WO2008131308A1 - Field-programmable gate array based accelerator system - Google Patents

Field-programmable gate array based accelerator system Download PDF

Info

Publication number
WO2008131308A1
WO2008131308A1 PCT/US2008/060942 US2008060942W WO2008131308A1 WO 2008131308 A1 WO2008131308 A1 WO 2008131308A1 US 2008060942 W US2008060942 W US 2008060942W WO 2008131308 A1 WO2008131308 A1 WO 2008131308A1
Authority
WO
WIPO (PCT)
Prior art keywords
fpga
data
algorithm
memory
relevance
Prior art date
Application number
PCT/US2008/060942
Other languages
French (fr)
Inventor
Ning-Yi Xu
Feng-Hsiung Hsu
Xiong-Fei Cai
Original Assignee
Microsoft Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Microsoft Corporation filed Critical Microsoft Corporation
Publication of WO2008131308A1 publication Critical patent/WO2008131308A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33Querying
    • G06F16/3331Query processing
    • G06F16/334Query execution

Definitions

  • Web search based ad services and search engines have become important tools for providing information to users.
  • One factor in attracting users and advertisers is providing relevant information and ads for a given search query.
  • Search relevance may be determined by a ranking function that ranks resultant documents according to their similarities to the input query.
  • Information retrieval (IR) researchers have studied search relevance for various search engines and tools. Representative methods include Boolean, vector space, probabilistic, and language models. Earlier search engines and tools were mainly based on such IR algorithms. These search engines and tools incorporate in varying degrees the concept of the ranking function. Many factors may affect the ranking function for search relevance. These factors may include page content, title, anchor, URL, spam, and page freshness. It is extremely difficult to manually tune ranking function parameters to accommodate these factors for large-scale data sets, such as those that are common in many applications including World Wide Web (“Web”) applications and speech and image processing. For these large data sets, machine based learning algorithms have been applied to learn complex ranking functions from large-scale data sets.
  • Web World Wide Web
  • Instruction level parallelism techniques somewhat improve the processing time. More particularly, distributed implementations with process level parallelism are faster than many of the PC central processing units (CPUs), which execute instructions in sequential manner. However, distributed implementations occupy many machines. Additionally, for some algorithms, distributed computing yields poor speed improvement per processor added due to communication cost.
  • a Graphics Processing Unit (GPU)-based accelerator could only accelerate a limited spectrum of machine learning algorithms due to its special hardware structure optimized for graphics applications. Thus, memory access bandwidth, communication cost, flexibility and granularity of parallelism remain bottlenecks for these solutions.
  • An accelerator system and method is provided that, according to one exemplary implementation, utilizes FPGA technology to achieve better parallelism and flexibility.
  • the FPGA-based accelerator uses a PCI controller to communicate with a host CPU.
  • a memory hierarchy composed of embedded Random Access Memory (RAM) in the FPGA, Static Random Access Memory (SRAM) and Synchronous Dynamic Random Access Memory (SDRAM), allows the FPGA assisted accelerator to take advantage of memory locality in algorithms.
  • RAM embedded Random Access Memory
  • SRAM Static Random Access Memory
  • SDRAM Synchronous Dynamic Random Access Memory
  • an FPGA-based accelerator system is combined with a relevance-ranking algorithm, such as the algorithm known as RankBoost, to increase the speed of a training process.
  • RankBoost a relevance-ranking algorithm
  • Using an approximated RankBoost algorithm reduces the computation and storage scale from O(N ) to O(N).
  • This algorithm could be mapped to the accelerator system to increase the speed of the pure software implementation by approximately 170 times.
  • the algorithm and related data structures associated with the FPGA-based accelerator may be organized to enable streaming data access and, thus, increase the training speed.
  • the data may be compressed to enable the system and method to be operable with larger data sets.
  • At least a portion of the approximated RankBoost algorithm may be implemented as a single instruction multiple data streams (SIMD) architecture with multiple processing engines (PEs) in the FPGA.
  • SIMD single instruction multiple data streams
  • PEs processing engines
  • FIG. 1 shows an exemplary architecture of an FPGA based accelerator system for machine learning.
  • FIG. 2 shows an exemplary deployment of the accelerator.
  • FIG. 3 shows an exemplary system architecture for an accelerator operable to perform relevance-ranking.
  • FIG. 4 shows an exemplary working flow of the accelerator system.
  • FIG. 5 shows an exemplary architecture of a processing engine (PE) for an accelerator operable to perform relevance-ranking.
  • FIG. 6 shows data sequences for the processing engine shown in FIG. 5.
  • FIG. 7 shows an exemplary data format and processing sequence for a First In First Out (FIFO) buffer.
  • FIFO First In First Out
  • the accelerator system may include an acceleration device, which may include a substrate, such as a Peripheral Component Interconnect (PCI) card, with a Field-Programmable Gate Array (FPGA) and memories acting as caches, e.g., SRAM, SDRAM, and so forth, connected to a computing device.
  • PCI Peripheral Component Interconnect
  • FPGA Field-Programmable Gate Array
  • One or more algorithms may be implemented on one or more of the FPGAs with direct parallel architecture and/or pipelined architecture to exploit both application parallelism and direct functional logic implementation.
  • the PCI could also be replaced by other computer buses, including but not limited to PCI-X, PCI-Express, HyperTransport, Universal Serial Bus (USB) and Front-Side Bus (FSB).
  • a training data set or other data may be loaded onto one or more memories on the accelerator board, or onto embedded memories in the FPGA, to increase memory access bandwidth and data locality.
  • the training data set may comprise information collected from Web searches to assess relevancy, and other characteristics.
  • the system may include or be associated with one or more PCs or other computing devices, each computing device having one or more accelerator cards.
  • the accelerator system 100 may include an acceleration device 102 comprising a Peripheral Component Interface (PCI) board 104 with a Field-Programmable Gate Array (FPGA) 106 and Double Data Rate (DDR) memory 108, e.g., SRAM 110, SDRAM 112, and so forth, connected to a computing device such as a host computer 114.
  • the PCI board 104 may interface with a PCI bus 116 on or associated with the host computing device 114.
  • the PCI board 104, and/or devices thereon, may communicate with the bus 116 thorough a PCI controller 118.
  • the FPGA 106 may comprise computation logic 120 that communicates to the DDR memory devices 108 and/or the PCI controller 118 through one or more interfaces 122.
