International Symposium on Microarchitecture 2018

4 papers
The 2018 International Symposium on Microarchitecture featured research focused on specialized hardware acceleration, network reliability, and architectural security. Several papers addressed the optimization of deep learning workloads. For instance, the MAESTRO framework was introduced as an open-source tool for analyzing dataflows in deep learning accelerators, while the EcoRNN project focused on improving the efficiency of Long Short-Term Memory (LSTM) networks on GPUs. Other research explored fundamental system design challenges and security. One study proposed methods for ensuring deadlock freedom in arbitrary network topologies through packet swaps. Additionally, the symposium addressed computer security by investigating the use of lazy, stackless functional programming architectures to mitigate vulnerabilities. Prolific authors contributing to these themes included Hyoukjun Kwon, Mayank Parasar, and Bojian Zheng. These works reflect a trend toward domain-specific architectures and the integration of security considerations into hardware design.

Papers

Title
EcoRNN: Efficient Computing of LSTM RNN on GPUs
Guaranteeing Deadlock Freedom in Arbitrary Network Topologies using Packet Swaps
MAESTRO: An Open-source Infrastructure for the Cost-Benefit Analysis of Dataflows within Deep Learning Accelerators
Tackling Computer Security Issues Through Lazy, Stackless Functional Programming Architectures