Massively Parallel AI unlocked now.
LightOn OPU: the Photonic AI Chip at scale

We develop photonic hardware called Optical Processing Units – or OPUs – of interest to Artificial Intelligence (AI).

The technology boosts some generic tasks in Machine Learning such as training and inference of high-dimensional data. It can be used in the context of supervised and unsupervised learning, with batch processing or streaming data.

OPUs are highly integrated with CPUs and GPUs so that it boosts their respective performance.

OPUs can be seamlessly accessed through an Open Source Python API called LightOnML and through a lower level, proprietary API called LightOnOPU.

Benchmarks to compare the performance of CPU and GPU with our OPU are available here: https://github.com/lightonai/opu-benchmarks

The OPU is the first large-scale hybrid digital / analog computing platform for AI.

Nitro, the first generation Photonic Core

The brain of the current OPUs is Nitro, the first generation Photonic Core – Nitro drives the Aurora 1.5 available in LightOn Cloud.

The Nitro photonic core implements specific matrix-vector multiplications in a massively parallel fashion, where inputs and outputs of dimension up to one million are processed at once, within a single clock cycle.

More precisely, a Nitro photonic core performs Random Projections at the equivalent of 2.1012 hybrid-precision MAC (multiply-accumulate) operations every clock cycle.

The Nitro photonic core thus yields a stunning performance of 3 PetaOP/S with only 30 Watts of power or about 100 TeraOP/J.

Use cases
Computer Vision

Why is Fast training of image classification models important?

 

  • Data Science team spends less time in training image classification models.
  • Allows the use of lower precision arithmetics for training/inference to further reduce training/test time.
  • Re-training SOTA models with few data is essential for businesses.
  • For data scientists: More iterations possible on higher-level tasks.

Up to 10x faster than backprop on GPU!

Background:

  • Transfer Learning applied on 2D CNNs such as VGG, Densenet, Resnet models for image classification
  • Library: LightOnML
  • Dataset: STL10, Skin Cancer, Flowers and other image datasets
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Why is Fast training of Video classification models important?

 

  • Training 3D CNNs is extremely time and energy-consuming.
  • Can get around the huge memory requirements of CNNs.
  • For Data scientists: More iterations possible on higher-level tasks.
  • Much lower variance to the hyperparameters’ choice.

Training time: Up to 9x faster than backprop on GPU at the same level of accuracy!

Background:

  • Transfer learning on 3D CNNs such as I3D for video action recognition
  • Library: LightOnML
  • Dataset: HMDB51, UCF101 and other action recognition video datasets
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Why is Fast training of simple models important?

  • State-of-the-art performance with theoretical guarantees in some tasks.
  • For Data scientists: More iterations possible on higher-level tasks.

1.3x to 23.6x faster extrapolating the curves to 1.000.000

Background:

  • Kernel ridge regression approximation for classification tasks.
  • Library: LightOnML
  • Dataset: qm7 (quantum chemistry), high energy physics data and others (image classification)
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Natural Language Processing

Why is Fast NLP important?

 

  • Data Science team spends less time in building NLP models and get results.
  • Re-training SOTA models with few data is essential for businesses.
  • For data scientists: more iterations possible on higher-level tasks.

Analysis of the French National Grand Débat data

Background:

  • NLP applied on Bag Of Random Embedding Projections (BOREP)
  • Library: LightOnML
  • Dataset: Grand Débat dataset
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Ranking & Recommendations

Why is important?

 

  • To offer a quick and easy baseline for large scale recommender systems.

MovieLens 20M database: 27000 movies x 138000 users, with 0.5%non-zero entries

 

Background:

 

Time Series

Why is Change detection in Molecular Dynamics (MD) simulation important?

 

  • MD simulations are used in drug design, new materials discovery

     

  • Applying an intelligent layer on top of HPC simulation enables metadynamics
  • 15x faster than FastFood on CPU at 50k atoms!
  • OPU enables analysis of samples containing a very large number of atoms due to overcoming the memory bottleneck of traditional architectures.
  • For 700k + atoms, NEWMA RP on OPU is expected to be 30x faster than NEWMA FF on CPU!

Background:

  • No-prior-knowledge EWMA: https://arxiv.org/abs/1805.08061
  • Library: LightOnML
  • Dataset: Molecular Dynamics simulations (HPC, Anton)
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Why is important?

 

  • Detect in real-time changes, without the need to store the whole history of the graph

  • Applications in community detection, fraud detection, biology and others

Reduced memory requirements and facilitating the analysis of Very large datasets

Background:

  • No-prior-knowledge EWMA: https://arxiv.org/abs/1805.08061
  • Library: LightOnML
  • Dataset: Facebook graph datasets, any time-evolving graph
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Machine Learning Techniques

Why is the training of Neural Networks with an OPU important?

 

  • Enables OPUs to become a cornerstone in the training process-optical training.

     

  • Great potential for faster training/larger models.

Demonstrated on a large range of tasks (but CNNs!) – Optical training demonstrated on MNIST

Background:

  • Background work: Direct Feedback Alignment https://arxiv.org/abs/1609.01596
  • Library: no dedicated library yet

  • Dataset: widely applicable (RecSys, graphs, NLP, etc.)

  • Next-generation photonic core Lapis Lazuli 1.5 (Available on LightOn Cloud by Q4 2020)

 

Deeper Insight

ArXiv

GitHub

Why is the fast class-aware low-D embedding of data
important?

  • Dimensionality reduction makes the whole processing pipeline faster.

  • Using information from the labels produces more useful embeddings.

  • It is a scalable method.

1.2 to 4x speedup at label dimensionality 100k to 700k

Background:

  • Supervised Random Projections from https://arxiv.org/abs/1811.03166
  • Library: LightOnML
  • Dataset: dimensionality reduction benchmark datasets
  • Nitro photonic core, Aurora 1.5 (LightOn Cloud)

 

Machine Intelligence

Why is important?

 

  • It can tackle high-dimensional control problems in robotics or trading with a more efficient “memory” for exploitation.
  • Accelerate early stages of learning of an agent with imitation learning.

Background:

 

A Deeper Insight

The Nitro photonic core leverages light scattering to perform a specific kind of matrix-vector operation called Random Projections.

Random Projections have a long history for the analysis of large-size data since they achieve universal data compression. In other words, you can use this tool to reduce the size of any type of data, while keeping all the important information that is needed for Machine Learning. There are well-established mathematical guarantees on why this works, in particular thanks to a result known as the Johnson-Lindenstrauss lemma.

Essentially, this states that compression through Random Projections approximately preserves distances: if two items are similar in the original domain, their compressed version will also be similar.

Going from mathematics to various fields of engineering, Random Projections have proved useful in a wide variety of application domains such as Compressed Sensing, Randomized Numerical Linear Algebra, Hashing, Streaming and Sketching algorithms. More recently, they have naturally found their way into Machine Learning, for instance in Deep Neural Networks, Reinforcement Learning or Reservoir Computing.

It’s important to stress that such Random Projections can be applied to any type of data: images or videos, sound, financial or scientific data or more abstract features … anywhere where one is “drowned in data but starved for wisdom”.

Besides developing the hardware, LightOn AI Research team investigates numerous use-cases for our unique digital-analog hybrid technology. We communicate to the science and tech community through journal papers, preprints, conference presentations, blog postsworkshops and meetups. We also have detailed examples in our API documentation pages and GitHub account. Have you done anything mind-blowing on LightOn Cloud? Let us know and we’d love to talk about it!

Stay connected
Subscribe for LightOn’s news, every month, in your inbox!
Subscribe