  • Training data or other data being accessed by the FPGA 106 may be loaded to DDR memory 108, including SRAM 110 or SDRAM 112, on the PCI board 104, or to embedded memories in the FPGA 106, in order to increase memory access bandwidth and data locality.
  • Software loaded on the computer 114 may be capable of programming or re-programming the FPGA 106 at any time during processing.
  • an acceleration system 200 may be composed of one or more computing devices 210, similar to computer 114, with each computing device 210 having one or more PCI cards 204, similar to PCI board 104.
  • the computing devices 210 may be connected through a network 206.
  • multiple cards 204 on multiple computing devices 210 may process data in parallel and thereby handle larger scale algorithms.
  • FIG. 3 shows a block diagram of a system 300 that is operable to implement relevance-ranking software 302 on an FPGA 304 residing on a substrate, such as a PCI card 305.
  • the relevance ranking software 302 may have, or be associated with, a driver 306 having a register read/write (R/W) module 308 and/or a direct memory access read/write (DMA R/W) module 310 for operation of the software 302 with the CPU 312 and memory store 314 through a PCI 316 and/or Northbridge (NB) 318.
  • the PCI card 305 may have a PCI 9054 Chip 320 or other 32-bit PCI bus mastering interface chip in order to facilitate communication between the FPGA 304 and the PCI 316.
  • the FPGA 304 may include a PCI local interface 322 for interfacing with the PCI 9054 chip 320.
  • the PCI local interface 322 may also connect to the processing engine (PE) units, e.g., PEO, PEl, and PEn.
  • the PE units implement the computation logic.
  • the FPGA 304 may also have a DDR interface 324 for interfacing with DDR memory 326.
  • the FPGA 304 may additionally have a control unit 328 for controlling the processing units PEO, PEl, PW, and PEn by sending a signal to the PE units.
  • the FPGA 304 may also have a memory management unit (MMU) 330 for aligning or managing data for faster processing.
  • the processing engines of the FPGA 304 may provide an output to the PCI local interface 320 for further implementation or use.
  • FIG. 4 illustrates an exemplary workflow 400 for processing and utilizing the training data in a system such as shown in FIG. 3.
  • the first column 402 represents actions taken by the application software.
  • the second column 404 represents driver-side activity.
  • the third column 406 describes actions performed on, or by, hardware, such as accelerator device 301.
  • steps (a)-(r) are steps for training.
  • application software will process the training data for hardware. The possible processing may include organizing data in the sequence of how the FPGA logic will access and utilize it.
  • application software will call the write routine in the driver (at (c)) to write the data to memories on the accelerator.
  • the write routine may be implemented with a direct memory access (DMA) method to achieve high bandwidth access to the accelerator.
  • DMA direct memory access
  • the PCI controller on the accelerator upon receiving the training data, the PCI controller on the accelerator will write the data to the FPGA. Then the memory management unit (MMU) in the FPGA will write the data to DDR memory (or other cache memories).
  • the MMU may set a register, or issue an interrupt through PCI controller, indicating that the data transfer has been finished and the application software may proceed.
  • the application software may check the status of the data transfer through register read routine in the driver (at (f)), or wait for the interrupt.
  • application software configures the hardware to begin the training process by calling register write routine in the driver (at (i)).
  • the control unit begins to control other hardware blocks to train using the training data.
  • At (k) application software may write (at (I)) some intermediate data to the accelerator hardware.
  • the MMU in the accelerator sends this data to participate the training.
  • this intermediate data may be generated from intermediate results of the training process (at (n)).
  • the software may check the status of the training (at (r))) to determine if the training process needs to be continued for another round. The software continues to monitor the training process to decide when the training rounds should be stopped.
  • the accelerator system supports hierarchical memory organization and access methods using SDRAM, SRAM and RAM/registers within the FPGA.
  • training data that will be iteratively used may be loaded onto SDRAM onboard an accelerator device, such as accelerator device 301.
  • the training data loaded in the SDRAM may be organized according to its access order in logic associated with the FPGA by a software tool so that the FPGA can fetch data in a so-called, and well-known, "burst" mode, thus enabling high bandwidth access to the data set.
  • Randomly used large-scale data structures could be loaded to SRAM onboard the accelerator device, such as accelerator device 301, and associated with an FPGA, such as FPGA 304.
  • the SRAM may be used as a large low latency cache.
  • Temporary data structures such as intermediate variables, parameters, and so forth, and results, e.g., the learned model, could be stored in distributed memory or registers inside the FPGA, which would act as high bandwidth, low latency cache. The data could be utilized without needing to access memory off of the FPGA, which would enhance the access speed of the cache.
  • a streaming data access architecture and method may be incorporated with the accelerator system and/or the data organization structure, such as described above, to enable fast access to data in the host memory hierarchy and to decrease the amount of hardware/software communication for selected algorithms.
  • Software which may be provided by or on a host CPU, may configure a PCI bridge chip on the accelerator board to fetch data from a host memory hierarchy.
  • the host memory hierarchy may provide a variety of memories including hard disks.
  • the data will contain necessary information (generated and inserted by software), with which FPGA logic can perform computing functions over the data stream without interaction with software or drivers.
  • the data will be organized in the sequence of how FPGA logic is accessing and utilizing it, such that input data is consumed at the time FPGA logic receives it.
  • the FPGA may buffer the result for a delayed read from the software, which reduces the time and processing cost of hardware/software communication.
  • a bit-map based data compression/decompression method for the architecture may be implemented to increase memory capacity and bandwidth available in the accelerator system.
  • Training data may be compressed by conventional compression software and stored in the memories associated with the acceleration device.
  • the FPGA may then read and decompress the data before performing computations.
  • Implementing compression and decompression techniques with the FPGA may increase the virtual bandwidth from a DDR to a PE by 2-4 times the virtual bandwidth for uncompressed data.
  • a machine learning, relevance ranking, algorithm may be implemented using the accelerator system.
  • the goal is to find a ranking function to order the given set of objects.
  • Such an object is denoted as an instance x in a domain (or instance space) X.
