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
| Item | Parameter |
|---|---|
| Architecture | Lightelligence Tianjic (Optical Matrix Multiplication) |
| Process | TSMC 7nm (photonic + electronic hybrid) |
| Photonic Cores | 64 MZIs (Mach-Zehnder Interferometers) |
| Digital Cores | 32x RISC-V processors (scheduling + activation) |
| HBM | 32GB HBM2E |
| Memory Bandwidth | 819 GB/s |
| FP16 | 1.6 PFLOPS |
| FP32 | ~0.8 PFLOPS |
| INT8 | 3.2 POPS |
| TDP | ~100 W |
| Efficiency | 16 TOPS/W (H100 ~3 TOPS/W, 5x advantage) |
| Form Factor | PCIe Gen4 x16 |
| Interconnect | proprietary optical interconnect (PCIe + fiber) |
| Mass Production | 2023 H2 |
| Unit Price | ~$3,500-5,500 |
Photonic Computing Principle
| Dimension | Traditional GPU | Lightelligence Tianjic |
|---|---|---|
| Compute Medium | Electronic (CMOS transistors) | Photonic (silicon photonic waveguides + MZI) |
| Matrix Multiplication | MAC arrays | MZI mesh (optical propagation matmul) |
| Signal Propagation | Copper wire + resistance | Optical waveguide (zero resistance) |
| Power Source | CV^2 (transistor switching) | MZI modulators only |
| Thermal Design | Complex liquid cooling | Natural air cooling (100W) |
| Frequency | 1-2 GHz | >40 GHz (theoretical) |
| Precision | FP64/FP32/FP16/INT8 | FP16/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
| Dimension | Spec |
|---|---|
| Architecture | RISC-V 64-bit (custom SiFive cores) |
| Core Count | 32 |
| Frequency | 1.5 GHz |
| Cache | L1 32KB + L2 512KB per core |
| Role | Scheduling + activation functions + non-matrix ops |
| Domestic | 100% 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
| Item | Content |
|---|---|
| Company | Lightelligence |
| Founder | Shen Zhihua (Chinese Academy of Sciences Shanghai Institute of Optics + Tsinghua) |
| Founded | 2017 |
| Headquarters | Jiading 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 Projects | National 863 Major Project, National Natural Science Foundation key |
| Fab | TSMC 7nm + proprietary silicon photonics process (CAS Institute of Microelectronics) |
| Patents | 150+ silicon photonics AI patents |
| Customers | Chinese supercomputing centers (Wuxi, Changsha, Shenzhen), government HPC, AI for Science |
Silicon Photonics Big Three
| Company | Country | Product | Compute | TDP | Mass Production |
|---|---|---|---|---|---|
| Lightmatter | US | Envise | 2 PF | 65W | 2023 H2 |
| Lightelligence | China | Tianjic | 1.6 PF | 100W | 2023 H2 |
| LightOn | France | Alfred | 1.2 PF | 80W | 2023 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
Related Products
- Lightmatter Envise - US competitor
- Cambricon MLU 590 - Chinese AI
- Huawei Ascend 910C - strongest Chinese chip
- Moore Threads MTT S5000 - Chinese GPU
- Tenstorrent - RISC-V AI
- NVIDIA H100 - traditional GPU comparison