Exabyte SSD in a 2.5" form factor is on the horizon.
Besides being large enough to fully populate a 48-bit address space,
they have 10 microsecond access times.

Currently it takes two 100 TB drives, which allows a fully populated system in 12+ cubic inches. Not cheap. An embedded form factor none the less.

If one equates one human synapse to one byte, at between 10^14 and 10^15 synapses, the above 256TB "embedded system" matches the capacity of the human brain.

Still amazed that a 4GB micro SD card an a pre Covid-19 Raspberry Pi make for a fully populated 32-bit address space and 32-bit computer for under $10.. Roughly one cubic inch. And have no idea what takes that much code, of course video easily takes that kind of space.

On 5/11/2023 9:50 PM, JimBrakefield wrote:
> Exabyte SSD in a 2.5" form factor is on the horizon.
> Besides being large enough to fully populate a 48-bit address space,
> they have 10 microsecond access times.
>
> Currently it takes two 100 TB drives, which allows a fully populated system in 12+ cubic inches. Not cheap. An embedded form factor none the less.
>
> If one equates one human synapse to one byte, at between 10^14 and 10^15 synapses, the above 256TB "embedded system" matches the capacity of the human brain.
>

Absent a significant increase in density (or a much cheaper substrate
and processes), this much storage is likely to remain mostly
unaffordable for the foreseeable future (with huge SSDs being the
exception rather than the rule).

Granted, maybe I am wrong and Moore's law will keep marching along; and
my general predictions about the future are wrong.

Granted, some of my recent sci-fi stories do include humanoid robots and
sentient AI, but as I see it, these should still be possible even
without much increase in process density (and I am mostly imagining with
levels of computational power not too far off from modern desktop PCs).

Just, sort of like if one can shove the power of a modern high-end
desktop PC into a humanoid form-factor and then make it all run with an
average power use of under 200W...

Also story assumes that they are using in-order CPUs, which makes this
part easier (one does need more cores though to compensate for worse
per-core performance); in story, they were described as essentially
running an RV64 variant with VLIW extensions (running on an OS "somewhat
resembling" a Linux distro). (Basically, like if one had an RV64 based
processor with Threadripper like stats ...).

Well, and an amount of storage that would be "fairly expensive" for
current SSDs (in story, the SSDs were imagined as having sizes more like
some of the larger "WD Red" drives; in a robot described as having been
built roughly 30 years in the future).

Though, in story, these older style computers are basically high-value
on the black market, as the idea is that future government makes it
illegal to have much more than a "kinda meh" modern PC (say, 70 years in
the future it being technically illegal to put more than 32GB of RAM in
a PC; or to have more than 8TB of total storage. With the ability to
compile software requiring a special license, side-loading software
being illegal, etc...).

But, then say, the black-market people are selling PCs with things like
64GB or 128GB of RAM, a compiler toolchain, a "jail broken" OS, ...

Granted, all this specifically excludes synapse-level simulation of a
human brain (assumed to be "basically impossible"), which is pretty far
outside what is feasible on a PC. But, for a hybrid of neural nets and
conventional algorithms, it seems like some form of "sentient" AI could
very well still be possible.

*1: Though, the definitions I am using are more in line with "Artificial
General Intelligence" + Self-Awareness; as opposed to the ability to
have subjective emotional experiences (but, this seems closer to
"sentience" in common use; presumably most people are not as concerned
with the "realness" of the subjective experiences of an AI... Vs its
ability to carry out tasks and decide for itself what actions to take, etc).

Though, it is likely some facsimile of this more subjective form of
sentience would emerge as well, but whether or not it would be
"equivalent" to that of humans or animals would be debatable; and it is
more of an open question of the likelihood of them developing a
self-preservation motive. However, any capable self-replication and
evolving would likely develop this as well; as survivor bias is likely
to favor those which act towards self preservation.

Depending on the type of training strategy, the development of
sentient-like behavior may be more or less likely (for example, I would
suspect AIs trained in competition-like settings using genetic
algorithms are more likely to gain sentient-like behaviors than those
trained to pass specific tests using back propagation and similar).

Though not expanded on in story, the idea was that the partial training
process for many of the AIs involved being put into various scenarios
(with some amount of randomization as to the details), and needing to
complete various objectives, using a genetic algorithm approach. With
each component and subsystem itself being trained with genetic algorithms.

Well, other than this, the assumption is that the robots in story are
mostly made out of things like 7075 aluminum and similar (with a lot of
cables and compact linear actuators). Mostly with an outer body made out
of materials like closed-cell polyurethane foam and silicone rubber.

Idea being the robot would probably be expensive enough "in general" to
justify the cost of 7075 (vs 6061), for the better strength-to-weight
ratio. Similarly, a steel skeleton would likely make it too heavy. Well,
can also debate between nylon and aramid for the internal tension cables
(aramid being stronger but a lot more expensive; the nylon likely being
in a similar form to heavy-duty fishing string), etc.

....

> Still amazed that a 4GB micro SD card an a pre Covid-19 Raspberry Pi make for a fully populated 32-bit address space and 32-bit computer for under $10. Roughly one cubic inch. And have no idea what takes that much code, of course video easily takes that kind of space.

? Including the case, the RasPi is well over a cubic inch. Granted,
maybe 1 in^3 is more if one excludes all the volume taken up by air and
similar.

For my project, I am using 16GB MicroSD cards, but ironically this is
because 16GB was the cheapest readily available option at the time
(IIRC, I had also bought a 64GB SDcard around the same time, mostly for
additional storage in my phone).

Now there are 256GB and 512GB SDcards for fairly reasonable prices, so
technology has seemingly still moved forwards in the past few years.

Otherwise:
Kinda funny, I have my laptops, and one of them has made it 20 years and
the batter still works... (Seemingly one of the early LiON packs; I
guess from not long after the move away from NiMH and NiCd packs).

Meanwhile, my phone is much newer and I already needed to replace the
battery because it puffed up (and newer laptops with non-removable
batteries also suffering from battery-puffing issues).

Seemingly, new tech isn't always better...

Otherwise, was working some on adding support for the newer "XG2RV" mode
to BGBCC, where:
XG2RV is effectively running a variant of my BJX2 XG2 ISA;
But, using the RISC-V RV64 mode's register scheme and RV64's C ABI.

Adding it to the CPU core and Emulator was easy enough.
But, working a new ABI into BGBCC is less easy.

Have ended up rewriting a big chunk of code related to handling of
argument passing to allow it to try to gloss over the ABI differences.

Probably not going to bother with the differences in struct-passing for
now (BGBCC uses "pass by reference" whereas the RV64 ABI is "pass by
copy on stack").

Maybe try to see if I can get around to trying to make basic RV64
programs work on top of BJX2.

Note that the purpose of XG2RV is to have a variant of BJX2 that can be
called more directly from RISC-V code without needing a bunch of awkward
thunks or similar (but, will still require appropriate tagging on
function pointers).