  • This feedback is denoted as function ⁇ : X x X ⁇ R, where ⁇ (x 0 , xi)>0 means X 1 should be ranked above X 0 , and ⁇ (x 0 , Xj) ⁇ 0 means x 0 should be ranked above X 1 .
  • a learner attempts to find a ranking function H: X ⁇ R, which is as consistent as possible to the given ⁇ , by asserting X 1 is preferred
  • a relevance-ranking algorithm may be used to learn the ranking function H by combining a given collection of ranking functions.
  • the relevance- ranking algorithm may be pair-based or document-based.
  • the psuedocode for one such relevance ranking algorithm is shown below:
  • the relevance-ranking algorithm is utilized in an iterative manner. In each round, a procedure named "WeakLearn” is called to select the best "weak ranker" from a large set of candidate weak rankers.
  • the weak ranker has the form h t : X ⁇ R and /z,(xy)> h t (x 0 ) means that instance X 1 is ranked higher than x 0 in round
  • final strong ranker H is a weighted sum of the selected weak rankers in each round.
  • the WeakLearn algorithm may be implemented to find the weak
  • the WeakLearn algorithm may be defined as follows:
  • training pairs may be generated and weak rankers may be defined.
  • the instance space for a search engine may be
  • the returned documents may be rated a relevance score, from 1 (means 'poor match 1 ) to 5 (means 'excellent match') using a manual or automated process. Unlabeled documents may be given a relevance score of 0. Based on the rating scores (ground truth), the training pairs for the relevance-ranking algorithm may be generated from the returned documents for each query.
  • So-called “weak rankers” may be defined as a transformation of a document feature, which is a one-dimensional real value number.
  • Document features can be classified into query dependent features, such as query term frequencies in a document and term proximity, and query independent features, such as PageRank, and so forth.
  • query dependent features such as query term frequencies in a document and term proximity
  • query independent features such as PageRank, and so forth.
  • PageRank query independent features
  • a document may be designated as d(q), a pair as ⁇ d ⁇ (q), fikf ⁇ ) ⁇ , and d) means a document for query q t .
  • the k th feature for document is denoted as f k (d' j ).
  • Nf features ⁇ f k (d l j )
  • WeakLearn may be defined as a routine that uses the N/ document features to form its weak rankers, attempting to find the one with the smallest pair-wise disagreement relative to distribution D over N pair document pairs.
  • the weak ranker may be defined by the following relationship: ⁇ ⁇ or f t (d) is undefined
  • the weak learner checks all of the possible combinations of feature ⁇ and threshold ⁇ .
  • the WeakLearn algorithm may be implemented to ascertain a maximum r(f, ⁇ ) by generating a temporary variable ⁇ (d) for each document.
  • contains information regarding labels and pair weights, and the weak ranker only needs to access ⁇ in a document-wise manner for each feature and each threshold, that is O(N doc Nf N ⁇ ), in a straightforward implementation.
  • an alternative weak learner may be utilized using an integral histogram to further reduce the computational complexity to 0(N ⁇ 0C N/). Because of this relatively low computational complexity, the algorithm may be implemented in both software and hardware, e.g., an accelerator system utilizing an FPGA, as described above.
  • r may be calculated in O(N doc Nf) time in each round using an integral histogram in O(N doc Nf) time.
  • feature values ⁇ f k (d) ⁇ in a dimension of the whole feature vector (J 1 ,..., f N ) may be classified into N b ⁇ n bms. The boundaries of these bins are:
  • each document d can be mapped to one of the bins according to the value of f k (d) :
  • Bin k (d) floor( J J M k ) k ) >-I J LTM ⁇ N bm - 1)
  • a relevance-ranking algorithm such as described above, may be implemented using an FPGA based accelerator system, also described above.
  • the main computing data structure is a histogram, mapped to the architecture of single instruction multiple data streams (SIMD) with distributed memories.
  • SIMD architecture is capable of separately building several integral histograms with multiple PE units at the same time, as described above.
  • Software provided on or to the host computer will send the quantized feature values to a DDR memory through the PCI bus, PCI controller and FPGA.
  • the data may be organized to enable streaming memory access, which can make full use of DDR memory bandwidth.
  • the software will call WeakLearn to compute ⁇ (d) for every document, and send ⁇ (d) to a First In First Out (FIFO) queue in the FPGA.
  • the control unit (CU) in the FPGA will direct the PE arrays to build histograms and integral histograms, and will then send the results r(f, ⁇ ) as output to the FIFO queue.
  • the CU is implemented as a finite state machine (FSM), which halts or resumes the pipeline in PE units according to the status of each FIFO.
  • FSM finite state machine
  • the software will read back these r values and select the maximum value. Then the software will update the distribution D(do, d ⁇ ) over all pairs and begin the next round.
  • Fig. 5 illustrates an exemplary micro-architecture of a processing engine 500, such as PEO, PEl, or PEn, previously described.
  • This micro-architecture may be used in building the integral histogram for a relevance ranking algorithm, such as RankBoost.
  • the dual port RAM 502 is used to store the histograms and integral histograms in the building process.
  • the base address of the RAM indexes the feature and the offset address indexes the bin of histogram or integral histogram as defined in the Hist k (i)and Integr ⁇ l ⁇ ) equations described above.
  • the shift registers 504 are used to delay the input feature fld). First, the histogram is built. Feature values f k (d) are input as the offset address to read out the corresponding intermediate values Hist ⁇ ) of the bin i. Then the other input jid) will be added to Hist k ( ⁇ ), and the result will be saved to the same bin where Hist k ( ⁇ ) is read out.
  • FIG. 6 An example data input 600 into 8 PE arrays with 16 features per PE is illustrated in FIG. 6.
  • software aligns a given amount data in its original format.
  • a software tool may be used to rearrange this data in the memory to generate binary data for storage in a memory block.
  • the data is ordered according to the order that the FPGA will access the data.
  • the input data may be organized to be aligned with the PE, as shown in FIG. 6, thereby mapping the data structure to the memories.
  • This data organization enables a streaming memory access.
  • the integral histogram can be implemented with this data structure based on the histogram stored in the dual port RAM.
  • the values are read out, added and then stored back to the memory according to the Integral ⁇ ) equation described above.. At last the final result r(f, ⁇ ) will be read out.
  • a streaming memory access organization can also be used for the FIFO buffer that will provide data from the DDR memory to the group of PE units.
  • the width of the FIFO associated with the PE array may be, for example, 128 bits, which is equivalent to 16 bytes.
  • the data in the FIFO can be arranged as shown in Fig. 7 to map the data to the PE units and further enable streaming memory access to data transferred from the DDR to the FPGA.

