🌐 Module 8 · Context and the Future · Chapter 8.4 · 9 min read

Y100 Photonics Vision

From electron to photon — SIDRA's 2031 chip.

Prerequisites

What you'll learn here

  • Explain photonic compute-in-memory
  • Compare optical MVM to electrical MVM
  • State silicon-photonics technology's current status
  • Detail Y100 photonic spec targets
  • Identify Türkiye's position in the post-2030 photonic AI race

Hook: Computing with Light

Y1-Y10: electrical memristor crossbar. Design target ~300 TOPS/W (INT8).

Y100 vision: hybrid photonic MVM. Theoretical ceiling 1000-10000 TOPS/W; state-of-the-art lab demos (Lightmatter, MIT 2023-25) measure in the ~10-100 TOPS/W range. SIDRA’s 10-year horizon aims to close that gap.

This chapter opens the photonic vision — an engineering target, not a proven product.

Intuition: Photon Beats Electron

Photons beat electrons on:

  • Propagation speed: light speed (2× electron drift).
  • No resistance → no IR drop.
  • Parallel wavelengths (WDM) → one waveguide carries 100 signals.
  • Low thermal effects.

Downsides: physical size large (wavelength ~1 µm), modulation slow (~50 GHz).

Formalism: Silicon Photonics + MVM

L1 · Başlangıç

Silicon photonics:

Optical devices in existing silicon fabs:

  • Waveguide: Si strip carries photons.
  • Modulator: electrical signal → photon amplitude.
  • Photodetector: photon → electrical.
  • Mach-Zehnder interferometer: two waves interfere → output controlled.

TSMC, Intel, GlobalFoundries have silicon-photonics services.

Photonic MVM:

MZI mesh:

  • N×N matrix = N×N Mach-Zehnder interferometers.
  • Each MZI tunable (with heater).
  • Input: N optical waves (wavelength division).
  • Output: N waves carrying the MVM result.

Speed:

  • MZI settle: 1 µs (heater thermal).
  • Light transit: 100 ps.
  • Modulation: 50 GHz.

Energy:

  • MZI heater: ~10 mW on.
  • Modulator: ~1 pJ/bit.
  • Photodetector: ~0.1 pJ/bit.

Y100 photonic design target (2031 roadmap):

  • 4096×4096 MZI matrix (target).
  • 16M MACs/step in parallel.
  • ~10 ns/MVM.
  • Theoretical ceiling ~1 TOPS/mJ = 1000 TOPS/W (assumes ideal light conversion + zero heater loss).
  • Realistic first-gen expectation: 100-300 TOPS/W (including heater, laser wall-plug efficiency, photodetector losses).

ℹ️ Comparison: Lightmatter Envise (2024, commercial): ~3-5 TOPS/W measured. MIT CSAIL research demos: 10-50 TOPS/W theoretical. 1000 TOPS/W has not yet been shown by anyone in production.

L2 · Tam

Hybrid: memristor + photonic:

Pure photonic is hard (heaters slow, no non-volatility).

Hybrid approach:

  • Weights in memristors (non-volatile, 256 levels).
  • MVM in photonics (fast, efficient).
  • Conversion: memristor → MZI heater control.

Y100 design: MZI mesh at the head of each crossbar. Electron + photon together.

Bandwidth:

Photonic waveguide: 100 channels × 100 Gbps = 10 Tbps. Electrical Cu wire: 100 Gbps max.

100× bandwidth advantage. Critical for chiplet-to-chiplet communication.

Heat:

MZI heaters run hot. 1000 MZIs × 10 mW = 10 W. Within Y100’s 100 W.

Alternative: electro-optic modulation (faster, cooler). MEMS also a candidate.

Thermal drift:

Temperature shifts MZI phase. Extra temperature-control circuitry.

Manufacture:

TSMC 7 nm + silicon photonics. Intel / GlobalFoundries also capable.

Türkiye: UNAM + BİLGEM silicon-photonics research exists. Y100 needs industrial-scale.

L3 · Derin

Rival photonic AI companies:

Lightmatter (US): Photonic AI chip. $300M funding. 2024 product. PsiQuantum (US): Quantum + photonic. Rain AI (US): Analog + photonic. Luminous (US): Silicon-photonics AI. Xanadu (Canada): Quantum-photonic.

Türkiye: none. Open territory.

Tech trends:

  • Integrated photonics: no longer a separate chip; CMOS + optics on one die.
  • Co-packaged optics (CPO): photons + electrons inside the package.
  • Neuromorphic photonics: spike-based photonic (next-gen).