On Thursday, May 11, 2023 at 9:50:18 PM UTC-5, JimBrakefield wrote:
> Exabyte SSD in a 2.5" form factor is on the horizon.
> Besides being large enough to fully populate a 48-bit address space,
<
This will prove to be a reason to have larger address spaces.
<
> they have 10 microsecond access times.
>
> Currently it takes two 100 TB drives, which allows a fully populated system in 12+ cubic inches. Not cheap. An embedded form factor none the less.
>
> If one equates one human synapse to one byte, at between 10^14 and 10^15 synapses, the above 256TB "embedded system" matches the capacity of the human brain.
>
> Still amazed that a 4GB micro SD card an a pre Covid-19 Raspberry Pi make for a fully populated 32-bit address space and 32-bit computer for under $10. Roughly one cubic inch. And have no idea what takes that much code, of course video easily takes that kind of space.

JimBrakefield <jim.brakefield@ieee.org> writes:
> If one equates one human synapse to one byte, at between 10^14 and 10^15 sy=
> napses, the above 256TB "embedded system" matches the capacity of the human=
> brain.

Did this research ever move further down the field?

https://www.newscientist.com/article/2288228-can-quantum-effects-in-the-brain-explain-consciousness/

Because it would calculations about the brain's computing scope harder to
quantify--if born out.

Andy Valencia
Home page: https://www.vsta.org/andy/
To contact me: https://www.vsta.org/contact/andy.html

MitchAlsup <MitchAlsup@aol.com> writes:
>On Thursday, May 11, 2023 at 9:50:18=E2=80=AFPM UTC-5, JimBrakefield wrote:
>> Exabyte SSD in a 2.5" form factor is on the horizon.=20
>> Besides being large enough to fully populate a 48-bit address space,=20
><
>This will prove to be a reason to have larger address spaces.

Which ARM64 anticipated with 52-bit physical and virtual spaces...

With optane gone, non-volatile main memory has lost steam, but
I can see it returning as faster and denser NVRAMs are developing.

Back to the days of powering on the PDP-8 and picking right up
where you left off, thanks to non-volatile magnetic core memory.

On Friday, May 12, 2023 at 12:56:32 PM UTC-5, Scott Lurndal wrote:
> MitchAlsup <Mitch...@aol.com> writes:
> >On Thursday, May 11, 2023 at 9:50:18=E2=80=AFPM UTC-5, JimBrakefield wrote:
> >> Exabyte SSD in a 2.5" form factor is on the horizon.=20
> >> Besides being large enough to fully populate a 48-bit address space,=20
> ><
> >This will prove to be a reason to have larger address spaces.
> Which ARM64 anticipated with 52-bit physical and virtual spaces...
<
A CLX linked server rack with 4 of those SSDs per motherboard
exceeds the ARM64 52-bit address space........
<
My point was that shooting less than 64-bit address spaces, looking
forward, is a losing proposition.
>
> With optane gone, non-volatile main memory has lost steam, but
> I can see it returning as faster and denser NVRAMs are developing.
>
> Back to the days of powering on the PDP-8 and picking right up
> where you left off, thanks to non-volatile magnetic core memory.

On 5/12/2023 10:08 AM, MitchAlsup wrote:
> On Thursday, May 11, 2023 at 9:50:18 PM UTC-5, JimBrakefield wrote:
>> Exabyte SSD in a 2.5" form factor is on the horizon.
>> Besides being large enough to fully populate a 48-bit address space,
> <
> This will prove to be a reason to have larger address spaces.

Why? This is peripheral storage, not, unless you are talking single
level storage as in the S/38, not directly addressable by the CPU. You
can already have lots more peripheral storage than can be addressed
within the CPUs address space, so how does the fact that it can be done
in less physical space and with fewer external devices make a difference?

--
- Stephen Fuld
(e-mail address disguised to prevent spam)

According to JimBrakefield <jim.brakefield@ieee.org>:
>Exabyte SSD in a 2.5" form factor is on the horizon.
>Besides being large enough to fully populate a 48-bit address space,
>they have 10 microsecond access times.
>
>Currently it takes two 100 TB drives,

Hm, is this a new sense of "two"? Ten 100 TB drives would do it, which
would currently cost $400,000, if you could get them which you
probably can't. Thirty two 32 TB drives would be about $200K. Access
time is 50us, not 10us, which is still pretty fast but not like a
cache.

I don't think that these are likely to be used as swap space any
time soon. So I agree that eventually we will want bigger physical
address spaces but it's not urgent.

--
Regards,
John Levine, johnl@taugh.com, Primary Perpetrator of "The Internet for Dummies",
Please consider the environment before reading this e-mail. https://jl.ly

On 5/12/2023 2:35 PM, Stephen Fuld wrote:
> On 5/12/2023 10:08 AM, MitchAlsup wrote:
>> On Thursday, May 11, 2023 at 9:50:18 PM UTC-5, JimBrakefield wrote:
>>> Exabyte SSD in a 2.5" form factor is on the horizon.
>>> Besides being large enough to fully populate a 48-bit address space,
>> <
>> This will prove to be a reason to have larger address spaces.
>
> Why?  This is peripheral storage, not, unless you are talking single
> level storage as in the S/38, not directly addressable by the CPU.  You
> can already have lots more peripheral storage than can be addressed
> within the CPUs address space, so how does the fact that it can be done
> in less physical space and with fewer external devices make a difference?
>

Agreed.

If your HDD or SSD is bigger than your address space, it doesn't really
matter all that much...

In my case, I figure a 48-bit address as probably going to be sufficient
for probably at least the next few decades for most purposes.

I had also noted recently that, as-is, getting 16TB of solid-state
storage (either via SDcards or SSDs) is currently in the area of $1k.

And, $16k before one matches the size of a 48-bit address space by these
means.

With the cheapest option at the moment seemingly being a brick of 256GB
SDcards, followed by 1TB M.2 SSDs.

For whatever reason, 2.5" SATA SSD's are worse in a GB/$ sense than M.2
SSDs.

Both 512GB SDcards and 2TB M.2 SSDs fall behind due to being over 2x the
cost as their smaller counterparts.

Granted, the cheaper option is still to throw HDDs at the problem.

Not sure how many SDcards one could fit into the same volume as a 3.5"
HDD. Probably a fair number of TB, main limiting factors likely being
cost and heat (even a single SDcard can put out a disturbing level of
heat; for a big brick of them, cooling is likely to be an issue).

....

On Friday, May 12, 2023 at 11:41:49 AM UTC-6, Andy Valencia wrote:

> Did this research ever move further down the field?
>
> https://www.newscientist.com/article/2288228-can-quantum-effects-in-the-brain-explain-consciousness/
>
> Because it would calculations about the brain's computing scope harder to
> quantify--if born out.