Abstract

Accelerator systems and methods are disclosed that utilize FPGA technology to achieve better parallelism and flexibility. The accelerator system may be used to implement a relevance-ranking algorithm, such as RankBoost, for a training process. The algorithm and related data structures may be organized to enable streaming data access and, thus, increase the training speed. The data may be compressed to enable the system and method to be operable with larger data sets. At least a portion of the approximated RankBoost algorithm may be implemented as a single instruction multiple data streams (SIMD) architecture with multiple processing engines (PEs) in the FPGA. Thus, large data sets can be loaded on memories associated with an FPGA to increase the speed of the relevance ranking algorithm.

Description

FIELD-PROGRAMMABLE GATE ARRAY BASED ACCELERATOR
SYSTEM
BACKGROUND
[0001] Web search based ad services and search engines have become important tools for providing information to users. One factor in attracting users and advertisers is providing relevant information and ads for a given search query. Search relevance may be determined by a ranking function that ranks resultant documents according to their similarities to the input query.
[0002] Information retrieval (IR) researchers have studied search relevance for various search engines and tools. Representative methods include Boolean, vector space, probabilistic, and language models. Earlier search engines and tools were mainly based on such IR algorithms. These search engines and tools incorporate in varying degrees the concept of the ranking function. Many factors may affect the ranking function for search relevance. These factors may include page content, title, anchor, URL, spam, and page freshness. It is extremely difficult to manually tune ranking function parameters to accommodate these factors for large-scale data sets, such as those that are common in many applications including World Wide Web ("Web") applications and speech and image processing. For these large data sets, machine based learning algorithms have been applied to learn complex ranking functions from large-scale data sets. [0003] Early algorithms for ranking function learning include Polynomial-based regression, Genetic Programming, RankSVM and classification- based SVM. However, these algorithms were only evaluated on a small-scale dataset due to the high computational cost. In fact, these traditional machine- learning algorithms operate slowly when searching large-scale data sets. Users often wait many hours, days, or even weeks to get results from these data sets. This slow computation time may be due, in part, to a typical personal computer (PC) being unable to exploit full parallelism in machine-learning algorithms efficiently.
[0004] Instruction level parallelism techniques somewhat improve the processing time. More particularly, distributed implementations with process level parallelism are faster than many of the PC central processing units (CPUs), which execute instructions in sequential manner. However, distributed implementations occupy many machines. Additionally, for some algorithms, distributed computing yields poor speed improvement per processor added due to communication cost. A Graphics Processing Unit (GPU)-based accelerator could only accelerate a limited spectrum of machine learning algorithms due to its special hardware structure optimized for graphics applications. Thus, memory access bandwidth, communication cost, flexibility and granularity of parallelism remain bottlenecks for these solutions. SUMMARY
[0005] An accelerator system and method is provided that, according to one exemplary implementation, utilizes FPGA technology to achieve better parallelism and flexibility. The FPGA-based accelerator uses a PCI controller to communicate with a host CPU. A memory hierarchy composed of embedded Random Access Memory (RAM) in the FPGA, Static Random Access Memory (SRAM) and Synchronous Dynamic Random Access Memory (SDRAM), allows the FPGA assisted accelerator to take advantage of memory locality in algorithms.
[0006] According to another exemplary implementation, an FPGA-based accelerator system is combined with a relevance-ranking algorithm, such as the algorithm known as RankBoost, to increase the speed of a training process. Using an approximated RankBoost algorithm reduces the computation and storage scale from O(N ) to O(N). This algorithm could be mapped to the accelerator system to increase the speed of the pure software implementation by approximately 170 times. Several techniques assist in achieving the acceleration rate. The algorithm and related data structures associated with the FPGA-based accelerator may be organized to enable streaming data access and, thus, increase the training speed. The data may be compressed to enable the system and method to be operable with larger data sets. At least a portion of the approximated RankBoost algorithm may be implemented as a single instruction multiple data streams (SIMD) architecture with multiple processing engines (PEs) in the FPGA. Thus, large data sets, such as a training set can be loaded on memories associated with an FPGA to increase the speed of the relevance ranking algorithm.
[0007] By virtue of this system, a user can train a ranking model with much less time and cost, so they can attempt different learning parameters of the algorithm in the same time, or carry out a study that depends on numerous ranking models.
[0008] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.
BRIEF DESCRIPTION OF THE DRAWINGS
[0009] FIG. 1 shows an exemplary architecture of an FPGA based accelerator system for machine learning.
[0010] FIG. 2 shows an exemplary deployment of the accelerator.
[0011] FIG. 3 shows an exemplary system architecture for an accelerator operable to perform relevance-ranking.
[0012] FIG. 4 shows an exemplary working flow of the accelerator system.
[0013] FIG. 5 shows an exemplary architecture of a processing engine (PE) for an accelerator operable to perform relevance-ranking. [0014] FIG. 6 shows data sequences for the processing engine shown in FIG. 5.
[0015] FIG. 7 shows an exemplary data format and processing sequence for a First In First Out (FIFO) buffer.
DETAILED DESCRIPTION
[0016] Overview
An FPGA-based accelerator system for machine learning as described and claimed herein accelerates selected algorithms by providing better processing parallelism and memory access. The accelerator system may include an acceleration device, which may include a substrate, such as a Peripheral Component Interconnect (PCI) card, with a Field-Programmable Gate Array (FPGA) and memories acting as caches, e.g., SRAM, SDRAM, and so forth, connected to a computing device. One or more algorithms may be implemented on one or more of the FPGAs with direct parallel architecture and/or pipelined architecture to exploit both application parallelism and direct functional logic implementation. The PCI could also be replaced by other computer buses, including but not limited to PCI-X, PCI-Express, HyperTransport, Universal Serial Bus (USB) and Front-Side Bus (FSB).
[0017] A training data set or other data may be loaded onto one or more memories on the accelerator board, or onto embedded memories in the FPGA, to increase memory access bandwidth and data locality. The training data set may comprise information collected from Web searches to assess relevancy, and other characteristics. The system may include or be associated with one or more PCs or other computing devices, each computing device having one or more accelerator cards.
[0018] Exemplary System
[0019] Accelerator System Architecture
An exemplary system for use as an accelerator is shown in FIG. 1. The accelerator system 100 may include an acceleration device 102 comprising a Peripheral Component Interface (PCI) board 104 with a Field-Programmable Gate Array (FPGA) 106 and Double Data Rate (DDR) memory 108, e.g., SRAM 110, SDRAM 112, and so forth, connected to a computing device such as a host computer 114. The PCI board 104 may interface with a PCI bus 116 on or associated with the host computing device 114. The PCI board 104, and/or devices thereon, may communicate with the bus 116 thorough a PCI controller 118. The FPGA 106 may comprise computation logic 120 that communicates to the DDR memory devices 108 and/or the PCI controller 118 through one or more interfaces 122.
[0020] Training data or other data being accessed by the FPGA 106 may be loaded to DDR memory 108, including SRAM 110 or SDRAM 112, on the PCI board 104, or to embedded memories in the FPGA 106, in order to increase memory access bandwidth and data locality. Software loaded on the computer 114 may be capable of programming or re-programming the FPGA 106 at any time during processing.
[0021] As shown in FIG. 2, an acceleration system 200 may be composed of one or more computing devices 210, similar to computer 114, with each computing device 210 having one or more PCI cards 204, similar to PCI board 104. The computing devices 210 may be connected through a network 206. Thus, multiple cards 204 on multiple computing devices 210 may process data in parallel and thereby handle larger scale algorithms.
