University of Pittsburgh

Hefei University of Technology

Jun Yang

Rami G Melhem

YouTao Zhang

Ziyu Zhang

Debashis Ganguly

Shuai Ding

Rujia Wang

Lei Zhao

A data management method wherein a working set is distributed between near and far memories includes migrating first data from the far to the near memory according to a prefetcher algorithm. The first data (a subset of the working set) is maintained in the near memory in data structures according to predetermined semantics of the prefetcher that dictate that certain of the first data is prefetched when a first function evaluates as true. The method further includes detecting that the near memory has reached capacity, and in response, adaptively migrating a portion of the first data out of the near and into the far memory according to an eviction algorithm that is based on the set of prefetcher semantics such that certain of the portion of the first data is evicted when a second function evaluates as true, wherein the second function equals the inverse of the first function.

Method and apparatus for replacing data from near to far memory over a slow interconnect for oversubscribed irregular applications

A method of managing data during execution of an application for use in a system that includes a host memory, a near memory, and a near device associated with the near memory. The application uses a working set of data that is distributed between the far memory and the near memory. The method includes counting a number of times that the near device accesses a unit of the working set of data from the far memory, determining whether the number of times exceeds a dynamically changing access counter threshold, wherein the dynamically changing access counter threshold is calculated dynamically based on a static threshold that is set for the system, and responsive to determining that the number of times exceeds the dynamically changing access counter threshold, migrating the unit of data from the far memory to the near memory.

Method and apparatus for adaptive page migration and pinning for oversubscribed irregular applications

A method of deploying a neural network on a target device method includes extracting a number of device dependent error masks from the target device, using the number of device dependent error masks in a training phase of the neural network to generate a device dependent weight matrix for the neural network, and storing the device dependent weight matrix in the target device for use by the target device to perform inference tasks using the neural network.

System and method of deploying an artificial neural network on a target device

An apparatus including (i) a processor including a plurality of main buffer on board (BOB) memory controllers (MCs) and a secure engine, (ii) a plurality of simple BOB MCs, (iii) a secure delegator, and (iv) a plurality of memory modules. The secure delegator coupled to a first main BOB MC and a first simple BOB MC creates a secure channel. A second main BOB MC coupled to a second simple BOB MC creates a non-secure channel. The plurality of main BOB MCs, the secure engine and the secure delegator are provided within a trusted computing base (TCB) of the apparatus and the plurality of simple BOB MCs and the plurality of memory modules are provided outside the TCB. The secure delegator is configured to: (i) secure communication between the first main BOB MC and the secure delegator, and (ii) perform Path ORAM accesses to the plurality of memory modules.

Growing community of inventors

Pittsburgh, PA, United States of America

Jun Yang