No. If it had, that would have been very big news.

John Savard

On Thursday, May 11, 2023 at 8:50:18 PM UTC-6, JimBrakefield wrote:
> Exabyte SSD in a 2.5" form factor is on the horizon.

It is? What news item did you see that led you to that
conclusion, since there seems to be some doubt here
as to whether that could be possible.

John Savard

On Friday, May 12, 2023 at 10:56:43 PM UTC-6, Quadibloc wrote:
> On Thursday, May 11, 2023 at 8:50:18 PM UTC-6, JimBrakefield wrote:
> > Exabyte SSD in a 2.5" form factor is on the horizon.

> It is? What news item did you see that led you to that
> conclusion, since there seems to be some doubt here
> as to whether that could be possible.

Ah. I found this article

https://www.techradar.com/news/at-100tb-the-worlds-biggest-ssd-gets-an-eye-watering-price-tag

about a 100 Terabyte solid state drive, which means that you're right,
and as it sells for a price of $40,000, the doubter who said that if
an exabyte SSD was available, it wouldn't have a reasonable price
tag for ordinary mortals... was also right.

John Savard

BGB <cr88192@gmail.com> writes:
>On 5/12/2023 2:35 PM, Stephen Fuld wrote:

>I had also noted recently that, as-is, getting 16TB of solid-state
>storage (either via SDcards or SSDs) is currently in the area of $1k.
>
>And, $16k before one matches the size of a 48-bit address space by these
>means.

A 5MB winchester cost $11,200 back in the day.

HP's first winchester cost between $25k and 36K for 12MB.

$25,000 in 1979 dollars is over $111,000 today.

On 5/13/2023 12:53 PM, Scott Lurndal wrote:
> BGB <cr88192@gmail.com> writes:
>> On 5/12/2023 2:35 PM, Stephen Fuld wrote:
>
>> I had also noted recently that, as-is, getting 16TB of solid-state
>> storage (either via SDcards or SSDs) is currently in the area of $1k.
>>
>> And, $16k before one matches the size of a 48-bit address space by these
>> means.
>
> A 5MB winchester cost $11,200 back in the day.
>
> HP's first winchester cost between $25k and 36K for 12MB.
>
> $25,000 in 1979 dollars is over $111,000 today.
>

Prices went down since then...
But doesn't necessarily mean prices will continue to go down now.

Back then, people needed to advance their manufacturing technology.
Now, they are pushing against fundamental physical limits.

Manufacturing technology means it is more a matter of time and
investment. Whereas, physical limits means one is basically screwed, and
anything going forwards will fall steadily more into "diminishing
returns" territory.

Say, if you can make a 1PB SSD as nearly a solid brick of silicon dies,
well, this can be done. Doesn't mean that making a brick of silicon dies
is going to become more affordable...

I can note for example, some time ago (5 or 6 years maybe?) I got an
SSD, it was 1TB.

How big are modern SSDs? Typically also 1TB.

The first 1TB HDDs came out roughly a few years after I graduated
high-school. Where are we, nearly 2 decades later?... Mostly a 1..4TB,
8, 12, or 16TB if one feels like burning some money...

Pretty much my entire adult life thus far has been watching progress in
these areas slowly grinding to a halt.

This is partly why my "future predictions" seem kinda lame, it doesn't
take too much to imagine that progress pretty much entirely grinds to a
halt, or maybe even back-slides to some extent when people stop throwing
as much money at the problem.

People who are young now may not even really know what it is like to
have lived in a world where the expectation was that computers would
keep getting bigger and faster, whereas in my younger years, the
assumption that computers would keep getting bigger and faster was taken
for granted...

....

Granted, humans exist within the confines of physical limits, with
relatively modest power requirements, and without a "particularly dense"
physical substrate (most of the brain being "staying alive" stuff,
rather than "performing computations" stuff).

So, I will speculate that something expressing human-like abilities
should very well be possible, but one will (likely) need to give up the
idea of it being a direct simulation of a human-like brain inside a
computer...

But, it is also amusing to think we might already be past the levels of
computational power needed for human-like AIs, and just have not
realized it yet?... (Because the existing strategies are just so
horridly inefficient?...).

Granted, it would be nice to see computers having native support for
things like low precision floating-point SIMD, etc...

On Saturday, May 13, 2023 at 1:35:32 PM UTC-5, BGB wrote:
> On 5/13/2023 12:53 PM, Scott Lurndal wrote:
> > BGB <cr8...@gmail.com> writes:
> >> On 5/12/2023 2:35 PM, Stephen Fuld wrote:
> >
> >> I had also noted recently that, as-is, getting 16TB of solid-state
> >> storage (either via SDcards or SSDs) is currently in the area of $1k.
> >>
> >> And, $16k before one matches the size of a 48-bit address space by these
> >> means.
> >
> > A 5MB winchester cost $11,200 back in the day.
> >
> > HP's first winchester cost between $25k and 36K for 12MB.
> >
> > $25,000 in 1979 dollars is over $111,000 today.
> >
> Prices went down since then...
> But doesn't necessarily mean prices will continue to go down now.
>
> Back then, people needed to advance their manufacturing technology.
> Now, they are pushing against fundamental physical limits.
>
>
> Manufacturing technology means it is more a matter of time and
> investment. Whereas, physical limits means one is basically screwed, and
> anything going forwards will fall steadily more into "diminishing
> returns" territory.
>
>
> Say, if you can make a 1PB SSD as nearly a solid brick of silicon dies,
> well, this can be done. Doesn't mean that making a brick of silicon dies
> is going to become more affordable...
>
>
>
> I can note for example, some time ago (5 or 6 years maybe?) I got an
> SSD, it was 1TB.
>
> How big are modern SSDs? Typically also 1TB.
>
> The first 1TB HDDs came out roughly a few years after I graduated
> high-school. Where are we, nearly 2 decades later?... Mostly a 1..4TB,
> 8, 12, or 16TB if one feels like burning some money...
>
>
>
> Pretty much my entire adult life thus far has been watching progress in
> these areas slowly grinding to a halt.
>
> This is partly why my "future predictions" seem kinda lame, it doesn't
> take too much to imagine that progress pretty much entirely grinds to a
> halt, or maybe even back-slides to some extent when people stop throwing
> as much money at the problem.
>
>
> People who are young now may not even really know what it is like to
> have lived in a world where the expectation was that computers would
> keep getting bigger and faster, whereas in my younger years, the
> assumption that computers would keep getting bigger and faster was taken
> for granted...
>
> ...
>
>
> Granted, humans exist within the confines of physical limits, with
> relatively modest power requirements, and without a "particularly dense"
> physical substrate (most of the brain being "staying alive" stuff,
> rather than "performing computations" stuff).
>
> So, I will speculate that something expressing human-like abilities
> should very well be possible, but one will (likely) need to give up the
> idea of it being a direct simulation of a human-like brain inside a
> computer...
>
> But, it is also amusing to think we might already be past the levels of
> computational power needed for human-like AIs, and just have not
> realized it yet?... (Because the existing strategies are just so
> horridly inefficient?...).
>
> Granted, it would be nice to see computers having native support for
> things like low precision floating-point SIMD, etc...

|> it is also amusing to think we might already be past the levels of
|>computational power needed for human-like AIs, and just have not
|>realized it yet?
The contrast between software lines of code (can't imagine how to organize 4GB of code)
and what the brain holds? leads one to conclude that the brain builds and holds
functional models, say one million of them. This corresponds to the idea of cortical columns
as the organizational structure of the brain? One of the biggest problems in brain structure
is the missing "middle ware", eg the micro-architecture in between the large scale
organization (cortex, cerebellum, thalamus, etc) and the neuron level processing.

