Skip to main content

Lightelligence Tianjic (Silicon Photonics AI Accelerator)

Product Overview

Lightelligence is China's leading silicon photonics AI chip company, founded 2017, incubated by Tsinghua University + Chinese Academy of Sciences Shanghai Institute of Optics and Fine Mechanics, headquartered in Shanghai. Tianjic is its first commercial silicon photonics AI accelerator, 2022 released, 7nm process, 1.6 PFLOPS FP16 compute, ~100W TDP. It is China's competitor to Lightmatter Envise, known alongside Lightmatter and LightOn as the "Silicon Photonics Big Three".

Strategic significance: Under US technology export controls on Lightmatter, Lightelligence is China's only commercial silicon photonics AI company, undertaking National 863 Major Project and National Natural Science Foundation key projects, with primary customers being Chinese supercomputing centers, government HPC, AI for Science projects.

Core Specs

ItemParameter
ArchitectureLightelligence Tianjic (Optical Matrix Multiplication)
ProcessTSMC 7nm (photonic + electronic hybrid)
Photonic Cores64 MZIs (Mach-Zehnder Interferometers)
Digital Cores32x RISC-V processors (scheduling + activation)
HBM32GB HBM2E
Memory Bandwidth819 GB/s
FP161.6 PFLOPS
FP32~0.8 PFLOPS
INT83.2 POPS
TDP~100 W
Efficiency16 TOPS/W (H100 ~3 TOPS/W, 5x advantage)
Form FactorPCIe Gen4 x16
Interconnectproprietary optical interconnect (PCIe + fiber)
Mass Production2023 H2
Unit Price~$3,500-5,500

Photonic Computing Principle

DimensionTraditional GPULightelligence Tianjic
Compute MediumElectronic (CMOS transistors)Photonic (silicon photonic waveguides + MZI)
Matrix MultiplicationMAC arraysMZI mesh (optical propagation matmul)
Signal PropagationCopper wire + resistanceOptical waveguide (zero resistance)
Power SourceCV^2 (transistor switching)MZI modulators only
Thermal DesignComplex liquid coolingNatural air cooling (100W)
Frequency1-2 GHz>40 GHz (theoretical)
PrecisionFP64/FP32/FP16/INT8FP16/INT8 primarily
Drawback-Non-matrix ops still need electronics, model limited

MZI Matrix Multiplication Principle

Input vector (8-dim): [x0, x1, x2, x3, x4, x5, x6, x7] (optical signals)
Weight matrix W (8x8): configured into 64 MZIs
Output: y = W x x (optical signal superposition)
ADC: convert to digital (only 1 electronic step)

Key advantages:

  • 64 MZIs compute simultaneously (1 cycle = 8x8 = 64 multiply-adds)
  • Speed of light = zero wait (vs GPU HBM-limited)
  • Efficiency 16 TOPS/W (GPU 5-10x advantage)

32 RISC-V Processors

DimensionSpec
ArchitectureRISC-V 64-bit (custom SiFive cores)
Core Count32
Frequency1.5 GHz
CacheL1 32KB + L2 512KB per core
RoleScheduling + activation functions + non-matrix ops
Domestic100% RISC-V (vs ARM proprietary)

Why RISC-V: Compared to ARM Cortex-A78, RISC-V saves 20% power at same performance, and is fully independent and controllable (meets China's Xinchuang requirements).

Vendor Information

ItemContent
CompanyLightelligence
FounderShen Zhihua (Chinese Academy of Sciences Shanghai Institute of Optics + Tsinghua)
Founded2017
HeadquartersJiading District, Shanghai
Funding$280M+ (Series C 2022 led by: Northern Light Venture Capital)
Valuation (2025)~¥15B
2024 Revenue~¥0.8B (early commercialization)
Employees~500
National ProjectsNational 863 Major Project, National Natural Science Foundation key
FabTSMC 7nm + proprietary silicon photonics process (CAS Institute of Microelectronics)
Patents150+ silicon photonics AI patents
CustomersChinese supercomputing centers (Wuxi, Changsha, Shenzhen), government HPC, AI for Science

Silicon Photonics Big Three

CompanyCountryProductComputeTDPMass Production
LightmatterUSEnvise2 PF65W2023 H2
LightelligenceChinaTianjic1.6 PF100W2023 H2
LightOnFranceAlfred1.2 PF80W2023 experimental

Lightelligence advantage: only commercial silicon photonics AI in China (Lightmatter under US export controls), key national policy support.

Use Cases

  • AI for Science (protein structure, climate, materials simulation)
  • Supercomputing centers (China Wuxi, Changsha, Shenzhen supercomputers)
  • Government HPC (national security, AI inversion)
  • Low-power AI inference (100W air-cooled deployment)
  • Domestic projects (RISC-V + silicon photonics)
  • ❌ AI training (inference only)
  • ❌ International market (export controls)
  • ❌ Non-matmul tasks (RNN weak performance)

Key Features

  • Silicon Photonics Computing: China's only commercial photonic AI chip
  • 16 TOPS/W: H100 5x efficiency advantage
  • RISC-V 32-core: fully independent and controllable
  • 100W TDP: air-cooled deployment
  • National policy support: 863 Major Project
  • Drawbacks: ecosystem only 1 year, 1 year behind Lightmatter