[0022] FIG. 3 shows a block diagram of a system 300 that is operable to implement relevance-ranking software 302 on an FPGA 304 residing on a substrate, such as a PCI card 305. The relevance ranking software 302 may have, or be associated with, a driver 306 having a register read/write (R/W) module 308 and/or a direct memory access read/write (DMA R/W) module 310 for operation of the software 302 with the CPU 312 and memory store 314 through a PCI 316 and/or Northbridge (NB) 318. The PCI card 305 may have a PCI 9054 Chip 320 or other 32-bit PCI bus mastering interface chip in order to facilitate communication between the FPGA 304 and the PCI 316.
[0023] The FPGA 304 may include a PCI local interface 322 for interfacing with the PCI 9054 chip 320. The PCI local interface 322 may also connect to the processing engine (PE) units, e.g., PEO, PEl, and PEn. The PE units implement the computation logic. The FPGA 304 may also have a DDR interface 324 for interfacing with DDR memory 326. The FPGA 304 may additionally have a control unit 328 for controlling the processing units PEO, PEl, PW, and PEn by sending a signal to the PE units. The FPGA 304 may also have a memory management unit (MMU) 330 for aligning or managing data for faster processing. The processing engines of the FPGA 304 may provide an output to the PCI local interface 320 for further implementation or use.
[0024] FIG. 4 illustrates an exemplary workflow 400 for processing and utilizing the training data in a system such as shown in FIG. 3. The first column 402 represents actions taken by the application software. The second column 404 represents driver-side activity. The third column 406 describes actions performed on, or by, hardware, such as accelerator device 301. In the following, steps (a)-(r) are steps for training. At (a), application software will process the training data for hardware. The possible processing may include organizing data in the sequence of how the FPGA logic will access and utilize it. At (b), application software will call the write routine in the driver (at (c)) to write the data to memories on the accelerator. The write routine may be implemented with a direct memory access (DMA) method to achieve high bandwidth access to the accelerator. At (d), upon receiving the training data, the PCI controller on the accelerator will write the data to the FPGA. Then the memory management unit (MMU) in the FPGA will write the data to DDR memory (or other cache memories). At (e), when all the data has been saved to the memory, the MMU may set a register, or issue an interrupt through PCI controller, indicating that the data transfer has been finished and the application software may proceed. At (g), the application software may check the status of the data transfer through register read routine in the driver (at (f)), or wait for the interrupt. At (h), application software configures the hardware to begin the training process by calling register write routine in the driver (at (i)). At (j), the control unit begins to control other hardware blocks to train using the training data. In the training process, at (k) application software may write (at (I)) some intermediate data to the accelerator hardware. At (m), the MMU in the accelerator sends this data to participate the training. At (p), this intermediate data may be generated from intermediate results of the training process (at (n)). At (q), the software may check the status of the training (at (r))) to determine if the training process needs to be continued for another round. The software continues to monitor the training process to decide when the training rounds should be stopped.
[0025] Data organization
The accelerator system supports hierarchical memory organization and access methods using SDRAM, SRAM and RAM/registers within the FPGA.
[0026] According to one example, training data that will be iteratively used may be loaded onto SDRAM onboard an accelerator device, such as accelerator device 301. The training data loaded in the SDRAM may be organized according to its access order in logic associated with the FPGA by a software tool so that the FPGA can fetch data in a so-called, and well-known, "burst" mode, thus enabling high bandwidth access to the data set. [0027] Randomly used large-scale data structures could be loaded to SRAM onboard the accelerator device, such as accelerator device 301, and associated with an FPGA, such as FPGA 304. According to this implementation, the SRAM may be used as a large low latency cache.
[0028] Temporary data structures, such as intermediate variables, parameters, and so forth, and results, e.g., the learned model, could be stored in distributed memory or registers inside the FPGA, which would act as high bandwidth, low latency cache. The data could be utilized without needing to access memory off of the FPGA, which would enhance the access speed of the cache.
[0029] Stream Data Processing Architecture
A streaming data access architecture and method may be incorporated with the accelerator system and/or the data organization structure, such as described above, to enable fast access to data in the host memory hierarchy and to decrease the amount of hardware/software communication for selected algorithms. Software, which may be provided by or on a host CPU, may configure a PCI bridge chip on the accelerator board to fetch data from a host memory hierarchy. The host memory hierarchy may provide a variety of memories including hard disks. The data will contain necessary information (generated and inserted by software), with which FPGA logic can perform computing functions over the data stream without interaction with software or drivers. Furthermore, the data will be organized in the sequence of how FPGA logic is accessing and utilizing it, such that input data is consumed at the time FPGA logic receives it. The FPGA may buffer the result for a delayed read from the software, which reduces the time and processing cost of hardware/software communication.
[0030] Data Compression/Decompression
A bit-map based data compression/decompression method for the architecture may be implemented to increase memory capacity and bandwidth available in the accelerator system. Training data may be compressed by conventional compression software and stored in the memories associated with the acceleration device. The FPGA may then read and decompress the data before performing computations. Implementing compression and decompression techniques with the FPGA may increase the virtual bandwidth from a DDR to a PE by 2-4 times the virtual bandwidth for uncompressed data.
[0031 ] Relevance-Ranking Algorithm
A machine learning, relevance ranking, algorithm may be implemented using the accelerator system. Generally, when ranking objects, the goal is to find a ranking function to order the given set of objects. Such an object is denoted as an instance x in a domain (or instance space) X. As a form of feedback, information about which instance should be ranked above (or below) one another is provided for every pair of instances. This feedback is denoted as function Φ: X x X →R, where Φ (x0, xi)>0 means X1 should be ranked above X0, andΦ (x0, Xj)<0 means x0 should be ranked above X1. A learner then attempts to find a ranking function H: X →R, which is as consistent as possible to the given Φ, by asserting X1 is preferred
Figure imgf000013_0001
[0032] A relevance-ranking algorithm may be used to learn the ranking function H by combining a given collection of ranking functions. The relevance- ranking algorithm may be pair-based or document-based. The psuedocode for one such relevance ranking algorithm, is shown below:
Initialize: Distribution D over Xx X
Do for t = \,...,T :
(1) Train WeakLearn using distribution Dt.
(2) WeakLearn returns a weak hypothesis ht,
(3) Choose α, e R
(4) Update weights: for each pair (do, dj): j Λ _ Dχd0,dλ) exρ(-αr, (ht (d0 ) - ht Jd1 )))
where Zt is the normalization factor:
Figure imgf000013_0002
T
Output: the final hypothesis: H(x) = ∑atht f=l
[0033] The relevance-ranking algorithm is utilized in an iterative manner. In each round, a procedure named "WeakLearn" is called to select the best "weak ranker" from a large set of candidate weak rankers. The weak ranker has the form ht: X →R and /z,(xy)> ht(x0) means that instance X1 is ranked higher than x0 in round
t. A distribution Dt over X x X is maintained in the training process. Weight Dt(x0,
Xy) will be decreased if ht ranks X0 and X1 correctly (ht (X1) > ht (xoj), and increased
otherwise. Thus, Dt will tend to concentrate on the pairs that are hard to rank. The
final strong ranker H is a weighted sum of the selected weak rankers in each round.
[0034] The WeakLearn algorithm may be implemented to find the weak
ranker with a maximum r(f, θ), by generating a temporary variable π(d) for each
document. The WeakLearn algorithm may be defined as follows:
