Lightmatter: A New Kind of Computer — Photonic Processor Running Production Transformers
First photonic processor to run unmodified transformers + CNNs + RL at near-32-bit-float accuracy — no fine-tuning, no quantization-aware training. 65.5 TOPS (ABFP16) at 78W electrical + 1.6W optical, with 3D-integrated six-chip design using ~1M photonic components
Lightmatter: A New Kind of Computer
Lead
Lightmatter demonstrated a hybrid photonic-electronic processor capable of running production-grade neural networks without modification — no fine-tuning, no quantization-aware training, no architecture changes. Achieves accuracy "approaching 32-bit floating-point digital systems" on transformers (BERT), CNNs (ResNet classification + segmentation), and Atari reinforcement learning. Six 3D-integrated chips, ~1 million photonic components, 65.5 TOPS at 78W electrical + 1.6W optical.
Key Claims
- Runs unmodified transformers + CNNs + RL — standard BERT, standard ResNet, standard Atari DQN; no model surgery.
- Near-32-bit-float accuracy out of the box — the threshold that defines "production-ready" vs "research demo."
- 65.5 trillion ABFP16 operations/second.
- 78W electrical + 1.6W optical total power.
- ~1M photonic components across six 3D-integrated chips.
- PyTorch + TensorFlow compatible — works with existing ML frameworks.
Architectural Approach
- Hybrid photonic-electronic — photonic tensor cores for matmul, electronic control + memory.
- Silicon photonics — standard process, not exotic materials.
- 3D vertical integration across six chips.
- Photonic tensor cores handle the compute-dense matmul; electronic layer handles what it does well.
Demonstrated Workloads
| Workload | Task |
|---|---|
| BERT | Transformer inference |
| ResNet | Image classification |
| ResNet (segmentation) | Semantic segmentation |
| Deep RL | Atari game-playing |
Novelty vs Prior Photonic Work
- First complex/real-world workloads — most prior photonic demos used MNIST-scale toy problems.
- Practical precision without simplified benchmarks — ABFP16 achieves near-FP32 accuracy without calibration tricks.
- Full system integration — not isolated components on an optical bench.
- Framework compatibility — PyTorch/TensorFlow work out of the box.
Why This Matters
Companion to the two 2026 Nature papers in this topic (photonic tensor processor and Ashtiani on-chip backprop). Lightmatter is the commercial endpoint of the research direction those papers establish: if you can run BERT without modification at near-FP32 accuracy, you have an inference-as-a-service offering that competes with GPU inference on throughput per watt, where photonics' ~20× energy advantage could justify significant silicon cost.
The critical gap is training — Lightmatter's chip is inference-only. Ashtiani et al. (Nokia Bell Labs) solve training on-chip but at toy scale. The 2027-2028 question is whether a production-scale photonic chip can also train, or whether inference-only remains the right product.
Limitations (Inferred)
- Blog claims, not peer-reviewed data; specific accuracy numbers on each workload not disclosed.
- Inference only; no training capability.
- 65.5 TOPS is modest vs GPU TFLOPS — competitive on energy/op, not raw throughput.
- ~1M photonic components is an engineering feat, but yield and reliability at scale not reported.
Source: A New Kind of Computer, Lightmatter, April 9 2025.