream but it may also be a nightmare. You also said
you wanted low power and no fan. If you want a lot of speed, you
really want a good cooling system and a whole lot of power. If you
want to make a faster system with low power, you want to make use of
things like lookup tables and hashes.

Couldn't the problem of excessive heat and large use of power be
solved [or at least mitigated] by using lower voltages while still
running things in real-time [and with the least amount of storage,
software, buffering, and latency possible] and not using fans?

I'm just musing, please correct where I'm right off track.

It's current that causes heat rise (I^2R). For things to work, current
must flow, and for current to flow, voltage must be applied. For
minaturisation, smaller currents will need higher voltages.
For things to go faster, larger currents are required, and so larger
voltages are required to push these currents. If you limit to non-fan
cooling, you limit current, and therefore limit speed and capacity.

Be gentle with me, I'm trying to learn. jack