Given: Distribution D(d0, d\) over all pairs
Initialize: (1) For each document d(q):
Compute π(d(q)) = ∑d,(?) (D(d' (q), d(q)) - D(d(q), d\q)))
(2) For every feature^ and every threshold θk s:
Compute r(fks k) = ∑π(d(q)) d(q)ft (d(q))>θ*
(3) Find the maximum]
Figure imgf000014_0001
1 1 + r* (4) Compute: a = — ln( -)
2 \ - r
Output: weak ranking (fk, , θ\ ) and a.
[0035] To extend the relevance-ranking algorithm to Web relevance
ranking, training pairs may be generated and weak rankers may be defined. To
generate the training pairs, the instance space for a search engine may be
partitioned according to queries issued by users. For each query q, the returned documents may be rated a relevance score, from 1 (means 'poor match1) to 5 (means 'excellent match') using a manual or automated process. Unlabeled documents may be given a relevance score of 0. Based on the rating scores (ground truth), the training pairs for the relevance-ranking algorithm may be generated from the returned documents for each query.
[0036] So-called "weak rankers" may be defined as a transformation of a document feature, which is a one-dimensional real value number. Document features can be classified into query dependent features, such as query term frequencies in a document and term proximity, and query independent features, such as PageRank, and so forth. Thus, the same document may be represented by different feature vectors for different queries based upon its query-dependent features.
[0037] In keeping with the previous algorithm example, a document may be designated as d(q), a pair as {d\(q), fikfø)}, and d) means a document for query qt. The kth feature for document is denoted as fk(d'j). With these notations, an alternative relevance-ranking algorithm may be implemented as follows.
Initialize: initial distribution D over XxX
Given: Nq queries {qt \ i=l ..., Nq).
N1 documents { d) | j=l,..., N1 } for each query qλ, where
Figure imgf000015_0001
-
Nf features {fk(dl j) | j= 1 , ... , N/ } for each document d1)- N*θ candidate thresholds {θk s\ s=l,..., NV} for each^. N pair pairs (d'jj, d^) generated by ground truth rating { Riq^d)) } or
[ R) )-
Initialize: initial distribution D(d)], dj?) over XxX
Do for t = \,...,T :
(1) Train WeakLearn using distribution Dt.
(2) WeakLearn returns a weak hypothesis ht, weighty
(3) Update weights: for each pair (d0, dj):
, Λ _ D1 (d0, dx)Qπp(-at(ht (d0) - ht Jd1))) ut+\ \ao > a\ ) - 7
where Zt is the normalization factor:
Figure imgf000016_0001
T
Output: the final hypothesis: H(x) = ∑atht t=l
[0038] For the relevance-ranking algorithms described by example above, WeakLearn may be defined as a routine that uses the N/ document features to form its weak rankers, attempting to find the one with the smallest pair-wise disagreement relative to distribution D over Npair document pairs. The weak ranker may be defined by the following relationship:
Figure imgf000016_0002
< θ or ft (d) is undefined
[0039] To find the best hid), the weak learner checks all of the possible combinations of feature β and threshold θ. The WeakLearn algorithm may be implemented to ascertain a maximum r(f, θ) by generating a temporary variable π(d) for each document. Intuitively, π contains information regarding labels and pair weights, and the weak ranker only needs to access π in a document-wise manner for each feature and each threshold, that is O(NdocNf Nθ), in a straightforward implementation. Based on this, an alternative weak learner may be utilized using an integral histogram to further reduce the computational complexity to 0(N^0CN/). Because of this relatively low computational complexity, the algorithm may be implemented in both software and hardware, e.g., an accelerator system utilizing an FPGA, as described above.
[0040] According to the implementation, r may be calculated in O(NdocNf) time in each round using an integral histogram in O(NdocNf) time. First, feature values { fk(d) } in a dimension of the whole feature vector (J1,..., fN ) may be classified into Nbιnbms. The boundaries of these bins are:
θk = J max J mm S + fin,s = 0X...,Nbm ,
Nbm
where /^x and /^1n are maximum and minimum values of all fk in the training data set. Then each document d can be mapped to one of the bins according to the value of fk(d) :
Bink(d) = floor( JJMk)k ) >-I JL™ Nbm - 1)
J max J mm
The histogram of π(d) over feature^ is then built using: Histk(i) = ∑π(d),i = 0,...,(Nbm - l) d Bιnt (d)=ι Then, an integral histogram can be determined by adding elements in the histogram from the right (i=Nb,n-l) to the left (i=0). That is, Integral^) = ∑Histk(a),i = 0,..., (JV4n, - 1)
[0041 ] Exemplary Implementation of Relevance-Ranking Algorithm
A relevance-ranking algorithm, such as described above, may be implemented using an FPGA based accelerator system, also described above. The main computing data structure is a histogram, mapped to the architecture of single instruction multiple data streams (SIMD) with distributed memories. The SIMD architecture is capable of separately building several integral histograms with multiple PE units at the same time, as described above.
[0042] Software provided on or to the host computer will send the quantized feature values to a DDR memory through the PCI bus, PCI controller and FPGA. As described above, the data may be organized to enable streaming memory access, which can make full use of DDR memory bandwidth. In each training round, the software will call WeakLearn to compute π(d) for every document, and send π(d) to a First In First Out (FIFO) queue in the FPGA. The control unit (CU) in the FPGA will direct the PE arrays to build histograms and integral histograms, and will then send the results r(f,θ) as output to the FIFO queue. The CU is implemented as a finite state machine (FSM), which halts or resumes the pipeline in PE units according to the status of each FIFO. When the CU indicates that the calculation of r is finished, the software will read back these r values and select the maximum value. Then the software will update the distribution D(do, d\) over all pairs and begin the next round.
[0043] It is noted that the micro-architecture of the PE supports fully- pipelined operation, which enhances the performance of hardware, particularly with regard to machine learning algorithms, such as a relevance-ranking algorithm. Fig. 5 illustrates an exemplary micro-architecture of a processing engine 500, such as PEO, PEl, or PEn, previously described. This micro-architecture may be used in building the integral histogram for a relevance ranking algorithm, such as RankBoost. The dual port RAM 502 is used to store the histograms and integral histograms in the building process. The base address of the RAM indexes the feature and the offset address indexes the bin of histogram or integral histogram as defined in the Histk(i)and Integrαlφ) equations described above. The shift registers 504 are used to delay the input feature fld). First, the histogram is built. Feature values fk(d) are input as the offset address to read out the corresponding intermediate values Hist^ϊ) of the bin i. Then the other input jid) will be added to Histk(ι), and the result will be saved to the same bin where Histk(ϊ) is read out.
[0044] An example data input 600 into 8 PE arrays with 16 features per PE is illustrated in FIG. 6. First, software aligns a given amount data in its original format. A software tool may be used to rearrange this data in the memory to generate binary data for storage in a memory block. The data is ordered according to the order that the FPGA will access the data. Moreover the input data may be organized to be aligned with the PE, as shown in FIG. 6, thereby mapping the data structure to the memories. This data organization enables a streaming memory access. The integral histogram can be implemented with this data structure based on the histogram stored in the dual port RAM. The values are read out, added and then stored back to the memory according to the Integralφ) equation described above.. At last the final result r(f, θ) will be read out.
[0045] A streaming memory access organization can also be used for the FIFO buffer that will provide data from the DDR memory to the group of PE units. The width of the FIFO associated with the PE array may be, for example, 128 bits, which is equivalent to 16 bytes. The data in the FIFO can be arranged as shown in Fig. 7 to map the data to the PE units and further enable streaming memory access to data transferred from the DDR to the FPGA. Thus, running the relevance- ranking algorithm utilizing an FPGA and incorporating streaming memory access provides fast processing of large data sets.
[0046] Although the invention has been described in language specific to structural features and/or methodological steps, it is to be understood that the invention defined in the appended claims is not necessarily limited to the specific features or steps described. Rather, the specific features and steps are disclosed as preferred forms of implementing the claimed invention.

