… 2 IOI DESIGN We propose IOI as a next-generation inference system using a novel transceiver
module with optical computation capabilities, combined with the power of programmable …

… IOI [11] has provisioned optical matrix multiplication in the PDP hardware switch to provide
low-latency machine learning inference… in-network computation of neural network inference in …

… In this section, we describe the most relevant photonic computing primitives for in-network …
learning inference on smartNICs. Similarly, IOI [70] proposed processing inference queries …

… We present pForest, a system for “As Soon As Possible” (ASAP) in-network classification …
We implemented a prototype of pForest in Python (training) and P4 (inference). Our evaluation …

… inference requests at the record-breaking 4.055 GHz with emerging photonic computing [1…
the IOI smart transceiver with the Tofino programmable switch to perform in-network optical …

… switches with smart transceivers, enabling in-network optical inference (15). Modern network
… switches to perform layer-by-layer inference with smart transceivers (15). The large data …

… inference, this work is a first step towards a long-term vision of … IOI and NetCast introduced
a smart transceiver module with … Lightning for in-network inference use cases to future work. …

… Taurus [59] and IOI [70] use modified ASIC to realize … the in-network inference. The results
show that the EB models consume less memory and M/A stages, but lead to longer inference …

… present an adaptive in-network inference mechanism to control the trade-off between accuracy
and inference load. To achieve this, we adopt a widely-used adaptive inference method, …

… The architecture consists of a server and a client, connected by an optical link. Since linear
algebra is the bottleneck step for DNN inference and training, we focus on how the optical …