From Newsgroup: comp.arch

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.

For you young folks, delay lines memories were common in the 1940s and
1950s before core memory replaced them, but they didn't use fiber.

Abstract

The rising pressure on DRAM availability and contract pricing reflects generative AI's massive high-performance memory requirements. This
pressure is heavily compounded by hyperscale data center expansion,
which now consumes a significant portion of global DRAM output. In
this work, we propose a new architecture: Fiber Memory, which
reimagines the role of optical fiber in a hyperscale data center,
deploying it as an active, recirculating delay-line memory for
immutable data, such as large language model (LLM) weights. We present
a data-parallel optical broadcast delay-line memory architecture that
accounts for fiber's physical realities. By incorporating
space-division multiplexed multi-core fibers (MCFs), passive optical tap-and-amplify interfaces, co-packaged optics (CPO), and regional
all-optical regeneration, our case study evaluation demonstrates that
Fiber Memory can eliminate redundant weight storage across 10,000 AI accelerators and reduce weight-delivery energy by over 70% compared to traditional HBM3e configurations.

https://arxiv.org/abs/2607.08407
--
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
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

John Levine <johnl@taugh.com> writes:

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.

For you young folks, delay lines memories were common in the 1940s and
1950s before core memory replaced them, but they didn't use fiber.

Abstract

The rising pressure on DRAM availability and contract pricing reflects >generative AI's massive high-performance memory requirements. This
pressure is heavily compounded by hyperscale data center expansion,
which now consumes a significant portion of global DRAM output. In
this work, we propose a new architecture: Fiber Memory, which
reimagines the role of optical fiber in a hyperscale data center,
deploying it as an active, recirculating delay-line memory for
immutable data, such as large language model (LLM) weights. We present
a data-parallel optical broadcast delay-line memory architecture that >accounts for fiber's physical realities. By incorporating
space-division multiplexed multi-core fibers (MCFs), passive optical >tap-and-amplify interfaces, co-packaged optics (CPO), and regional >all-optical regeneration, our case study evaluation demonstrates that
Fiber Memory can eliminate redundant weight storage across 10,000 AI >accelerators and reduce weight-delivery energy by over 70% compared to >traditional HBM3e configurations.

https://arxiv.org/abs/2607.08407

Interesting. From the summary, it's more like EPROM than DRAM.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch


John Levine <johnl@taugh.com> posted:

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.

For you young folks, delay lines memories were common in the 1940s and
1950s before core memory replaced them, but they didn't use fiber.

Abstract

The rising pressure on DRAM availability and contract pricing reflects generative AI's massive high-performance memory requirements. This
pressure is heavily compounded by hyperscale data center expansion,
which now consumes a significant portion of global DRAM output.

Most of the memory being used for AI compute loads is necessary only
because of the interface between SSDs and DRAM. There are ways where
the same amount of data needs to be read off the SSDs, but less than
0.15% needs to arrive in DRAM.

For example, say one put FMAC units down on the SSD-FLASH controller
so that when SSD ships read-data, it ships it through said FMAC unit
to perform the cosine-similarity calculation and only ships only the
comparison result up to DRAM for CPU uses.

Vector length vary between 768 to 4096 elements (FP32) so cosine
results are between 0.15% and 0.025% of vector data size.

In
this work, we propose a new architecture: Fiber Memory, which
reimagines the role of optical fiber in a hyperscale data center,
deploying it as an active, recirculating delay-line memory for
immutable data, such as large language model (LLM) weights. We present
a data-parallel optical broadcast delay-line memory architecture that accounts for fiber's physical realities. By incorporating
space-division multiplexed multi-core fibers (MCFs), passive optical tap-and-amplify interfaces, co-packaged optics (CPO), and regional all-optical regeneration, our case study evaluation demonstrates that
Fiber Memory can eliminate redundant weight storage across 10,000 AI accelerators and reduce weight-delivery energy by over 70% compared to traditional HBM3e configurations.

https://arxiv.org/abs/2607.08407

--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.
For you young folks, delay lines memories were common in the 1940s and
1950s before core memory replaced them, but they didn't use fiber.

Right, but this is specially designed for the case where "every"
processor in your large system needs repeatedly the same sequence
of data. That seems potentially applicable to today's LLMs, but it's
*highly* specialized.

I don't like the direction this is going since it seems to imply that
running such LLMs would be inherently limited to large data-centers,
thereby taking control out of the end-users hands.


=== Stefan
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

In article <jwvpl0umsyt.fsf-monnier+comp.arch@gnu.org>, monnier@iro.umontreal.ca (Stefan Monnier) wrote:

I don't like the direction this is going since it seems to imply
that running such LLMs would be inherently limited to large
data-centers, thereby taking control out of the end-users hands.

That seems entirely consistent with the agendas of the companies running
most of the world's LLMs.

John
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

According to Stefan Monnier <monnier@iro.umontreal.ca>:

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.
For you young folks, delay lines memories were common in the 1940s and
1950s before core memory replaced them, but they didn't use fiber.

Right, but this is specially designed for the case where "every"
processor in your large system needs repeatedly the same sequence
of data. That seems potentially applicable to today's LLMs, but it's >*highly* specialized.

I don't like the direction this is going since it seems to imply that
running such LLMs would be inherently limited to large data-centers,
thereby taking control out of the end-users hands.

I don't think it's much of a change from the current situation, since
the fiber loop is intended to replace gigabyte HBMs which already only
can work in large data centers. If they needed fewer HBMs so the RAM manufacturers could make more DRAM and SD, we'd all be better off.

This is separate from the issue of whether you really need a giant
data centre to do useful inference, or the Apple model where you do
it mostly on your device and only call out to the cloud occasionally
will work.
--
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
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

On Fri, 10 Jul 2026 02:42:59 -0000 (UTC), John Levine
<johnl@taugh.com> wrote:

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.

For you young folks, delay lines memories were common in the 1940s and
1950s before core memory replaced them, but they didn't use fiber.

In the 70's and 80's there was considerable interest in magnetic
bubble memory which essentially is a delay line technology. It was
expensive, and bit density was a problem, but since it was
non-volatile, it often was used in machines designed for harsh
environments.

In the 80's Olivetti sold as a (IBM)PC expansion board a solid state
drive that used bubble memory.

In the 2K's IBM's "racetrack" memory was a form of bubble memory.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

On Sat, 11 Jul 2026 14:44 +0100 (BST), John Dallman wrote:

That seems entirely consistent with the agendas of the companies
running most of the world's LLMs.

On the bright side, they seem to be having real trouble making money
from the massive investments and expenditure being made in equipment
and software/training and, in particular, on electricity bills.

The pesky users don’t seem to want to pay $£€¥ anywhere near what it costs to provide the service.
--- Synchronet 3.22a-Linux NewsLink 1.2

From Newsgroup: comp.arch

John Levine [2026-07-10 02:42:59] wrote:

This paper proposes using an optical fiber ring as a delay line to
store and distribute data in AI data centers. I think they're serious.

BTW, I think the core of the idea is to recognize that some of the
workload in large AI datacenters is not just "SIMT" but that some part
of the data is exactly the same for some of those threads (e.g. say most
of the threads are performing a "vector x matrix" multiply with
a different vector each but all with the same matrix), so you might be
able to get similar benefits with some kind of "broadcast"
DRAM/SSD reads.


=== Stefan
--- Synchronet 3.22a-Linux NewsLink 1.2