Claims

1. A system comprising: a Field Programmable Gate Array (FPGA) provided on a substrate; a memory connected to the substrate and the FPGA; an interface for connecting the FPGA to a computing device; and a relevance-ranking algorithm associated with the logic in FPGA.
2. A system as recited in claim 1, wherein the substrate comprises a Peripheral Component Interface (PCI) board, PCI-X board, PCI-Express board, HyperTransport board, Universal Serial Bus (USB) board or Front-Side Bus (FSB) board.
3. A system as recited in claim 1, wherein the FPGA has at least one processing engine, and the processing engine is controlled by a control unit.
4. A system as recited in claim 3, wherein the memory comprises Double Data Rate (DDR) memory.
5. A system as recited in claim 1, wherein the relevance ranking algorithm incorporates a RankBoost algorithm.
6. A system as recited in claim 4, wherein the FPGA comprises a number of processing engine (PE) unit, and wherein data is arranged in a First In First Out (FIFO) buffer to map the data onto the PE unit.
7. A system as recited in claim 1, wherein the FPGA associated with a computing device is a first FPGA associated with a first computing device, the system further comprising a network connecting the first FPGA associated with the first computing device to a second FPGA associated with the second computing device.
8. A system as recited in claim 7, wherein the first and second computing devices are each associated with multiple FPGA devices
9. A method comprising: mapping data in one or more data structures to one or more memories associated with a Field Programmable Gate Array (FPGA); and enabling the performance of a relevance ranking-algorithm with respect to the data structures.
10. A method as recited in claim 9, wherein the relevance-ranking algorithm comprises at least a portion of a RankBoost algorithm.
11. A method as recited in claim 9, further comprising loading a training dataset into a memory associated with the FPGA.
12. A method as recited in claim 9, further comprising utilizing memory within the FPGA as a cache memory.
13. The method of claim 9, further comprising compressing data in the one or more data structures prior to mapping the one or more data structures in the one or more memories.
14. The method of claim 9, wherein the relevance-ranking algorithm is document based.
15. The method of claim 9, further comprising organizing the data according to input data structure of processing units in the FPGA and according to the order by which the FPGA will access the data.
16. A system comprising: an FPGA logic device operable to perform a machine learning algorithm; a PCI controller to communicate with a Central Processing Unit (CPU) of a host computing device, and a memory hierarchy composed of Static Random Access Memory (SRAM) and Synchronous Dynamic Random Access Memory (SDRAM) associated with the FPGA and embedded Random Access Memory (RAM) within the FPGA.
17. A system according to claim 16, wherein the machine learning algorithm comprises a document-based relevance-ranking algorithm.
18. A system according to claim 16, wherein: training data that will be iteratively used is loaded onto the SDRAM and organized according to its usage pattern in logic associated with the FPGA; randomly used large-scale data structures are loaded onto the SRAM to be used as a large low latency cache; and temporary data structures are stored in the embedded RAM to act as high bandwidth, low latency cache.
19. A system according to claim 16, wherein multiple FPGA devices are provided for connection to the computer.
20. A system according to claim 19, wherein multiple computers are connected together through a network.
PCT/US2008/060942 2007-04-19 2008-04-18 Field-programmable gate array based accelerator system WO2008131308A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US11/737,605 2007-04-19
US11/737,605 US8117137B2 (en) 2007-04-19 2007-04-19 Field-programmable gate array based accelerator system