On 5/13/2023 3:08 PM, JimBrakefield wrote:
> On Saturday, May 13, 2023 at 1:35:32 PM UTC-5, BGB wrote:
>> On 5/13/2023 12:53 PM, Scott Lurndal wrote:
>>> BGB <cr8...@gmail.com> writes:
>>>> On 5/12/2023 2:35 PM, Stephen Fuld wrote:
>>>
>>>> I had also noted recently that, as-is, getting 16TB of solid-state
>>>> storage (either via SDcards or SSDs) is currently in the area of $1k.
>>>>
>>>> And, $16k before one matches the size of a 48-bit address space by these
>>>> means.
>>>
>>> A 5MB winchester cost $11,200 back in the day.
>>>
>>> HP's first winchester cost between $25k and 36K for 12MB.
>>>
>>> $25,000 in 1979 dollars is over $111,000 today.
>>>
>> Prices went down since then...
>> But doesn't necessarily mean prices will continue to go down now.
>>
>> Back then, people needed to advance their manufacturing technology.
>> Now, they are pushing against fundamental physical limits.
>>
>>
>> Manufacturing technology means it is more a matter of time and
>> investment. Whereas, physical limits means one is basically screwed, and
>> anything going forwards will fall steadily more into "diminishing
>> returns" territory.
>>
>>
>> Say, if you can make a 1PB SSD as nearly a solid brick of silicon dies,
>> well, this can be done. Doesn't mean that making a brick of silicon dies
>> is going to become more affordable...
>>
>>
>>
>> I can note for example, some time ago (5 or 6 years maybe?) I got an
>> SSD, it was 1TB.
>>
>> How big are modern SSDs? Typically also 1TB.
>>
>> The first 1TB HDDs came out roughly a few years after I graduated
>> high-school. Where are we, nearly 2 decades later?... Mostly a 1..4TB,
>> 8, 12, or 16TB if one feels like burning some money...
>>
>>
>>
>> Pretty much my entire adult life thus far has been watching progress in
>> these areas slowly grinding to a halt.
>>
>> This is partly why my "future predictions" seem kinda lame, it doesn't
>> take too much to imagine that progress pretty much entirely grinds to a
>> halt, or maybe even back-slides to some extent when people stop throwing
>> as much money at the problem.
>>
>>
>> People who are young now may not even really know what it is like to
>> have lived in a world where the expectation was that computers would
>> keep getting bigger and faster, whereas in my younger years, the
>> assumption that computers would keep getting bigger and faster was taken
>> for granted...
>>
>> ...
>>
>>
>> Granted, humans exist within the confines of physical limits, with
>> relatively modest power requirements, and without a "particularly dense"
>> physical substrate (most of the brain being "staying alive" stuff,
>> rather than "performing computations" stuff).
>>
>> So, I will speculate that something expressing human-like abilities
>> should very well be possible, but one will (likely) need to give up the
>> idea of it being a direct simulation of a human-like brain inside a
>> computer...
>>
>> But, it is also amusing to think we might already be past the levels of
>> computational power needed for human-like AIs, and just have not
>> realized it yet?... (Because the existing strategies are just so
>> horridly inefficient?...).
>>
>> Granted, it would be nice to see computers having native support for
>> things like low precision floating-point SIMD, etc...
>
> |> it is also amusing to think we might already be past the levels of
> |>computational power needed for human-like AIs, and just have not
> |>realized it yet?
> The contrast between software lines of code (can't imagine how to organize 4GB of code)
> and what the brain holds? leads one to conclude that the brain builds and holds
> functional models, say one million of them. This corresponds to the idea of cortical columns
> as the organizational structure of the brain? One of the biggest problems in brain structure
> is the missing "middle ware", eg the micro-architecture in between the large scale
> organization (cortex, cerebellum, thalamus, etc) and the neuron level processing.
>

Yeah, for something like a human-like mind, pure software isn't going to
work either. One needs a system capable of dynamic adaptation and
learning based on experiences, as opposed to purely static binaries.

On the other hand, trying to do everything using free-form neural nets
wouldn't really be all that practical in a computational sense.

One would likely need something where some parts (that don't really need
any learning ability) are "hard logic" (namely, a mix of conventional
algorithms and more fixed-form neural nets); and other parts are
dynamically adaptive nets will some sort of mimicry of organic
neuroplasticity (IIRC, not really a thing in conventional neural nets).

For example, in a human, much of the functionality of the occipital and
parietal lobe and similar could be dedicated to hard-logic, with mostly
the frontal and temporal lobe areas being flexible logic.

Similarly, not all neurons are uniform in such a system, as one may need
special types of neurons whose inputs or outputs connect directly to
hard-logic elements (sort of like a neural-net equivalent of how things
work in an FPGA).

As opposed to conventional "deep" or "convolutional" nets where the
overall structure is fairly uniform (one doesn't just randomly have some
neurons using a different activation function or plugging into external
logic, etc).

I have had some moderate success (on a small scale) building these sorts
of hybrid nets using genetic algorithms, but one big drawback is that
"experience-based learning" (as typically seen in humans and animals)
isn't really compatible with a genetic algorithm approach (which more
favors something like back-propagation and a reward/punishment system).

Though, possibly both approaches could be used, where the overall
architecture is built using a mix of conventional programming and
genetic algorithms, but then back-propagation is used to fine-tune the
parts of the nets which need a more "dynamic" structure.

Likely, also some amount of partitioning/segmentation of the net to try
to keep things more manageable, etc...

Though, the overall topology and functional areas could still be built
using the human brain as a design template, with various known non-local
signaling pathways being included in the design.

....

