Advanced Micro Devices Corporation

T-head (shanghai) Semiconductor Co., Ltd.

Advanced Micro Devices (shanghai) Co., Ltd.

Ati Technologies Ulc

Shanghai Iluvatar Corex Semiconductor Co., Ltd.

YunXiao Zou

Jiasheng Chen

Michael Mantor

Bin He

Qingcheng Wang

Liang Han

Angel E Socarras

Wei Yuan

Brian D Emberling

Timour T Paltashev

Mark M Leather

Allen H Rush

Radhakrishna Giduthuri

Yang Jiao

Pingping Shao

Syed Zohaib M Gilani

Chengyuan Wu

Jinshan Zheng

Yile Sun

Rong Zhong

Guoyu Zhu

Yongliu Wang

Eric J Finger

Haihua Zhai

The presented systems enable efficient and effective network communications. In one embodiment, a system comprises a parallel processing unit (PPU) included in a chip and a plurality of interconnects in an inter-chip network (ICN) configured to communicatively couple a plurality of PPUs that communicate with one another via the ICN. Corresponding communications are configured in accordance with routing tables. The routing tables can be stored and reside in registers of an ICN subsystem included in the PPU and include indications of minimum links available to forward a communication from the PPU and another PPU. Respective ones of the routing tables include indications of a correlation between the minimum links and respective ones of a plurality of egress ports that are available for communication coupling to the respective ones of PPUs that are possible destination PPUs. The routing tables can include indications of a single path between two PPUs per respective information communication flow.

Parallel dataflow routing scheme systems and methods

The time required to read from and write to multi-terabyte memory chips in an inter-chip network can be reduced by breaking each memory chip into a number of memory spaces, and then individually addressing each memory space with an address that identifies the memory space, a line number within the memory space, and a number of transmit/receive ports to be used to access the line number.

Routing scheme for heterogeneous interconnected-chip networks using distributed shared memory

Systems, apparatuses, and methods for implementing a low power parallel matrix multiply pipeline are disclosed. In one embodiment, a system includes at least first and second vector register files coupled to a matrix multiply pipeline. The matrix multiply pipeline comprises a plurality of dot product units. The dot product units are configured to calculate dot or outer products for first and second sets of operands retrieved from the first vector register file. The results of the dot or outer product operations are written back to the second vector register file. The second vector register file provides the results from the previous dot or outer product operations as inputs to subsequent dot or outer product operations. The dot product units receive the results from previous phases of the matrix multiply operation and accumulate these previous dot or outer product results with the current dot or outer product results.

Stream processor with low power parallel matrix multiply pipeline

A system includes a high-bandwidth inter-chip network (ICN) that allows communication between parallel processing units (PPUs) in the system. For example, the ICN allows a PPU to communicate with other PPUs on the same compute node or server and also with PPUs on other compute nodes or servers. In embodiments, communication may be at the command level (e.g., at the direct memory access level) and at the instruction level (e.g., the finer-grained load/store instruction level). The ICN allows PPUs in the system to communicate without using a PCIe bus, thereby avoiding its bandwidth limitations and relative lack of speed. The respective routing tables comprise information of multiple paths to any given other PPU.

Systems and methods for multi-branch routing for interconnected chip networks

Systems, apparatuses, and methods for efficiently processing arithmetic operations are disclosed. A computing system includes a processor capable of executing single precision mathematical instructions on data sizes of M bits and half precision mathematical instructions on data sizes of N bits, which is less than M bits. At least two source operands with M bits indicated by a received instruction are read from a register file. If the instruction is a packed math instruction, at least a first source operand with a size of N bits less than M bits is selected from either a high portion or a low portion of one of the at least two source operands read from the register file. The instruction includes fields storing bits, each bit indicating the high portion or the low portion of a given source operand associated with a register identifier specified elsewhere in the instruction.

Packed 16 bits instruction pipeline

A processing element is implemented in a stage of a pipeline and configured to execute an instruction. A first array of multiplexers is to provide information associated with the instruction to the processing element in response to the instruction being in a first set of instructions. A second array of multiplexers is to provide information associated with the instruction to the first processing element in response to the instruction being in a second set of instructions. A control unit is to gate at least one of power or a clock signal provided to the first array of multiplexers in response to the instruction being in the second set.

Pipeline including separate hardware data paths for different instruction types

A data reuse method based on a convolutional neural network accelerator includes a tile scanning module receiving command information of a command module, the command information comprising a size of a CNN job to be divided into tile blocks; a tile scanning module according to a tile. The size of the tile generates the coordinates of the tile block and sends it to the memory request module; the memory request module generates a memory read request and sends the memory read request to the memory module; the memory module sequentially returns the tile block data to the input activation In the weight buffer unit, the input activation weight buffer unit saves the received tile block data to implement data reuse and transmits the received tile block data to the calculation processing unit PE.

Data reuse method based on convolutional neural network accelerator

A method and apparatus for performing a multi-precision computation in a plurality of arithmetic logic units (ALUs) includes pairing a first Single Instruction/Multiple Data (SIMD) block channel device with a second SIMD block channel device to create a first block pair having one-level staggering between the first and second channel devices. A third SIMD block channel device is paired with a fourth SIMD block channel device to create a second block pair having one-level staggering between the third and fourth channel devices. A plurality of source inputs are received at the first block pair and the second block pair. The first block pair computes a first result, and the second block pair computes a second result.

Pairing SIMD lanes to perform double precision operations

Systems, apparatuses, and methods implementing a hybrid matrix multiplication pipeline are disclosed. A hybrid matrix multiplication pipeline is able to execute a plurality of different types of instructions in a plurality of different formats by reusing execution circuitry in an efficient manner. For a first type of instruction for source operand elements of a first size, the pipeline uses N multipliers to perform N multiplication operations on N different sets of operands, where N is a positive integer greater than one. For a second type of instruction for source operand elements of a second size, the N multipliers work in combination to perform a single multiplication operation on a single set of operands, where the second size is greater than the first size. The pipeline also shifts element product results in an efficient manner when implementing a dot product operation.

Hybrid matrix multiplication pipeline

Systems, apparatuses, and methods for implementing a high bandwidth, low power vector register file for use by a parallel processor are disclosed. In one embodiment, a system includes at least a parallel processing unit with a plurality of processing pipeline. The parallel processing unit includes a vector arithmetic logic unit and a high bandwidth, low power, vector register file. The vector register file includes multi-bank high density random-access memories (RAMs) to satisfy register bandwidth requirements. The parallel processing unit also includes an instruction request queue and an instruction operand buffer to provide enough local bandwidth for VALU instructions and vector I/O instructions. Also, the parallel processing unit is configured to leverage the RAM's output flops as a last level cache to reduce duplicate operand requests between multiple instructions. The parallel processing unit includes a vector destination cache to provide additional R/W bandwidth for the vector register file.

Processor support for bypassing vector source operands

Described is a method and processing apparatus to improve power efficiency by gating redundant threads processing. In particular, the method for gating redundant threads in a graphics processor includes determining if data for a thread and data for at least another thread are within a predetermined similarity threshold, gating execution of the at least another thread if the data for the thread and the data for the at least another thread are within the predetermined similarity threshold, and using an output data from the thread as an output data for the at least another thread.

Growing community of inventors

Shanghai, China

YunXiao Zou