Empire Technology Development LLC

North Carolina State University

Empire Technologies Development LLC

Mercury Kingdom Assets Limited

Yan Solihin

Ahmad A Samih

Amro Awad

Mazen Kharbutli

Yipeng Wang

Techniques described herein generally include methods and systems related to cache partitioning in a chip multiprocessor. Cache-partitioning for a single thread or application between multiple data sources improves energy or latency efficiency of a chip multiprocessor by exploiting variations in energy cost and latency cost of the multiple data sources. Partition sizes for each data source may be selected using an optimization algorithm that minimizes or otherwise reduces latencies or energy consumption associated with cache misses.

Cache partitioning in a multicore processor

Techniques described herein generally include methods and systems related to circuit switching in a network-on-chip. According to embodiments of the disclosure, a network-on-chip may include routers configured to pre-reserve circuit-switched connections between a source node and a destination node before requested data are available for transmission from the source node to the destination node. Because the circuit-switched connection is already established between the source node and the destination node when the requested data are available for transmission from the source node, the data can be transmitted without the delay or with reduced delay caused by setup overhead of the circuit-switched connection. A connection setup message may be transmitted together with a memory request from the destination node to facilitate pre-reservation of the circuit-switched connection.

Circuit switch pre-reservation in an on-chip network

A processor may comprise a plurality of cores operating at heterogeneous frequencies communicatively coupled by a network of routers also operating at heterogeneous frequencies. A core may be prioritized for thread execution based on operating frequencies of routers on a path from the core to a memory controller. Relatively higher priority may be assigned to cores having a path comprising only routers operating at a relatively higher frequency. A combined priority for thread execution may be based on core frequency, router frequency, and the frequency of routers on a path from the core to a memory controller. A core may be selected based primarily on core operating frequency when cache misses fall below a threshold value.

Core prioritization for heterogeneous on-chip networks

Techniques described herein generally include methods and systems related to designing and operating a DRAM device that has significantly reduced refresh energy use. A method for designing a DRAM optimizes or otherwise improves the DRAM for energy efficiency based on a measured or predicted failure probability of memory cells in the DRAM. The DRAM may be configured to operate at an increased refresh interval, thereby reducing DRAM refresh energy but causing a predictable portion of the memory cells in the DRAM to leak electrical energy too quickly to retain data. The DRAM is further configured with a selected number of spare memory cells for replacing the 'leaky' memory cells, so that operation of the DRAM at the increased refresh interval may result in little or no reduction in capacity of the DRAM.

Increased refresh interval and energy efficiency in a DRAM

Methods and systems to assign threads in a multi-core processor are disclosed. A method to assign threads in a multi-core processor may include determining data relating to memory controllers fetching data in response to cache misses experienced by a first core and a second core. Threads may be assigned to cores based on the number of cache misses processed by respective memory controllers. Methods may further include determining that a thread is latency-bound or bandwidth-bound. Threads may be assigned to cores based on the determination of the thread as latency-bound or bandwidth-bound. In response to the assignment of the threads to the cores, data for the thread may be stored in the assigned cores.

Thread and data assignment in multi-core processors based on cache miss data and thread category

Techniques described herein generally include methods and systems related to dynamic cache-sizing used to reduce the energy consumption of a DRAM cache in a chip multiprocessor. Dynamic cache sizing may be performed by adjusting the refresh interval of a DRAM cache or by combining way power-gating of the DRAM cache with adjusting the refresh interval.

Methods and systems for dynamic DRAM cache sizing

Technologies are generally described manage MRAM cache writes in processors. In some examples, when a write request is received with data to be stored in an MRAM cache, the data may be evaluated to determine whether the data is to be further processed. In response to a determination that the data is to be further processed, the data may be stored in a write cache associated with the MRAM cache. In response to a determination that the data is not to be further processed, the data may be stored in the MRAM cache.

Magnetoresistive random-access memory cache write management

A computing system to compress an array using hardware-based compression and to perform various instructions on the compressed array is generally described. The computing system may receive an instruction adapted to access an address in an array. The computing system may determine whether the address is compressible. If the address is compressible, then the computing system may determine a compressed address of a compressed array based on the address. The compressed array may represent a compressed layout of the array, where a reduced size of each compressed element in the compressed array is smaller than an original size of each element in the array. The computing system may access the compressed array at the compressed address in accordance with the instruction.

Hardware-based array compression

Techniques described herein generally include methods and systems related to improving energy efficiency in a chip multiprocessor by reducing the energy consumption of a DRAM cache for such a multi-chip processor. Methods of varying refresh interval may be used to improve the energy efficiency of such a DRAM cache. Specifically, a per-set refresh interval based on retention time of memory blocks in the set may be determined, and, starting from the leakiest memory block, memory blocks stored in the DRAM cache that are associated with data also stored in a lower level of cache are not refreshed.

Energy-efficient dynamic dram cache sizing via selective refresh of a cache in a dram

Technologies are generally described for methods and systems effective to implement a memory allocation accelerator. A processor may generate a request for allocation of a requested chunk of memory. The request may be received by a memory allocation accelerator configured to be in communication with the processor. The memory allocation accelerator may process the request to identify an address for a particular chunk of memory corresponding to the request and may return the address to the processor.

Memory allocation accelerator

A cache coherence mechanism may comprise a bit-to-cache map for processor cores operable up to a maximum frequency for cores of a multicore processor. Entries in a cache coherence directory may include a bit field identifying cores operable at or near the maximum frequency that share a memory block corresponding to the entry. An additional field may indicate sharing by cores operating at lower frequencies. The additional field may be indicative of the bit-field corresponding to a bit-to-cache map representative of cores other than those operating at or near the maximum frequency.

Directory coherence for multicore processors

Technologies are generally described for methods and systems effective to store data in a memory module. The memory module may include a volatile portion and a non-volatile portion. The methods may comprise receiving, by a processor, a request to store the data. The request may include an indication of a virtual address. The methods may further include determining, by the processor, a persistency of the data based on the virtual address. The methods may further include performing a first operation of identifying a particular portion of the memory module based on the virtual address. The methods may further include generating a command to store the data in the particular portion of the memory module. The methods may further include controlling the operating system to perform a second operation of updating a translation lookaside buffer to indicate the persistency of the data.

Data storage based on memory persistency

Technologies are generally described for systems, devices and methods relating to multicore processors. The multicore processors may include first and second tiles with first and second caches, respectively. The first cache may include first magnetoresistive random access memory (MRAM) cells with first storage characteristics. The second cache may include second MRAM cells with second storage characteristics different from the first storage characteristics. In some examples, an interconnect structure may be coupled to the first and second tiles and may be configured to provide communication between the first tile and the second tile. Methods for handling migration between tiles and cores are also described.

Heterogeneous magnetic memory architecture

Systems and methods for modeling memory access behavior and memory traffic timing behavior are disclosed. According to an aspect, a method includes receiving data indicative of memory access behavior resulting from instructions executed on a processor. The method also includes determining a statistical profile of the memory access behavior, the profile including tuple statistics of memory access behavior. Further, the method includes generating a clone of the executed instructions based on the statistical profile for use in simulating the memory access behavior.

Growing community of inventors

Raleigh, NC, United States of America

Yan Solihin