top of page

📢 STT 訂閱專區已上線

免費文章會照常更新,一篇都不會少。訂閱是「加強版」——每週深度週評、財報法說的完整判讀、所有長篇深度報告全包。

免費讓你跟上,訂閱讓你看懂、能做判斷。

月訂 NT$199|年訂 NT$2,000(約 NT$167/月)
👉 立即訂閱: vocus.cc/salon/simpletechtrend

HOT CHIPS 2025: How Celestial AI's Photonic Fabric Module Redefines SoC I/O Architecture

2 days ago
3 min read

The Real SoC Bottleneck Is No Longer Compute, but I/O

For more than a decade, the semiconductor industry has improved system performance by stacking ever more compute. That still holds in the AI era, but clear structural limits are starting to appear.

  • Whether CPU, GPU or AI accelerator, the design focus has been to keep adding compute units, cache capacity and internal bandwidth

  • As die area grows, compute can expand toward the center of the chip, but I/O can only sit along the chip edge

  • Die area grows much faster than perimeter, so I/O resources cannot scale in step with compute

  • In large AI SoCs, even when compute can still increase, the ability to move data on and off the chip becomes the performance bottleneck

This is not a design choice; it is an unavoidable consequence of physical geometry.

Why Electrical I/O Alone Can No Longer Solve the Problem

Electrical interconnect still works over short distances, but at high bandwidth and longer reach its limitations are quickly amplified.

  • As operating frequency rises, electrical I/O power grows non-linearly

  • Signal integrity and noise margins shrink, making system stability harder to maintain

  • Package and routing design grow ever more complex, raising both design cost and risk

  • Even advanced process nodes, packaging and SerDes technology can only delay the problem

Celestial AI's conclusion is blunt: if the bottleneck comes from physical scale itself, continuing to optimize electrical interconnect is not a long-term answer.


Celestial AI's Core Idea: I/O Should Not Live Only at the Chip Edge

Most optical interconnect approaches start by moving optics as close to the chip edge as possible to shorten electrical traces. Celestial AI instead chose to redefine the entire I/O architecture.

  • The Photonic Fabric Module is not simply co-packaged optics

  • Silicon photonics is deeply integrated into the interposer, not just a peripheral module

  • Optical I/O can be placed inside the die footprint, close to the cache, memory controller or internal fabric

  • I/O is no longer forced to line up along the perimeter; it becomes a freely placeable architectural resource

This means the SoC's I/O floorplan is no longer dictated by the physical edge but by system architecture needs.


Why Celestial AI Chose EAM over Micro-Rings

For its optical modulator, Celestial AI took a very pragmatic engineering approach rather than chasing peak efficiency under lab conditions.

  • Micro-rings offer high energy efficiency and wavelength density

  • But they depend heavily on resonance conditions and are extremely sensitive to temperature changes

  • In high-TDP AI SoCs they require continuous wavelength tuning and control, adding system complexity and risk

By comparison, the Electro-Absorption Modulator (EAM) is a better fit for system-level requirements.

  • No reliance on resonant structures, with high tolerance to temperature variation

  • Simpler drive and control circuitry, with lower integration cost

  • Predictable behavior and high stability in large-scale systems

In very large AI systems, stability often matters more than theoretical peak energy efficiency.


What the Photonic Fabric Module Means at the Architecture Level

The Photonic Fabric Module is not about the speed of a single link; it addresses the fundamental question of how SoCs keep scaling.

  • I/O bandwidth can scale with die area instead of being limited by edge perimeter

  • Replacing long electrical interconnects with optics helps cut power and improve signal integrity

  • It gives multi-die and heterogeneously integrated SoCs much more design freedom

From a packaging perspective, this is advanced integration; from an architecture perspective, it is a shift in design philosophy.


How It Differs from Conventional Co-Packaged Optics

Conventional co-packaged optics mainly address front-panel density and module power, but they keep the edge-bound I/O structure.

  • Optical resources are concentrated at the package edge

  • I/O remains a perimeter-limited resource

  • It cannot fundamentally solve the I/O scaling problem

Celestial AI's Photonic Fabric Module takes a completely different direction.

  • Optical resources are distributed across the entire chip plane

  • I/O becomes part of the internal architecture rather than an add-on

  • It is closer to the concept of in-die optical I/O


Simple Tech Trend's Take

Celestial AI is not offering a solution ready for volume production in today's data centers; it is paving the way for the next generation of SoC architecture.

  • Once electrical I/O can no longer support compute scaling, bringing optics inside the chip becomes inevitable

  • The Photonic Fabric Module offers a direction that is both engineering-feasible and architecturally sound

  • I/O is no longer a secondary resource at the chip edge but a core element of system design

This Hot Chips 2025 talk showed how optical interconnect is moving from a "packaging option" to an "architectural foundation."

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page