Has anyone heard of this idea? It apparently delegates some
lower-level computing functions directly to the memory itself, to get
a speedup from doing everything in the CPU. It seems to be an
outgrowth of the “memristor” component that was discovered/invented by
some researchers at HP a few decades ago.

The researchers in this paper have come up with a Python library to
make the technology easier to write programs for.

<https://www.tomshardware.com/pc-components/cpus/researchers-develop-python-code-that-is-compatible-with-in-memory-computing-python-commands-converted-into-machine-code-to-be-executed-in-the-computers-memory>

Lawrence D'Oliveiro wrote:
> Has anyone heard of this idea? It apparently delegates some
> lower-level computing functions directly to the memory itself, to get
> a speedup from doing everything in the CPU. It seems to be an
> outgrowth of the “memristor” component that was discovered/invented by
> some researchers at HP a few decades ago.

Delegating memory operations to lower layers in the hierarchy is one of
those wheel of re-incarnation ideas that pop back up every decade or two.

You typically start with shared atomic operations and very simple
computation, like a LOCK XADD, then once you are on this slippery slope
you quickly decide to add more advanced capabilities, quickly ending up
with something like Bunny Chang's distributed virtual machine which can
securely distribute its code anywhere in the cluster.

Taken to its extreme, any cloud datacenter works this way, but at a far
higher granularity.

Terje

--
- <Terje.Mathisen at tmsw.no>
"almost all programming can be viewed as an exercise in caching"

On Wed, 13 Nov 2024 07:46:28 +0100, Terje Mathisen wrote:

> Taken to its extreme, any cloud datacenter works this way, but at a far
> higher granularity.

Maybe you’re thinking of supercomputers. They have a processor
interconnect which is sufficiently fast as to make the collection of nodes
behave more like a single machine. Just having a cloud data centre filled
with nodes on a conventional network isn’t quite the same thing.

Terje Mathisen <terje.mathisen@tmsw.no> schrieb:

> You typically start with shared atomic operations and very simple
> computation, like a LOCK XADD, then once you are on this slippery slope
> you quickly decide to add more advanced capabilities, quickly ending up
> with something like Bunny Chang's distributed virtual machine which can
> securely distribute its code anywhere in the cluster.

What is Bunny Chang's distributed virtual machine?

On Wed, 13 Nov 2024 06:30:34 -0000 (UTC)
Lawrence D'Oliveiro <ldo@nz.invalid> wrote:

> <https://www.tomshardware.com/pc-components/cpus/researchers-develop-python-code-that-is-compatible-with-in-memory-computing-python-commands-converted-into-machine-code-to-be-executed-in-the-computers-memory>

Quoth the article:

> Compute code written for conventional computers has purportedly
> "barely changed" since the 1940s.

I'm curious to hear what the practical implications of the real thing
are, but that right there is some laughable clickbait garbage written
by someone who has absolutely no idea what they're talking about.

On Wed, 13 Nov 2024 6:30:34 +0000, Lawrence D'Oliveiro wrote:

> Has anyone heard of this idea? It apparently delegates some
> lower-level computing functions directly to the memory itself, to get
> a speedup from doing everything in the CPU. It seems to be an
> outgrowth of the “memristor” component that was discovered/invented by
> some researchers at HP a few decades ago.

Denelcore: 1980:: had atomic memory ops in memory; so at least 40 YO.

Thomas Koenig wrote:
> Terje Mathisen <terje.mathisen@tmsw.no> schrieb:
>
>> You typically start with shared atomic operations and very simple
>> computation, like a LOCK XADD, then once you are on this slippery slope
>> you quickly decide to add more advanced capabilities, quickly ending up
>> with something like Bunny Chang's distributed virtual machine which can
>> securely distribute its code anywhere in the cluster.
>
> What is Bunny Chang's distributed virtual machine?
>

Oops, typo! I meant Andrew "Bunnie" Huang, not Chang. :-(

He's the guy who got famous for reverse engineering all the encrypted
Xbox stuff. He's done a lot since then, including one time when he
talked about hacking thumb drives to do basically anything:

Every single flash drive contains a full 32-bit CPU, upon first boot it
surveys all the connected flash blocks and test them, eventually
deciding how many are good and then picking the correct size of the drive.

