Tuesday, August 18, 2026
Ryan Scherbarth (nvidia) joined the channel
Sayan Ghosh joined the channel
David Ozog joined the channel
Ryan Scherbarth (nvidia) renamed the channel from "2026-d1-1030-sponsor-lightmatter" to "2026-d1-1030-sponsor-talk-lightmatter"
Ryan Scherbarth (nvidia) renamed the channel from "2026-d1-1030-sponsor-talk-lightmatter" to "2026-d1-1015-sponsor-talk-lightmatter"
Jiaqi Lou joined the channel
bsahu joined the channel
Fabian joined the channel
Francois Labonte joined the channel
Daud Arslan joined the channel
Shahbozbek Hakimov joined the channel
Tong Xu joined the channel
myoung-gyun.suh joined the channel
xi.chen joined the channel
Mahfuz joined the channel
Varun Chotalia joined the channel
pierre-louis.benard joined the channel
Matthew Fricke joined the channel
Advm joined the channel
ahmadrazarehman033 joined the channel
Howard Wang joined the channel
rashid-ahmed.kukkady joined the channel
Wednesday, August 19, 2026
S
Hello lightmatter team, u guys are doing great work:)
A
question: if we use different colors, how do you handle challenges that comes with different wavelength and in turn different speeds
A
my question is now answered
P
Hi @Ayush Chaturvedi, take a look at the performance metrics on Nick's next slide.
✅ 1
S
With BiDi using 8 wavelengths each way on a single fiber, what do you see as the biggest scaling challenge as you move toward 400G per fiber and eventually higher bandwidths?
5 replies
S
@Nick Harris can u answer this too ?
N
Scaling challenges were inventing Guide for 16 wavelengths per fiber at 100 Gbps modulation
N
Need great power and spectral stability
S
That’s really interesting, Nick. I’m actually using 8-wavelength WDM in my photonic CNN to parallelize pooling operations, so the power and spectral stability challenge you mentioned is directly relevant to my work. I’d be very interested in understanding how Lightmatter approaches wavelength-level stability and calibration as you scale from 8 to 16+ wavelengths.
P
These are measured results.
P
I'll leave your 400G / 8 lambda question to Nick.
👍 1
M
What material in Guide generates the light?
2 replies
N
not patterned. just material.
R
What is the power for 400Gbps, 800Gbps and 1.6Tbps? What cooling is required?
2 replies
N
2pJ/bit. Cooling is air or liquid for the chips.
D
does this efficiency include laser power?
H
OCI MSA adopts CWDM for BiDi, whereas LM adopts interleaving. Can you introduce pros and cons of those two approaches?
2 replies
N
Interleaving is better and uses less optical power but you need better lasers.
We can do band interleaved and already do 🙂
H
Thank you so much for a kind reply 🙂
S
To Nick: If photonics is going to become the dominant interconnect technology for AI scale-up, what is the one remaining problem that you think the industry is still underestimating?
4 replies
N
completed optical engine test and rack-scale infra definition are the last big things. working on both.
S
Thanks Nick — that’s really interesting. On the rack-scale side, as you move toward 512-XPU scale-up domains, do you see the bigger challenge being the physical optical architecture — fiber routing, connector count, BiDi/WDM scaling — or the system-level co-design between the optical engine, XPU tray, and switch fabric?
N
See the vision whitepaper at OCP that covers how the racks are going to be designed. we've been leading this effort with our 19 partners.
S
thanks Nick - I'll definitely look at the OCP vision paper.
A
can light matter also provide overview about their software stack and pre-requisites about making their hardware work with other OEMs
C
Where can we use photonic interconnect today, can we replace all network links between nodes, between chips or inside chip package? What are the limitation or constrain we need to consider? How far is PIC integrate with EIC on die?
R
One issue with photonic is that it is still point-to-point. Will there ever be a Photonic switch that can split the SMF's wavelength/color, to support different routing, similar to spine-and-leaf based switches, to create a MESH network.
1 reply
N
photonic switches are built into passage. see m1000 demo on
lightmatter.co > products > passage M1000 EVK
A
The 2× decode result caught my attention. The HOTI paper is prefill-focused, so what model, batch and context regime, and copper baseline produced the decode number? Does it include the complete optical path, laser and control power, failures and retries, and KV-cache movement, or is it an analytical interconnect-only result?
L
Are you considering multiple rack as scale out?
S
@Leila Rashidi do you want to ask the question here #C0BPQE5R44X