On Saturday, May 13, 2023 at 4:46:28 PM UTC-5, BGB wrote:
> On 5/13/2023 3:08 PM, JimBrakefield wrote:

> > The contrast between software lines of code (can't imagine how to organize 4GB of code)
> > and what the brain holds? leads one to conclude that the brain builds and holds
> > functional models, say one million of them. This corresponds to the idea of cortical columns
> > as the organizational structure of the brain? One of the biggest problems in brain structure
> > is the missing "middle ware", eg the micro-architecture in between the large scale
> > organization (cortex, cerebellum, thalamus, etc) and the neuron level processing.
> >
> Yeah, for something like a human-like mind, pure software isn't going to
> work either. One needs a system capable of dynamic adaptation and
> learning based on experiences, as opposed to purely static binaries.
<
LISP has those properties. It agglomerates new procedures as it goes along.
>
> On the other hand, trying to do everything using free-form neural nets
> wouldn't really be all that practical in a computational sense.
<
We are still looking for a road map--neural nets do some of this, but we are
still missing a big key.
>
> One would likely need something where some parts (that don't really need
> any learning ability) are "hard logic" (namely, a mix of conventional
> algorithms and more fixed-form neural nets); and other parts are
> dynamically adaptive nets will some sort of mimicry of organic
> neuroplasticity (IIRC, not really a thing in conventional neural nets).
<
But those hard parts are simply what/how the brains learns from 0 years old
to 4 years old. {language, colors, fine motor skills, societal norms,...}.
>
>
> For example, in a human, much of the functionality of the occipital and
> parietal lobe and similar could be dedicated to hard-logic, with mostly
> the frontal and temporal lobe areas being flexible logic.
>
>
> Similarly, not all neurons are uniform in such a system, as one may need
> special types of neurons whose inputs or outputs connect directly to
> hard-logic elements (sort of like a neural-net equivalent of how things
> work in an FPGA).
>
> As opposed to conventional "deep" or "convolutional" nets where the
> overall structure is fairly uniform (one doesn't just randomly have some
> neurons using a different activation function or plugging into external
> logic, etc).
>
Deep neural nets work for some things, but we remain missing a key
{step, insight, parameter, organization.}
>
> I have had some moderate success (on a small scale) building these sorts
> of hybrid nets using genetic algorithms, but one big drawback is that
> "experience-based learning" (as typically seen in humans and animals)
> isn't really compatible with a genetic algorithm approach (which more
> favors something like back-propagation and a reward/punishment system).
>
> Though, possibly both approaches could be used, where the overall
> architecture is built using a mix of conventional programming and
> genetic algorithms, but then back-propagation is used to fine-tune the
> parts of the nets which need a more "dynamic" structure.
>
>
> Likely, also some amount of partitioning/segmentation of the net to try
> to keep things more manageable, etc...
>
> Though, the overall topology and functional areas could still be built
> using the human brain as a design template, with various known non-local
> signaling pathways being included in the design.
>
> ...

On 5/13/2023 5:13 PM, MitchAlsup wrote:
> On Saturday, May 13, 2023 at 4:46:28 PM UTC-5, BGB wrote:
>> On 5/13/2023 3:08 PM, JimBrakefield wrote:
>
>>> The contrast between software lines of code (can't imagine how to organize 4GB of code)
>>> and what the brain holds? leads one to conclude that the brain builds and holds
>>> functional models, say one million of them. This corresponds to the idea of cortical columns
>>> as the organizational structure of the brain? One of the biggest problems in brain structure
>>> is the missing "middle ware", eg the micro-architecture in between the large scale
>>> organization (cortex, cerebellum, thalamus, etc) and the neuron level processing.
>>>
>> Yeah, for something like a human-like mind, pure software isn't going to
>> work either. One needs a system capable of dynamic adaptation and
>> learning based on experiences, as opposed to purely static binaries.
> <
> LISP has those properties. It agglomerates new procedures as it goes along.
>>
>> On the other hand, trying to do everything using free-form neural nets
>> wouldn't really be all that practical in a computational sense.
> <
> We are still looking for a road map--neural nets do some of this, but we are
> still missing a big key.

Yeah.

There was also some hype a few years back about "memristors" and their
application to neural net hardware. I was skeptical as this approach (as
I had seen it presented) is ultimately non scalable.

One would likely be "better" served trying to figure out how to
integrate large numbers of low-power vector-processors with flash-memory
chips (sort of like many-core processors integrated into an SDcard or
similar; holding the general logic for neural-net tasks in firmware with
each core able to make repeating sweeps over designated areas of the
Flash storage).

Or, say, some sort of "Analog FGMOS", where the state is used for
holding an N-bit analog value, rather than trying to store an exact bit
pattern (would likely get longer cycle-life out of the FGMOS cells as
well, since ability to exactly preserve values is less important in this
case).

....

>>
>> One would likely need something where some parts (that don't really need
>> any learning ability) are "hard logic" (namely, a mix of conventional
>> algorithms and more fixed-form neural nets); and other parts are
>> dynamically adaptive nets will some sort of mimicry of organic
>> neuroplasticity (IIRC, not really a thing in conventional neural nets).
> <
> But those hard parts are simply what/how the brains learns from 0 years old
> to 4 years old. {language, colors, fine motor skills, societal norms,...}.

One doesn't necessarily need the AI to be able to learn these things, if
a lot of the parts can be reduced to hard-coded elements.

If the vast majority of low-level sensory processing and physical
control parts can be eliminated from the neural net, this reduces
computational cost and leaves more processing power available for the
"cognitive" parts.

A human would learn these things, but the amount of neurons and
connection density in the human brain would also be impractical for
anything along similar lines of a modern PC.

Sort of like "AI upsampling" vs a bicubic filter:
AI upsampling, using neural nets, eats a GPU;
Bicubuc is fast enough for real-time video processing;
Visually, there isn't enough difference to matter.

If one were designing something, it would be more efficient to plug the
output of a bicubic filter as an input into the net, than have the net
figure out itself how to do this.

Or, noting the similarities, parts of the visual cortex functionality
could be replaced by a loop performing the discrete wavelet transform,
noting how it is much cheaper to run the DWT on an image, than it is to
run this logic in a neural net.

Then have virtual neurons whose sole purpose is to represent output
coefficients from the DWT filtered image, etc.

Though, it is a question of how much can be stripped away while still
being able to give something "human-like".

Or, if the goal is to be human-like in essence, or merely in terms of
some basic superficial aspects (so, more like a glorified video-game
character than a person).

This latter case is likely even easier (basically just computer vision,
"mostly sufficient" levels of articulation and motor control, things
like path-finding algorithms and similar, and some state machines and
Markov models and similar). This would be more a coding challenge than a
computational-limits one.

But, I am assuming something a bit beyond this...

>>
>>
>> For example, in a human, much of the functionality of the occipital and
>> parietal lobe and similar could be dedicated to hard-logic, with mostly
>> the frontal and temporal lobe areas being flexible logic.
>>
>>
>> Similarly, not all neurons are uniform in such a system, as one may need
>> special types of neurons whose inputs or outputs connect directly to
>> hard-logic elements (sort of like a neural-net equivalent of how things
>> work in an FPGA).
>>
>> As opposed to conventional "deep" or "convolutional" nets where the
>> overall structure is fairly uniform (one doesn't just randomly have some
>> neurons using a different activation function or plugging into external
>> logic, etc).
>>
> Deep neural nets work for some things, but we remain missing a key
> {step, insight, parameter, organization.}

I am not claiming to have figured it out.