There is nothing that prevents such a cpu to be reprogrammed to also act
as a keyboard or mouse input device, or to hide half the available disk
space and use it to keep copies of everything that gets deleted from the
visible part.

Terje

--
- <Terje.Mathisen at tmsw.no>
"almost all programming can be viewed as an exercise in caching"

On Wed, 13 Nov 2024 19:38:13 +0000, MitchAlsup1 wrote:

> On Wed, 13 Nov 2024 6:30:34 +0000, Lawrence D'Oliveiro wrote:
>
>> Has anyone heard of this idea? It apparently delegates some
>> lower-level computing functions directly to the memory itself, to get
>> a speedup from doing everything in the CPU. It seems to be an
>> outgrowth of the “memristor” component that was discovered/invented by
>> some researchers at HP a few decades ago.
>
> Denelcore: 1980:: had atomic memory ops in memory; so at least 40 YO.

They didn’t have memristors back then, though. This paper uses memristors
in place of traditional DRAM/SRAM memory cells. The resulting read/write
networks look remarkably like old-style magnetic-core memories, except
that these cells can act as logic gates to perform operations in parallel.

On Fri, 15 Nov 2024 3:19:55 +0000, Lawrence D'Oliveiro wrote:

> On Wed, 13 Nov 2024 19:38:13 +0000, MitchAlsup1 wrote:
>
>> On Wed, 13 Nov 2024 6:30:34 +0000, Lawrence D'Oliveiro wrote:
>>
>>> Has anyone heard of this idea? It apparently delegates some
>>> lower-level computing functions directly to the memory itself, to get
>>> a speedup from doing everything in the CPU. It seems to be an
>>> outgrowth of the “memristor” component that was discovered/invented by
>>> some researchers at HP a few decades ago.
>>
>> Denelcore: 1980:: had atomic memory ops in memory; so at least 40 YO.
>
> They didn’t have memristors back then, though. This paper uses
> memristors
> in place of traditional DRAM/SRAM memory cells. The resulting read/write
> networks look remarkably like old-style magnetic-core memories, except
> that these cells can act as logic gates to perform operations in
> parallel.

In a Sph. project, we were given a ferrite core (~1 pound) and were told
to use it as a counter, adding up when a new car entered a parking lot,
and subtracting down when a car left. So, doing arithmetic in ferrite
cores has been around for a very long time, indeed. {{OH, BTW, the
purpose of the count was to prevent overflowing of the parking lot}}

On Sun, 17 Nov 2024 21:32:29 +0000, MitchAlsup1 wrote:

> ... doing arithmetic in ferrite cores has been around for a very long
> time, indeed.

Memristors are a new kind of electronic component, where the resistance is
proportional to the integral of applied voltage over time.

According to Lawrence D'Oliveiro <ldo@nz.invalid>:
>On Sun, 17 Nov 2024 21:32:29 +0000, MitchAlsup1 wrote:
>
>> ... doing arithmetic in ferrite cores has been around for a very long
>> time, indeed.
>
>Memristors are a new kind of electronic component, where the resistance is
>proportional to the integral of applied voltage over time.

This is a rather capacious version of "new" since memristors were invented in 1971.

My impression is that they are real, they work, but they don't work well enough to
replace conventional components.