Y100 timeline:

  • 2026: photonics research begins (UNAM + METU).
  • 2028: MZI prototype (FPGA-emulated).
  • 2030: silicon-photonics test chip (TSMC).
  • 2031-2033: Y100 photonic product.
  • 2035+: Y100 generation widespread.

Critical for Türkiye:

The photonic boom is just starting. Within 5 years the US + China will lead. If Türkiye enters now, it can be a player 2030-2035 in photonic AI.

Missed: SIDRA Y100 loses its global leadership potential.

Energy savings (scenario):

Y100 photonic design target: 100 W, 3 PFLOPS (inference, INT8). For comparison, NVIDIA B200 (2024, measured): 1000 W, ~10 PFLOPS FP8. If Y100 hits its design targets, it lands ~3× more efficient (FLOP/W) on inference workloads.

At datacenter scale: 30-50% electricity savings potential for inference. Actual rate depends on workload profile — SIDRA is not suitable for training.

Challenges:

  • Yield: photonic manufacturing is hard (target >10%).
  • Ecosystem: silicon-photonics design tools scarce.
  • Talent: photonic engineers few in Türkiye.
  • Investment: Y100 requires $1B+.

Experiment: Y100 vs NVIDIA 2030 Scenario

⚠️ The numbers below compare two roadmap scenarios. NVIDIA’s 2030 GPU is not yet announced; we use B200 (2024 measured) as the baseline and derive the 2030 estimate from the historical trend line (~2× energy efficiency every 4-5 years).

NVIDIA 2030 scenario (linear extrapolation from B200):

  • ~1500-2000 W, ~25-50 PFLOPS FP4.
  • ~25-50 TOPS/W design target.
  • ~$40-60K/chip.

SIDRA Y100 design target:

  • 100 W, 3 PFLOPS (inference, INT8).
  • Electronic layer ~300 TOPS/W + hybrid photonic first-gen target 500-1000 TOPS/W.
  • Target unit cost ~$2K/chip (assuming domestic production + simpler packaging).

Comparison (if both scenarios materialize):

  • FLOP/W: SIDRA ~10-20× more efficient (on inference).
  • Cost/chip: SIDRA ~20-30× cheaper (domestic + niche segment).
  • Raw speed (FLOPS): NVIDIA ~10× faster (still the pick for training).

Use segment:

  • Datacenter training: NVIDIA (dense FP4/FP8).
  • Datacenter inference: SIDRA Y100 target (energy + cost).
  • Edge: SIDRA already suitable from Y1 onward.

Market scenario 2033:

  • NVIDIA GPU revenue estimate (Gartner/McKinsey scenarios): $100-150B/year.
  • SIDRA target market share of 1-3% of the inference segment = $1-5B/year.

Very large at Türkiye’s scale, niche globally. The likelihood of realization hinges on disciplined execution of the roadmap.

Quick Quiz

1/6Y100 photonic advantage?
Y100 photonic advantage?

Lab Exercise

Y100 photonic investment plan.

2026-2028: Foundations.

  • UNAM + METU silicon photonics R&D.
  • 5 PhD students.
  • $2M fund.

2028-2030: Prototype.

  • TSMC silicon-photonics MPW ($500K).
  • FPGA emulator.
  • 20 engineers.

2030-2033: Product.

  • Full fab + photonics ($1B investment).
  • 200+ engineers.
  • First Y100 shipments 2031.

2033+: Scale.

  • Global market.
  • $5B/year revenue.

Total 10-year investment: $2B. ROI: 5-10 years.

Large for Türkiye but realistic strategic investment.

Cheat Sheet

  • Y100 hybrid: memristor + silicon photonics (roadmap).
  • Photonic MVM: MZI mesh, theoretically light-speed transit (~10 ns/MVM target).
  • Design target: 3 PFLOPS, 100 W, 300-1000 TOPS/W — a roadmap for 2031+.
  • Rivals: Lightmatter (commercial, ~3-5 TOPS/W measured), Rain, Luminous. No equivalent player in Türkiye.
  • Timeline: 2031-2033 prototype target.
  • Investment: estimated $2B / 10 years (public + private).

Vision: The Photonic Era

  • Y100 (2031): hybrid photonic prototype.
  • Y200 (2033): fully photonic MVM.
  • Y1000 (2040+): photonic + quantum + bio-compatible triad.

AI’s physics revolution. Türkiye’s chance to be in the room.

Further Reading

  • Next chapter: 8.5 — Your Place
  • Previous: 8.3 — Ethics
  • Silicon photonics: Reed, Silicon Photonics, Wiley.
  • Lightmatter: lightmatter.co official.
  • Photonic AI: Shen et al., Deep learning with coherent nanophotonic circuits, Nature Photonics 2017.