2026-d1-1400-panel-interconnects-at-the-edge-data-center

ESNET question: what is driving real-time requirements for scientific data streaming? it would seem to me that most data can be recorded and then analyzed offline, and on the slide I saw data flowing in one direction only.

Does NTT uses scale across for inference? Or remote storage?

Does scale up and out have different requirements at edge than central cloud?

Are NTT or Geico edge DCs liquid cooled?

What are opinions about workload aware or compute aware routing in network?

GEICO question: how are business applications driving latency requirements? or other requirements such as throughput or reliability more important?

What would be potentials of AI grid (telco infra) in inference market??

(comment: AI crawlers seem to have evolved to resemble DDoS, unfortunately)

@Bob Lantz Real-time streaming is an important requirement for many facilities. I will give some examples:
1. A scientist at a beamline looking at a sample wants to see the 3-D movie reconstruction of the sample so they can change the depth of the view, or angle of the sample, or verify the hypothesis or come up with a different approach to view the sample.
2. SLAC’s LCLSII and BNL’s EIC are producing data at a rate that is very hard to store..they are projecting 1 Tbps of streaming of data which is very hard to store as raw data in storage given that the growth of storage i/o has not been at the same scale as network bandwidth.
3. For astronomy, there are people wanting real-time AI algorithms to find supernovae so all the telescopes in the world can point to it and get information at a deeper level…
In many of these cases, the raw data is not important to sotre, or hard to store, or expensive to store.

With automated labs, robotics, they also want this rapid reconstruction to steer the experiment. Think of a fusion reactor and shots, the plasma has to be managed as real-time as possible so it can continue to operate.

I am not sure if this helps @Bob Lantz

yes streaming is mostly one way - that was not a full engineering diagram.

@Ofer Shapiro, @Amy Leeland welcome

Some of the kinds of things we want to observe is the vibration of an atom and how that affects material properties

that requires real-time view as well

or how the protien unfolds when getting cut, how does the structure change

Hi Everyone. This is <mailto:ofer@resolight.ai|ofer@resolight.ai>

Yosuke is not on slack so I am posting replies on his behalf:
Q: Does NTT uses scale across for inference? Or remote storage?
A: NTT is trying to develop inter-DC AI inference infrastructure including remote GPU and remote storage. it comes from requirements from users and infrastructure environments.

Q: Does scale up and out have different requirements at edge than central cloud?
A: Yes they are different. For scale-out, securing multi-tennancy and data-sovereignity is very important requirement in edge-DC. Therefore, dynamic inter-rack interconnect is an essential capability.
For scale-up, low-power consumption is very important in the edge-DC.
Anyway, orchestration between scale-up/out/across is a key for AI inference era.

How agentic AI impacts traffic of scale out and across?

(comment: this is another answer to Dan's [I think] question earlier about what the differences are between scale-across and WAN)

On behalf of Amy Leeland:

At GEICO, many of the workloads we're evaluating for AI, automation, and platform services can tolerate hundreds of milliseconds of latency and still deliver a good user experience. What often matters more is reliability, economics, and operational simplicity
For example, if an associate uses a copilot for document summarization, claims analysis, knowledge retrieval, or content generation, whether the response comes back in 300 milliseconds or 700 milliseconds usually isn't what determines success.
What determines success is:

Is the service available?
Is it accurate?
Is it cost effective?
Can it scale reliably?
Can we operate it securely?
That's why I often say the more interesting question isn't latency, it's data movement.

On behalf of Yosuke Aragane:

Q: How agentic AI impacts traffic of scale out and across?
A: It makes a great impact. Traffic amount is increasing as an explosion. From edge-DC perspective, scale-across is more impactful rather than scale-out. Edge-DC does not have ultimate computing capability. It should cooperate with center DC so that scale-across is essential capability between edge-DC and center-DC.

Resolight question: copper has the advantages of avoiding optical-electrical conversion and wires being cheaper than fibers, as well as lower complexity, easier manufacturing, etc.; what is the rationale for replacing copper within the rack? is the argument that it is justified overall by power savings?

For Resolight: What does it mean for the rack to become a scope?

@Ofer Shapiro @mari

@Ofer Shapiro @mari

Resolight: @Ofer Shapiro if the GPU-&gt;Optical Interface-&gt;Optical Switch, where is the "switching logic" being done?

@Rabindra Guha (Cerio) switching logics - still happen at the switch, but by request from endpoint, rather than by packet header read

is this more of a circuit switch than package switch / grainularity?

@Bob Lantz the issue is - can copper carry these signal as far as the distance between multiple racks (e.g. across the aisle) - it can’t

@Dan Pitt poor choice of words (‘scope’) - the intent is that the workload is not happing all inside the rack as ‘compute chassis’ but across multiple racks.

Has anyone deployed OCS in edge data center?

@Ofer Shapiro

thanks for the interesting sessions and thoughtful question reponses!