There is a very long article about them in Wikipedia.
--
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 Mon, 18 Nov 2024 15:25:37 -0000 (UTC)
John Levine <johnl@taugh.com> wrote:

> According to Lawrence D'Oliveiro <ldo@nz.invalid>:
> >On Sun, 17 Nov 2024 21:32:29 +0000, MitchAlsup1 wrote:
> >
> >> ... doing arithmetic in ferrite cores has been around for a very
> >> long time, indeed.
> >
> >Memristors are a new kind of electronic component, where the
> >resistance is proportional to the integral of applied voltage over
> >time.
>
> This is a rather capacious version of "new" since memristors were
> invented in 1971.
>
> My impression is that they are real, they work, but they don't work
> well enough to replace conventional components.
>
> There is a very long article about them in Wikipedia.

My impression from Wikipedia article is different. Memristors are not
real.
I.e. there are no physical devices that approximate mathematical
abstraction proposed in 1971. There are some devices taht look like
that, but only before researcher starts to pay attention to details.
After researcher starts to pays attention to details it typically turns
out that device resistance does not really depend on charge, but on
something else that happens to correlate with charge on bigger or
smaller parts of characteristic curves.

What does exist and does work and does not work well enough relatively
to conventional tech are various variants of ReRAM. But memory elements
of those various ReRAMs are *not* memristors. That applies as much to
HP's not quite working "memristor" ReRAM as to all others ReRAMs in
existence including those that work relatively better.

On Mon, 18 Nov 2024 15:25:37 -0000 (UTC), John Levine wrote:

> ... memristors were invented in 1971.

They were theorized in 1971, but there was no physical component was
created that came close to the theoretical behaviour until somewhat more
recently.

> There is a very long article about them in Wikipedia.

Yes <https://en.wikipedia.org/wiki/Memristor>. Among other things, it says
their distinguishing characteristic is a linear relationship between the
rate of change of flux and the rate of change of charge.

But the paper that I linked to in the posting that started this thread
shows a connection between the resistance, and the integral of voltage
over time.

Ah, I get it: the Wikipedia article says the “flux” thing is indeed the
integral of voltage with respect to time.

On 18/11/2024 19:09, Michael S wrote:
> On Mon, 18 Nov 2024 15:25:37 -0000 (UTC)
> John Levine <johnl@taugh.com> wrote:
>
>> According to Lawrence D'Oliveiro <ldo@nz.invalid>:
>>> On Sun, 17 Nov 2024 21:32:29 +0000, MitchAlsup1 wrote:
>>>
>>>> ... doing arithmetic in ferrite cores has been around for a very
>>>> long time, indeed.
>>>
>>> Memristors are a new kind of electronic component, where the
>>> resistance is proportional to the integral of applied voltage over
>>> time.
>>
>> This is a rather capacious version of "new" since memristors were
>> invented in 1971.
>>
>> My impression is that they are real, they work, but they don't work
>> well enough to replace conventional components.
>>
>> There is a very long article about them in Wikipedia.
>
> My impression from Wikipedia article is different. Memristors are not
> real.
> I.e. there are no physical devices that approximate mathematical
> abstraction proposed in 1971. There are some devices taht look like
> that, but only before researcher starts to pay attention to details.
> After researcher starts to pays attention to details it typically turns
> out that device resistance does not really depend on charge, but on
> something else that happens to correlate with charge on bigger or
> smaller parts of characteristic curves.
>

All electronic devices are approximations. There is no such thing as a
pure resistor, or a pure capacitor, or pure inductor. Current
memristors are no different in principle, but are - for now, at least -
poorer approximations than the more common components. Whether they
will ever be close enough to be of practical use, remains to be seen.

> What does exist and does work and does not work well enough relatively
> to conventional tech are various variants of ReRAM. But memory elements
> of those various ReRAMs are *not* memristors. That applies as much to
> HP's not quite working "memristor" ReRAM as to all others ReRAMs in
> existence including those that work relatively better.
>

Yes, that is my understanding too - there are a variety of memory
devices that have been made with different properties and niches, but I
don't believe any of them are based on devices that are close enough to
ideal memristors to justify using the term.