Has anyone heard of this idea? It apparently delegates someDelegating memory operations to lower layers in the hierarchy is one of
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.
Taken to its extreme, any cloud datacenter works this way, but at a far higher granularity.
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.
<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>
Compute code written for conventional computers has purportedly
"barely changed" since the 1940s.
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.
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 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.
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.
... doing arithmetic in ferrite cores has been around for a very long
time, indeed.
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.
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}}
MitchAlsup1 <mitchalsup@aol.com> schrieb:
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}}
Sounds like an interesting project, I assume you could add some
extra logic :-)
Were there enough cores so you could use a one-hot representation,
or did you have to do something more elaborate?
... memristors were invented in 1971.
There is a very long article about them in Wikipedia.
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.
| Sysop: | DaiTengu |
|---|---|
| Location: | Appleton, WI |
| Users: | 993 |
| Nodes: | 10 (0 / 10) |
| Uptime: | 207:04:13 |
| Calls: | 12,972 |
| Calls today: | 1 |
| Files: | 186,574 |
| Messages: | 3,268,391 |