But, rather, that it is possible that something "may" be possible within
the limits of the currently available computational power.

Not that it will necessarily be all that easy to pull off...

>>
>> I have had some moderate success (on a small scale) building these sorts
>> of hybrid nets using genetic algorithms, but one big drawback is that
>> "experience-based learning" (as typically seen in humans and animals)
>> isn't really compatible with a genetic algorithm approach (which more
>> favors something like back-propagation and a reward/punishment system).
>>
>> Though, possibly both approaches could be used, where the overall
>> architecture is built using a mix of conventional programming and
>> genetic algorithms, but then back-propagation is used to fine-tune the
>> parts of the nets which need a more "dynamic" structure.
>>
>>
>> Likely, also some amount of partitioning/segmentation of the net to try
>> to keep things more manageable, etc...
>>
>> Though, the overall topology and functional areas could still be built
>> using the human brain as a design template, with various known non-local
>> signaling pathways being included in the design.
>>
>> ...

On Sunday, 14 May 2023 at 00:38:08 UTC+1, BGB wrote:
> On 5/13/2023 5:13 PM, MitchAlsup wrote:
> > On Saturday, May 13, 2023 at 4:46:28 PM UTC-5, BGB wrote:
> >> On 5/13/2023 3:08 PM, JimBrakefield wrote:
> >
> >>> The contrast between software lines of code (can't imagine how to organize 4GB of code)
> >>> and what the brain holds? leads one to conclude that the brain builds and holds
> >>> functional models, say one million of them. This corresponds to the idea of cortical columns
> >>> as the organizational structure of the brain? One of the biggest problems in brain structure
> >>> is the missing "middle ware", eg the micro-architecture in between the large scale
> >>> organization (cortex, cerebellum, thalamus, etc) and the neuron level processing.
> >>>
> >> Yeah, for something like a human-like mind, pure software isn't going to
> >> work either. One needs a system capable of dynamic adaptation and
> >> learning based on experiences, as opposed to purely static binaries.
> > <
> > LISP has those properties. It agglomerates new procedures as it goes along.
> >>
> >> On the other hand, trying to do everything using free-form neural nets
> >> wouldn't really be all that practical in a computational sense.
> > <
> > We are still looking for a road map--neural nets do some of this, but we are
> > still missing a big key.
> Yeah.
>
>

There is a project at the University of Manchester under Prof. Steve Furber that is researching into computer simulations that mimic biological neural nets (as opposed to those used in machine learning) to try to further knowledge in this area: Project SpiNNaker. They use (or aim to?) a machine with a million ARM cores targeting biological neural nets at around the complexity of a mouse brain.

https://apt.cs.manchester.ac.uk/projects/SpiNNaker/

On Saturday, May 13, 2023 at 12:35:32 PM UTC-6, BGB wrote:

> People who are young now may not even really know what it is like to
> have lived in a world where the expectation was that computers would
> keep getting bigger and faster, whereas in my younger years, the
> assumption that computers would keep getting bigger and faster was taken
> for granted...

Arthur C. Clarke, near the beginning of his book _Profiles of the Future_, which
collected essays which originally appeared in a famous American magazine
which shall remain nameless, noted how the speed of transportation had been
rapidly increasing over the past several decades...

but the trend, from the railroad to the jet airplane to the spaceship, could not
continue, as soon we would be travelling faster than the speed of light.

He used that example to illustrate the phenomenon which we are running into
with computers - the "S-curve"; a new technology ushers in a phase where there
is rapid improvement, at an exponential rate, but that eventually comes to an
end, slowing down with diminishing returns.

A new invention, such as quantum computers, could usher in another phase
of rapid improvement.

John Savard

On 5/15/2023 6:03 PM, Quadibloc wrote:
> On Saturday, May 13, 2023 at 12:35:32 PM UTC-6, BGB wrote:
>
>> People who are young now may not even really know what it is like to
>> have lived in a world where the expectation was that computers would
>> keep getting bigger and faster, whereas in my younger years, the
>> assumption that computers would keep getting bigger and faster was taken
>> for granted...
>
> Arthur C. Clarke, near the beginning of his book _Profiles of the Future_, which
> collected essays which originally appeared in a famous American magazine
> which shall remain nameless, noted how the speed of transportation had been
> rapidly increasing over the past several decades...
>
> but the trend, from the railroad to the jet airplane to the spaceship, could not
> continue, as soon we would be travelling faster than the speed of light.
>

Yeah.

Physical limits are annoying.

For vehicles, there are many practical limits as to how quickly one can
make them go.

Cars: Ability to maintain stability and the response speed of the person
driving them, etc. Even if driven by computers, it is unlikely they
could remain stable at 200-300 mph (and even then, things like
encountering random road debris would likely end in disaster, etc).

Trains: Mechanical issues with the rails or debris/etc moving onto the
track. Having a maglev train inside of a tunnel with the tunnel held at
reduced pressure (with a helium atmosphere, *) could potentially allow
for absurdly fast trains.

*: Holding a tunnel at a full vacuum would be impractical, but 0.1 bar
of helium would have nearly the same effect. One might need multiple
layers of tunnel though (with a system of 1-way valves, such that sudden
pressure spikes in the inner tunnel can be deflected into the outer tunnel).

Seems like building a tunnel that doesn't crush under external air
pressure would itself be an engineering challenge.

Might be simpler/cheaper just to use 1 bar of helium (or hydrogen...).

For space, plasma thrusters can be useful, but one needs a good power
source (such as a nuclear reactor).

For interstellar travel, things are more of an issue...

I had imagined an idea resembling a hybrid of a plasma thruster,
cyclotron, and a linear particle accelerator. For initial acceleration,
a plasma thruster stage would be used (feeding through a linear
accelerator stage).

For higher speed operation, it would switch from using primarily plasma
ionization to primarily using the cyclotron (idea being to try to push
the exhaust velocity to up nearly light speed, *).

Though, less obvious is where to send the heat from the reactor at this
stage (if using a reactor for a plasma thruster, they could use the
propellant as a heat-sink for the reactor; for cyclotron operation the
amount of propellant used would likely be far less than the amount
needed to effectively cool a reactor).

*: As a ship approaches its exhaust speed, the relative thrust from the
exhaust approaches zero. If the ship were moving faster than its own
exhaust, then it would lose more velocity by firing its thrusters than
it would gain

Say, from an external observer, the exhaust would be chasing the ship.
If the ship is traveling at 0.3c and fires a thruster with a velocity of
0.2c, then the exhaust is still traveling in the same direction as the
ship at 0.1c, and the effect on the ship's total velocity would be
analogous to a thruster firing in the opposite direction at 0.1c,
pulling the ship to a point where its forwards velocity and exhaust
velocity add up to 0.

