HOT CHIPS 2025: Lightmatter's Passage M1000 and Why a 3D Photonic Interposer Is Inevitable for AI Systems

Oct 5

3 min read

As AI systems keep scaling, the real bottleneck is moving to the package level

As AI model sizes and system compute keep expanding, the problem is no longer just single-chip performance, but how the whole system is effectively "assembled."

A single GPU or ASIC cannot hold an entire model and its data flow
System performance depends heavily on chip-to-chip interconnect bandwidth and latency
Packaging is no longer just putting chips together; it is part of system performance
Conventional packaging and interconnect are approaching physical limits

Lightmatter's core thesis: if AI systems are to keep scaling, the packaging architecture itself must be redesigned.

Why the conventional 2D interposer is nearing its limits

Today's mainstream advanced packaging is still largely built on a two-dimensional, planar interposer architecture.

Signals and I/O resources are placed mainly along the chip edges
Bandwidth scaling depends heavily on perimeter length
Routing density and signal crossing problems are getting worse
Power and latency rise quickly with distance

Even with Co-Packaged Optics, the approach is essentially an extension of the 2D plane and does not truly change the "edge-limited" structure.

The core idea of Passage M1000: a true 3D photonic interposer

Passage M1000, unveiled by Lightmatter at Hot Chips 2025, is not an optical module in the traditional sense, but a 3D photonic interposer.

The photonic layer becomes the core of the package structure, not an add-on
Compute dies are bonded face-down directly onto the photonic layer
Optical TX/RX become built-in resources beneath the chip
I/O is no longer concentrated at the edge but can be distributed across the entire die area

This design fundamentally redefines the roles of packaging and interconnect.

Why "3D" is the key, not just optics

Lightmatter emphasizes 3D rather than photonics alone, because architecture-level limits cannot be solved by changing materials alone.

The bottleneck of 2D architectures lies in edges and planar routing
3D architectures can introduce vertical I/O and bandwidth
Optics provides reach and energy efficiency; 3D provides density and scalability

The value of Passage M1000 is that it addresses both problems at once rather than optimizing only one.

System capabilities enabled by Passage M1000

At the system level, Passage M1000 offers an entirely new interconnect platform rather than a single product.

Package bandwidth on the order of hundreds of Tbps
Supports thousands of SerDes lanes and optical channels
Can connect directly to high-density fiber
Supports reconfigurable optical circuit switching

This frees AI system interconnect topologies from the constraints of fixed electrical wiring.

Why this path is not chasing near-term volume production

Lightmatter has not positioned Passage M1000 as a solution for immediate high-volume production, which is a very honest and important signal.

Manufacturing and yield barriers for 3D photonic interposers are extremely high
Packaging cost is far higher than existing 2.5D or CPO architectures
System adoption requires highly customized design

But that does not make it unimportant. Quite the opposite: it means it is aimed at the next unavoidable stage.

What Passage M1000 means for hyperscale AI systems

In hyperscale AI systems, the traditional "more racks, more switches" strategy is starting to break down.

Accumulated system latency hurts synchronization efficiency
Power consumption rises fast while marginal returns decline
Topology design keeps getting more complex

The 3D photonic interposer offers a different path.

Pull more interconnect capability back inside the package
Let the system operate like one much larger chip
Reduce reliance on external switching tiers

This is exactly the future direction Passage M1000 aims to demonstrate.

How it differs from the Celestial AI and Ayar Labs approaches

Put these three companies on the same optical interconnect map and the differences become clear.

Celestial AI brings optics inside the SoC
Ayar Labs extends the existing electrical ecosystem with retimers
Lightmatter rewrites the structure of the package itself

Passage M1000 is the most radical and the most long-term of the three approaches.

Simple Tech and Trend's take

What Lightmatter is showing is not a next-generation product, but the next physical and architectural boundary.

When 2D packaging can no longer support system growth
When interconnect becomes the main limit on performance and power
3D photonic interposers will move from research topic to engineering option

Passage M1000 sends a clear signal: the future of AI systems must move from planar to three-dimensional.