for our open-source software please visit the "VAJRA-main" directory
We have built the "VAJRA" an edge server for computing deep learning workloads which are usually heavy to deploy at the extreme edge devices of the network. VAJRA mimics an Edge-Cloud scenario, operating between edge devices and the cloud, co-optimizing latency and throughput.
Hardware: VAJRA hosts a Raspberry Pi-5 as a master node, that acts like a gateway router to the 3 Intel DE-10 SoC FPGAs. The aim is to scale the number of parameters for resource-constrained edge devices, by employing HPC techniques at the Edge.
Figure-1: VAJRA System
Figure-2: System Architecture
Team VAJRA is building a lightweight scheduler on the master node. This will aim to co-optimize latency and throughput.
Team VAJRA focuses on delivering reconfigurable compute facilities to users to elevate performance at the edge. We aim to deploy openCL-based kernels on the device to ensure Power-Time-Memory are optimized throughout the inference cycle.
VMA enables model parallelism for deep learning on our multi-node system. It optimally analyzes the memory requirement for each model, and optimally partitions it to fit on the system. We scale our parameters up to 400 million parameters using the VMA for VAJRA.
