A5003878521- Recommender Systems and Techniques
- Stochastic Gradient Optimization Techniques
- Radiation Effects in Electronics
- Parallel Computing and Optimization Techniques
- Advanced Data Storage Technologies
- Distributed systems and fault tolerance
- Advanced Bandit Algorithms Research
- Advanced Graph Neural Networks
- Caching and Content Delivery
- Advanced Malware Detection Techniques
- Data Quality and Management
- Cryptographic Implementations and Security
- Cloud Computing and Resource Management
- Security and Verification in Computing
- Web Data Mining and Analysis
- VLSI and Analog Circuit Testing
- Low-power high-performance VLSI design
- Scientific Computing and Data Management
- Software Reliability and Analysis Research
- Distributed and Parallel Computing Systems
- Machine Learning in Materials Science
Advanced Micro Devices (United States)
2023
Harvard University Press
2018-2021
Northeastern University
2014
Neural personalized recommendation models are used across a wide variety of datacenter applications including search, social media, and entertainment. State-of-the-art comprise large embedding tables that have billions parameters requiring memory capacities. Unfortunately, fast DRAM-based memories levy high infrastructure costs. Conventional SSD-based storage solutions offer an order magnitude larger capacity, but worse read latency bandwidth, degrading inference performance. RecSSD is near...
Modern smart NICs provide little isolation between the network functions belonging to different tenants. These also do not protect from datacenter-provided management OS which runs on NIC. We describe concrete attacks allow a function's state leak (or be modified by) another function or OS. then introduce S-NIC, new hardware design for that provides strong guarantees. S-NIC pervasively virtualizes accelerators, enforces single-owner semantics each line in on-NIC cache and RAM, dedicated bus...
Reliability is an important design constraint in modern microprocessors, and one of the fundamental reliability challenges combating effects transient faults. This requires extensive analysis, including significant fault modelling allow architects to make informed tradeoffs. Recent data shows that multi-bit faults are becoming more common, increasing from 0.5% static random-access memory (SRAM) 180nm 3.9% 22nm. Such predicted be even prevalent smaller technology nodes. Therefore, accurately...
Deep learning recommendation systems must provide high quality, personalized content under strict tail-latency targets and system loads. This paper presents RecPipe, a to jointly optimize quality inference performance. Central RecPipe is decomposing models into multi-stage pipelines maintain while reducing compute complexity exposing distinct parallelism opportunities. implements an scheduler map engines onto commodity, heterogeneous platforms (e.g., CPUs, GPUs). While the hardware-aware...
Deep learning based recommendation systems form the backbone of most personalized cloud services. Though computer architecture community has recently started to take notice deep inference, resulting solutions have taken wildly different approaches - ranging from near memory processing at-scale optimizations. To better design future hardware for we must first systematically examine and characterize underlying systems-level impact decisions across levels execution stack. In this paper, eight...
The pace of advancement the top-end supercomputers historically followed an exponential curve similar to (and driven in part by) Moore's Law. Shortly after hitting petaflop mark, community started looking ahead next milestone: Exascale. However, many obstacles were already looming on horizon, such as slowing Law, and others like end Dennard Scaling had arrived. Anticipating significant challenges for overall high-performance computing (HPC) achieve 1000x improvement, U.S. Department Energy...
Reliability is a significant design constraint for supercomputers and large-scale data centers. Modeling the effects of faults on applications targeted to such systems allows system architects software designers provision resilience features, that improve fidelity results reduce runtimes. In this paper, we propose mechanisms existing techniques model effect transient realistic applications. First, extend Program Vulnerability Factor metric multi-threaded Then demonstrate how measure PVF an...
Neural personalized recommendation models are used across a wide variety of datacenter applications including search, social media, and entertainment. State-of-the-art comprise large embedding tables that have billions parameters requiring memory capacities. Unfortunately, fast DRAM-based memories levy high infrastructure costs. Conventional SSD-based storage solutions offer an order magnitude larger capacity, but worse read latency bandwidth, degrading inference performance. RecSSD is near...
Deep learning based recommendation systems form the backbone of most personalized cloud services. Though computer architecture community has recently started to take notice deep inference, resulting solutions have taken wildly different approaches - ranging from near memory processing at-scale optimizations. To better design future hardware for we must first systematically examine and characterize underlying systems-level impact decisions across levels execution stack. In this paper, eight...
Deep learning recommendation systems must provide high quality, personalized content under strict tail-latency targets and system loads. This paper presents RecPipe, a to jointly optimize quality inference performance. Central RecPipe is decomposing models into multi-stage pipelines maintain while reducing compute complexity exposing distinct parallelism opportunities. implements an scheduler map engines onto commodity, heterogeneous platforms (e.g., CPUs, GPUs).While the hardware-aware...