But, say, if the exhaust velocity is 0.999c or similar, then it can
still keep accelerating; even with only a relatively small number of
exhaust particles being ejected.

Put then one basically needs to use a particle accelerator or similar to
push each exhaust particle as fast as it can go (rather than it being
about the volume or power that one can push through the engine). Say,
the ship slowly transitioning from plasma to particle acceleration once
it starts crossing 0.1c or so...

But, say, the ship needs to get above roughly 0.9c before the effects of
time-dilation and similar become "useful".

Well, and then the ship would need to expend a similar amount of energy
decelerating before it reaches the target.

Say, a ship crashing into a planet at a good part of light-speed being
"not useful" (well, apart from possibly making a "world breaking"
explosion).

Though, for whatever reason, planetary assault with high velocity
projectiles isn't really a thing in sci-fi (usually all the emphasis is
on directed energy weapons or similar, rather than kinetic projectiles).

Or, say, space combat with people firing cannons (or railguns) at each
other and randomly changing velocity in an attempt to throw off the aim
of the enemy ships (since the travel-time of the projectile would become
a lot more significant in space-based combat).

> He used that example to illustrate the phenomenon which we are running into
> with computers - the "S-curve"; a new technology ushers in a phase where there
> is rapid improvement, at an exponential rate, but that eventually comes to an
> end, slowing down with diminishing returns.
>

Yep.

I suspect my childhood was in the sharply rising part of the S-curve,
but pretty much my entire adult life has been watching it slowly fizzle out.

There were occasional "hype" technologies:
Carbide semiconductors;
Now displaced by Gallium Nitride;
Graphene chips;
Photonics;
...

But, thus far, little has really amounted to much; each having severe
drawbacks which limit either to niche applications (such as
power-electronics and LEDs for Carbide and GaN) or seemingly kill it
entirely (general situation with graphene).

> A new invention, such as quantum computers, could usher in another phase
> of rapid improvement.
>

I have some skepticism at this point about the "general purpose
viability" of quantum computers.

But, as-is, we don't have anything that is really a viable (or even
"particularly promising") alternative to the existing CMOS technology,
which is concerning...

Sadly, beyond just "general stagnation", the existence of "high end"
chips will likely only continue so long as there is a market for them,
which mostly depends on the assumption that "new/shiny device is better
than old aging device".

Once progress comes to an end, it is likely that the "high end" will not
see enough revenue to justify its continued existence (say, people no
longer buying new CPUs and GPUs quite as often), and things will begin
to back-slide until things stabilize on whatever is the "local optimum".

Say, if progress stopped "right now", then we might see a situation
where CPUs from 2045 or 2050 would be made with a "worse" process node
than modern CPUs.

Say, 32nm, because anything, smaller nodes were too expensive to
maintain relative to their gain in terms of density or performance. And,
keeping a 10 or 14nm fab going only makes sense if one is moving enough
"product" to justify its maintenance; Which in turn depends on large
numbers of people buying a new PC every few years or so (vs using the
same PC for 10 or 15 years); ...

So, if a wall was hit, consumers would realize fairly quickly, and in a
decade, the PC manufactures would notice their PC sales effectively
tank, then cut back production, and also kill off any "enthusiast"
parts, fabs beginning to shut down, etc.

Well, that or "planned obsolescence" becomes even more of a thing than
it is already (can't keep using the same PC for 10 years because there
is a built-in time-bomb or similar that renders it unusable after 8 years).

....

