Project case study · AI Engineering / Neuromorphic Computing
Hardware-aware GPT models for photonic computing
A hardware-aware causal language modelling pipeline that simulates photonic/electronic constraints inside selected linear layers of GPT-style models. The goal was to investigate how quantization and analog noise affect the viability of language models on alternative computing hardware.
By Berke Katrancı
- My role
- AI engineering, machine learning and neuromorphic computing research
- Context
- Hardware-aware simulation of GPT-style language models
- Status
- Finished
The research question
Can GPT-style language models tolerate hardware-inspired constraints from photonic/electronic neuromorphic systems? The project explored this question through simulation, examining model behaviour beyond conventional digital accelerator assumptions.
The broader motivation was to evaluate whether language models could remain workable under constraints relevant to future energy-efficient computing substrates.
What the simulation models
I designed and implemented a causal language modelling pipeline that applies hardware-inspired effects inside selected linear layers during training. The implementation used PyTorch, Hugging Face Transformers and small GPT-2-style models.
- Limited spatial light modulator (SLM) weight precision, represented through weight quantization.
- microLED/digital-to-analog converter (DAC) activation quantization.
- Analog-to-digital converter (ADC) readout quantization at the layer output.
- Analog noise introduced alongside the quantization effects during training.
Findings from the tested setup
Experiments with small GPT-2-style models showed that 8-bit hardware-aware settings remained workable in the tested setup. More aggressive low-bit quantization caused significant degradation.
The results also suggested that feed-forward network layers may be a promising initial target for future photonic AI acceleration. This is a direction indicated by the experiments, rather than a demonstrated hardware deployment.
Scope and limitations
These findings come from a hardware-inspired software simulation with small GPT-style models. They do not establish measured energy savings, physical-device performance or equivalent behaviour in larger language models.
The bit-width observations are specific to the tested model and simulated constraints. Extending the conclusions would require further experiments and validation against physical hardware.
Discuss this project
Get in touch to discuss the approach, my contribution or related AI and software work.
Contact Berke