Photonic AI Accelerator – Envise
CLEO/Europe and the World of Photonics Congress
We are grateful to the EPSRC (UK) for funding this work under Grant No.
**Publisher\’s note:** Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This research was partially supported by the European Regional Development Fund (ERDF E-Infrastructure) and the Greek National Funds through the Operational Program Education and Lifelong Learning, under Project “Fundamental and Applied Nanomaterials of Functional Materials” (EFOP-3. 3-CM-ONLUS) hosted by the ERDF of the Turkish Higher Education Authority.
![*G*~2~ images with a 2.
(**a**) Photon induced emission spectra of Tb^3+^-doped (red circles) and non-doped (black open circles) CdS. (**b**) Cross-sections for the red and the green spectral lines shown in (**a**) in the *T* = 3 K, 60 K (black dotted line) and 300 K (red line) thermal field. (**c**) Simulated cross-section of the spectral line of the (**a**) CdS sample and its fit to Lorentzian. (**d**) Simulated cross-section of the green spectral line in the *T* = 3 K, 60 K (black dotted line) and 300 K (red line) thermal field. The dashed line is a fit to Lorentzian. (**e**) Simulated *G*~2~ image for the CdS (red) and CdS (green) samples with 2. The insets show the corresponding cross-sections.
Envise: A Photonic AI Accelerator
Envise is our photonic AI accelerator. It has a general purpose—it is not just for image recognition or natural language processing; it’s general across AI. Its very high performance in terms of output and very low energy consumption. Our goal with this study is to try really to help with scaling AI and its minimizing of environmental footprint. If you look at chips that are out there right now, the Nvidia chip draws about 450 watts—that’s an insanely hot computer chip. Our chip is about 80 watts and is multiple times faster.
We are thrilled to introduce Envise: a new photonic accelerator for machine learning. Envise is a self-contained modular, autonomous AI research reactor that combines an ultra-high-powered laser, a cloud-based networking stack and digital signal processing hardware. Envise uses a variety of techniques for photonic signal processing to enhance the performance of AI, including laser-powered, continuous-wave self-phase modulation (CSPM) frequency-modulation, self-phase modulation non-reciprocating optical amplitude-shift keying, and time-to-digital converters.
Overcoming latency issues.
We found that the latency in latency-sensitive machine learning models is typically a problem for the machine learning platform, particularly for larger models and deep learning models.
For the first time, we have come up with a way to alleviate this latency, and that is with the use of self-phase modulation on the data stream coming from the neural network during the training process.
We have tried to eliminate the latency by using the phase of the current phase-locked loop (PLL) to generate frequency-modulation to increase the phase coherence, and in doing that we are leveraging the inherent nature of the PLL to be able to achieve very high-precision of frequency and phase.
This concept is applied to every type of neural network that is used in AI, such as the convolutional neural network, the recurrent neural network, the multi-layer neural network, the hierarchical deep learning network, the fully connected neural network.
We have developed a novel concept called phase-locked loop with the use of an analog phase shifter that is placed in between the digital inputs and the analog outputs of the neural network of interest.
In this way, all the signals are phase-locked to that particular PLL so that the frequency modulation is generated in the PLL at the exact same time the signals from the neural network are injected into the PLL.
Leave a Comment