> John Savard

On Monday, May 15, 2023 at 9:26:06 PM UTC-5, BGB wrote:
> On 5/15/2023 6:03 PM, Quadibloc wrote:
> > On Saturday, May 13, 2023 at 12:35:32 PM UTC-6, BGB wrote:
> >
> >> People who are young now may not even really know what it is like to
> >> have lived in a world where the expectation was that computers would
> >> keep getting bigger and faster, whereas in my younger years, the
> >> assumption that computers would keep getting bigger and faster was taken
> >> for granted...
> >
> > Arthur C. Clarke, near the beginning of his book _Profiles of the Future_, which
> > collected essays which originally appeared in a famous American magazine
> > which shall remain nameless, noted how the speed of transportation had been
> > rapidly increasing over the past several decades...
> >
> > but the trend, from the railroad to the jet airplane to the spaceship, could not
> > continue, as soon we would be travelling faster than the speed of light..
> >
> Yeah.
>
> Physical limits are annoying.
>
> For vehicles, there are many practical limits as to how quickly one can
> make them go.
>
> Cars: Ability to maintain stability and the response speed of the person
> driving them, etc. Even if driven by computers, it is unlikely they
> could remain stable at 200-300 mph (and even then, things like
> encountering random road debris would likely end in disaster, etc).
>
> Trains: Mechanical issues with the rails or debris/etc moving onto the
> track. Having a maglev train inside of a tunnel with the tunnel held at
> reduced pressure (with a helium atmosphere, *) could potentially allow
> for absurdly fast trains.
>
> *: Holding a tunnel at a full vacuum would be impractical, but 0.1 bar
> of helium would have nearly the same effect. One might need multiple
> layers of tunnel though (with a system of 1-way valves, such that sudden
> pressure spikes in the inner tunnel can be deflected into the outer tunnel).
>
> Seems like building a tunnel that doesn't crush under external air
> pressure would itself be an engineering challenge.
>
> Might be simpler/cheaper just to use 1 bar of helium (or hydrogen...).
>
>
>
> For space, plasma thrusters can be useful, but one needs a good power
> source (such as a nuclear reactor).
>
> For interstellar travel, things are more of an issue...
>
>
> I had imagined an idea resembling a hybrid of a plasma thruster,
> cyclotron, and a linear particle accelerator. For initial acceleration,
> a plasma thruster stage would be used (feeding through a linear
> accelerator stage).
>
>
> For higher speed operation, it would switch from using primarily plasma
> ionization to primarily using the cyclotron (idea being to try to push
> the exhaust velocity to up nearly light speed, *).
>
> Though, less obvious is where to send the heat from the reactor at this
> stage (if using a reactor for a plasma thruster, they could use the
> propellant as a heat-sink for the reactor; for cyclotron operation the
> amount of propellant used would likely be far less than the amount
> needed to effectively cool a reactor).
>
> *: As a ship approaches its exhaust speed, the relative thrust from the
> exhaust approaches zero. If the ship were moving faster than its own
> exhaust, then it would lose more velocity by firing its thrusters than
> it would gain
>
> Say, from an external observer, the exhaust would be chasing the ship.
> If the ship is traveling at 0.3c and fires a thruster with a velocity of
> 0.2c, then the exhaust is still traveling in the same direction as the
> ship at 0.1c, and the effect on the ship's total velocity would be
> analogous to a thruster firing in the opposite direction at 0.1c,
> pulling the ship to a point where its forwards velocity and exhaust
> velocity add up to 0.
>
> But, say, if the exhaust velocity is 0.999c or similar, then it can
> still keep accelerating; even with only a relatively small number of
> exhaust particles being ejected.
>
> Put then one basically needs to use a particle accelerator or similar to
> push each exhaust particle as fast as it can go (rather than it being
> about the volume or power that one can push through the engine). Say,
> the ship slowly transitioning from plasma to particle acceleration once
> it starts crossing 0.1c or so...
>
> But, say, the ship needs to get above roughly 0.9c before the effects of
> time-dilation and similar become "useful".
>
>
> Well, and then the ship would need to expend a similar amount of energy
> decelerating before it reaches the target.
>
> Say, a ship crashing into a planet at a good part of light-speed being
> "not useful" (well, apart from possibly making a "world breaking"
> explosion).
>
>
> Though, for whatever reason, planetary assault with high velocity
> projectiles isn't really a thing in sci-fi (usually all the emphasis is
> on directed energy weapons or similar, rather than kinetic projectiles).
>
> Or, say, space combat with people firing cannons (or railguns) at each
> other and randomly changing velocity in an attempt to throw off the aim
> of the enemy ships (since the travel-time of the projectile would become
> a lot more significant in space-based combat).
> > He used that example to illustrate the phenomenon which we are running into
> > with computers - the "S-curve"; a new technology ushers in a phase where there
> > is rapid improvement, at an exponential rate, but that eventually comes to an
> > end, slowing down with diminishing returns.
> >
> Yep.
>
> I suspect my childhood was in the sharply rising part of the S-curve,
> but pretty much my entire adult life has been watching it slowly fizzle out.
>
>
>
> There were occasional "hype" technologies:
> Carbide semiconductors;
> Now displaced by Gallium Nitride;
> Graphene chips;
> Photonics;
> ...
>
> But, thus far, little has really amounted to much; each having severe
> drawbacks which limit either to niche applications (such as
> power-electronics and LEDs for Carbide and GaN) or seemingly kill it
> entirely (general situation with graphene).
> > A new invention, such as quantum computers, could usher in another phase
> > of rapid improvement.
> >
> I have some skepticism at this point about the "general purpose
> viability" of quantum computers.
>
> But, as-is, we don't have anything that is really a viable (or even
> "particularly promising") alternative to the existing CMOS technology,
> which is concerning...
>
>
> Sadly, beyond just "general stagnation", the existence of "high end"
> chips will likely only continue so long as there is a market for them,
> which mostly depends on the assumption that "new/shiny device is better
> than old aging device".
>
> Once progress comes to an end, it is likely that the "high end" will not
> see enough revenue to justify its continued existence (say, people no
> longer buying new CPUs and GPUs quite as often), and things will begin
> to back-slide until things stabilize on whatever is the "local optimum".
>
>
> Say, if progress stopped "right now", then we might see a situation
> where CPUs from 2045 or 2050 would be made with a "worse" process node
> than modern CPUs.
>
> Say, 32nm, because anything, smaller nodes were too expensive to
<
20nm is the current best price performance place. Can still be done with
planar transistors, and only polysilicon layer needs double masks. Smaller
than 20nm required FinFETs, and FinFETs require months in the "front end"
of the fab making fins 20- atoms wide and 200+ atoms tall.
<
> maintain relative to their gain in terms of density or performance. And,
> keeping a 10 or 14nm fab going only makes sense if one is moving enough
> "product" to justify its maintenance; Which in turn depends on large
> numbers of people buying a new PC every few years or so (vs using the
> same PC for 10 or 15 years); ...
>
> So, if a wall was hit, consumers would realize fairly quickly, and in a
> decade, the PC manufactures would notice their PC sales effectively
> tank, then cut back production, and also kill off any "enthusiast"
> parts, fabs beginning to shut down, etc.
>
> Well, that or "planned obsolescence" becomes even more of a thing than
> it is already (can't keep using the same PC for 10 years because there
> is a built-in time-bomb or similar that renders it unusable after 8 years).
>
> ...
>
>
> > John Savard

On Monday, May 15, 2023 at 6:03:46 PM UTC-5, Quadibloc wrote:
> On Saturday, May 13, 2023 at 12:35:32 PM UTC-6, BGB wrote:
>
> > People who are young now may not even really know what it is like to
> > have lived in a world where the expectation was that computers would
> > keep getting bigger and faster, whereas in my younger years, the
> > assumption that computers would keep getting bigger and faster was taken
> > for granted...
> Arthur C. Clarke, near the beginning of his book _Profiles of the Future_, which
> collected essays which originally appeared in a famous American magazine
> which shall remain nameless, noted how the speed of transportation had been
> rapidly increasing over the past several decades...
>
> but the trend, from the railroad to the jet airplane to the spaceship, could not
> continue, as soon we would be travelling faster than the speed of light.
>
> He used that example to illustrate the phenomenon which we are running into
> with computers - the "S-curve"; a new technology ushers in a phase where there
> is rapid improvement, at an exponential rate, but that eventually comes to an
> end, slowing down with diminishing returns.
>
> A new invention, such as quantum computers, could usher in another phase
> of rapid improvement.
>
> John Savard

It is amazing that Moore's law has held for as long as it has.
The economic and social repercussions will continue to reverberate for years to come.

On Monday, May 15, 2023 at 8:26:06 PM UTC-6, BGB wrote:

> But, as-is, we don't have anything that is really a viable (or even
> "particularly promising") alternative to the existing CMOS technology,
> which is concerning...

The main reason we don't is that the CMOS technology has been
developed, over the years, so extensively that even a better
semiconductor material than silicon, just starting out, is not able to
produce better chips. To attain parity with silicon in its current state
requires extensive development now.

So improving silicon further is the line of least resistance, despite
progress now having slowed down. While this is unfortunate for
potential future progress, it isn't that the cost of such progress has
gone up - it's just that such progress can't pay for itself in the early
stages because silicon has gotten so far ahead. So we're not running
out of less common materials to make chips with.

The situation is, therefore, too much of a _good_ thing. If we really
need much faster chips than can be made with silicon, I guess that
we might end up needing large-scale government funding to get them.

A... breathing space... in the pace of technical advancement, during
which we can buy a computer, and benefit from our investment for
years, and during which innovation is in the form of architectural
improvement, so as to get the most out of the level of chip technology
that we have... is not all bad. It would mean that when a better
semiconductor is ready, we would be prepared to get more performance
out of it as well.

John Savard

On Thursday, May 11, 2023 at 9:50:18 PM UTC-5, JimBrakefield wrote:

I sent you an e-mail query about time and location for March 25 lecture.
I have not received reply.