Skip to main content

Lightmatter Envise (Silicon Photonics AI Inference Chip)

Product Overview

Lightmatter Envise is the world's first Silicon Photonics AI inference chip, 2022-11 released, 2023 mass production. Based on 3nm process, integrating Photonic Compute Cores + MIPS RISC-V processors + silicon photonic interconnect. 2 PFLOPS FP16 compute (65W TDP), one of the most power-efficient AI inference chips (30 TOPS/W, 5-10x higher than NVIDIA H100). Paired with an 8-card system at 16 PFLOPS + optical interconnect.

Revolutionary significance: first AI chip to replace electronic matrix multiplication with photonics, speed of light = 0 latency + 0 power consumption (theoretical), actual 80-90% power savings. Represents a paradigm shift in AI chip architecture, known alongside Lightelligence and LightOn as the "Silicon Photonics Big Three".

Core Specs

ItemParameter
ArchitectureLightmatter Passage Silicon Photonics (opto-electronic hybrid)
ProcessTSMC 3nm
Photonic Cores4 Photonic Tensor Cores (each 4x4 array)
Digital Cores64x MIPS RISC-V processors
HBM64GB HBM3
Memory Bandwidth1.6 TB/s
FP162 PFLOPS
FP32~1 PFLOPS
INT84 POPS
TDP65 W (one of the lowest AI inference TDP in industry)
Efficiency30 TOPS/W (H100 ~3 TOPS/W, 10x advantage)
Form FactorOAM
InterconnectLightmatter Passage optical interconnect (PCIe + fiber)
Mass Production2023 H2
Unit Price~$5,000-8,000 (estimated)

Photonic Computing Principle

DimensionTraditional GPULightmatter Envise
Compute MediumElectronic (CMOS transistors)Photonic (silicon photonic waveguides + modulators)
Matrix MultiplicationMAC arrays (transistor switching)Mach-Zehnder Interferometer (MZI) mesh
Signal PropagationCopper wire + resistanceOptical waveguide + all-optical propagation
LatencyNanosecond-level0 latency (speed of light)
PowerQuadratic law (CV^2)Linear (modulator only)
Frequency1-2 GHz>100 GHz (theoretical)
Thermal DesignComplex liquid coolingNatural air cooling (65W)
Drawback-Not general-purpose, only for matmul (activation functions still need digital)

MZI Matrix Multiplication

Input vector (4-dim): [x0, x1, x2, x3] (optical signals)
Weight matrix W (4x4): configured via MZI mesh
Output: y = W x x (optical signal superposition)
ADC: convert to digital (the only electronic step)

Key advantages:

  • Matrix mul 0 power (light itself)
  • Only ADC/DAC consume energy (10% of total compute)
  • Reconfigurable: MZI mesh is programmable (same hardware for different matrices)

MIPS RISC-V Processors

DimensionSpec
ArchitectureMIPS RISC-V compatible (64-bit)
Core Count64 cores (multi-threaded)
Frequency2 GHz
CacheL1 32KB + L2 1MB per core
RoleScheduling + activation functions + preprocessing
RationaleMIPS lower power than ARM Cortex-A78 (4x cores, more power efficient)

Why MIPS RISC-V: Silicon photonic cores only handle matmul, non-matrix ops (softmax, layernorm, attention) still require digital processors. MIPS vs ARM same perf, 30% power savings.

8-Card System 16 PFLOPS

ItemConfiguration
Envise Cards8
Total Compute16 PFLOPS FP16
Total HBM512GB
Optical Interconnect BW1.6 TB/s intra-domain (Passage)
Rack TDP~520 W (8 cards only)
Rack Form Factor1U / 2U server
Price~$50-70K (8-card system)

vs NVIDIA H100 8-card:

  • Compute 16 PF vs 32 PF (H100 FP8 dense 2 PF x 8)
  • Power 520W vs 5600W (H100 700W x 8) -- Envise 10x advantage
  • Efficiency 30 TOPS/W vs 3 TOPS/W -- 10x advantage
  • Latency 50% lower (optical interconnect)

Vendor Information

ItemContent
CompanyLightmatter, Inc.
FounderNicholas Harris (MIT PhD)
Founded2017
HeadquartersBoston, Massachusetts, USA
Funding$420M+ (Series D 2023-Q2 led by: USIT + Google Ventures)
Valuation (2025)~$1.2B (unicorn)
2025 Revenue~$30M (early commercialization)
CustomersGoogle internal testing, Meta, Anthropic, government HPC
CEONicholas Harris
CTODarius Bunandar
FabTSMC 3nm + AIM Photonics (silicon photonics process)
Patents200+ silicon photonics AI patents

Lightmatter Product Line

ProductPositioningComputeReleased
EnviseAI inference chip2 PF FP162022-11
Passage M1000optical interconnect chip (PCIe Gen5 fiber)1.6 TB/s intra-domain2022-11
Idiomsoftware stack (PyTorch / TensorFlow compatible)-continuously updated
Envise X (estimated)2nd-gen silicon photonics, 5 PF FP165 PF2026 H2 estimated

Software Stack Idiom

LayerToolDescription
AI frameworkIdiomPyTorch 1:1 compatible (auto-maps matmul to photonics)
TensorFlow (experimental)compatible
CompilerLightflow Compilermatrix ops -> silicon photonics config + MIPS code
RuntimeIdiom Runtimephotonic + digital core co-scheduling
Model ZooModelZooLLaMA / Mistral / Qwen / SDXL

Warning: Ecosystem limitation: Silicon photonics only handles matmul, softmax / attention / layernorm etc handled by MIPS. Models must go through Idiom compiler optimization. PyTorch models run directly at 50-70% lower performance, requiring manual lightmatter.optimize(model).

Use Cases

  • Low-power AI inference (data center PUE optimized)
  • Edge AI (65W air-cooled, no liquid cooling needed)
  • Green AI (carbon-neutral target enterprises)
  • LLM inference (2 PF FP16 fits 70B FP16 + KV Cache)
  • Government/research HPC (US DOE, NSA interest)
  • AI training (Envise inference only, training needs GPU)
  • CUDA proprietary workloads (requires Idiom porting)
  • Non-matmul tasks (e.g., RNN weak performance)

Lightmatter vs NVIDIA H100

MetricLightmatter EnviseNVIDIA H100
Compute FP162 PF2 PF FP8 sparse
TDP65W700W
Efficiency30 TOPS/W3 TOPS/W
Memory64GB HBM380GB HBM3
Latency0 latency matmulnanosecond-level
Training SupportNoYes
EcosystemIdiom (new)CUDA (mature)
Production Maturityearly mass productionfully mature
Price~$5-8K~$25-30K

Envise killer feature: TDP 65W vs H100 700W = 10x efficiency advantage, deployment cost (rack density + cooling) far below H100. It is one of the best HW solutions for hyperscale AI inference.

Silicon Photonics AI Big Three

CompanyCodenameComputeTDPMass Production
LightmatterEnvise2 PF65W2023 H2
LightelligenceTianjic1.6 PF~100W2022 pilot
LightOnAlfred1.2 PF80W2023 experimental

Key Features

  • Silicon Photonics Computing: first commercial photonic AI inference chip
  • 30 TOPS/W: H100 10x efficiency advantage
  • 65W TDP: air-cooled deployment, no liquid cooling
  • MIPS RISC-V: 64-core digital processor for non-matrix ops
  • Optical Interconnect: Passage 1.6 TB/s intra-domain
  • Drawbacks: inference only, no training support, new ecosystem