Publications (1)

Publication Number Publication Date
WO2008131308A1 true WO2008131308A1 (en) 2008-10-30

Family

ID=39873236

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2008/060942 WO2008131308A1 (en) 2007-04-19 2008-04-18 Field-programmable gate array based accelerator system

Country Status (2)

Country Link
US (2) US8117137B2 (en)
WO (1) WO2008131308A1 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103729320A (en) * 2013-12-20 2014-04-16 天津光电通信技术有限公司 Method for implementing CY7C68013 communication on basis of FPGA (field programmable gate array)
CN103813125A (en) * 2014-03-03 2014-05-21 安庆师范学院 Multi-branch digital image processing system

Families Citing this family (43)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8117137B2 (en) 2007-04-19 2012-02-14 Microsoft Corporation Field-programmable gate array based accelerator system
US8131659B2 (en) * 2008-09-25 2012-03-06 Microsoft Corporation Field-programmable gate array based accelerator system
US8755515B1 (en) 2008-09-29 2014-06-17 Wai Wu Parallel signal processing system and method
US10270709B2 (en) 2015-06-26 2019-04-23 Microsoft Technology Licensing, Llc Allocating acceleration component functionality for supporting services
CN103324559A (en) * 2013-06-27 2013-09-25 成都林海电子有限责任公司 Speed testing method for data transmitting from PCIExpress master to equipment based on FPGA (field programmable gate array)
US20150261724A1 (en) * 2014-03-14 2015-09-17 Emilio Billi Massive parallel exascale storage system architecture
US10332008B2 (en) 2014-03-17 2019-06-25 Microsoft Technology Licensing, Llc Parallel decision tree processor architecture
US10216448B2 (en) * 2014-09-11 2019-02-26 Hitachi, Ltd. Storage system with read request accelerator having dynamic internal data memory allocation
CN104467909B (en) * 2014-12-23 2016-11-30 天津光电通信技术有限公司 A kind of transmission circuit of configurable pci bus based on FPGA technology
US10303622B2 (en) 2015-03-06 2019-05-28 Hewlett Packard Enterprise Development Lp Data write to subset of memory devices
US9792154B2 (en) * 2015-04-17 2017-10-17 Microsoft Technology Licensing, Llc Data processing system having a hardware acceleration plane and a software plane
US10198294B2 (en) 2015-04-17 2019-02-05 Microsoft Licensing Technology, LLC Handling tenant requests in a system that uses hardware acceleration components
US10296392B2 (en) 2015-04-17 2019-05-21 Microsoft Technology Licensing, Llc Implementing a multi-component service using plural hardware acceleration components
US10606651B2 (en) 2015-04-17 2020-03-31 Microsoft Technology Licensing, Llc Free form expression accelerator with thread length-based thread assignment to clustered soft processor cores that share a functional circuit
US10511478B2 (en) * 2015-04-17 2019-12-17 Microsoft Technology Licensing, Llc Changing between different roles at acceleration components
US10216555B2 (en) 2015-06-26 2019-02-26 Microsoft Technology Licensing, Llc Partially reconfiguring acceleration components
US10452995B2 (en) 2015-06-29 2019-10-22 Microsoft Technology Licensing, Llc Machine learning classification on hardware accelerators with stacked memory
US10540588B2 (en) * 2015-06-29 2020-01-21 Microsoft Technology Licensing, Llc Deep neural network processing on hardware accelerators with stacked memory
US11157800B2 (en) 2015-07-24 2021-10-26 Brainchip, Inc. Neural processor based accelerator system and method
US20180005059A1 (en) * 2016-07-01 2018-01-04 Google Inc. Statistics Operations On Two Dimensional Image Processor
CN107102962B (en) * 2017-04-27 2019-10-18 科大讯飞股份有限公司 Board circuit and computer equipment based on programmable logic device
US10838902B2 (en) * 2017-06-23 2020-11-17 Facebook, Inc. Apparatus, system, and method for performing hardware acceleration via expansion cards
US11429848B2 (en) * 2017-10-17 2022-08-30 Xilinx, Inc. Host-directed multi-layer neural network processing via per-layer work requests
US11222256B2 (en) * 2017-10-17 2022-01-11 Xilinx, Inc. Neural network processing system having multiple processors and a neural network accelerator
TWI672924B (en) * 2017-11-23 2019-09-21 財團法人資訊工業策進會 Platform as a service cloud server and machine learning data processing method thereof
US10970080B2 (en) 2018-02-08 2021-04-06 Marvell Asia Pte, Ltd. Systems and methods for programmable hardware architecture for machine learning
US11016801B1 (en) * 2018-05-22 2021-05-25 Marvell Asia Pte, Ltd. Architecture to support color scheme-based synchronization for machine learning
US10929779B1 (en) * 2018-05-22 2021-02-23 Marvell Asia Pte, Ltd. Architecture to support synchronization between core and inference engine for machine learning
US10891136B1 (en) 2018-05-22 2021-01-12 Marvell Asia Pte, Ltd. Data transmission between memory and on chip memory of inference engine for machine learning via a single data gathering instruction
US10929760B1 (en) 2018-05-22 2021-02-23 Marvell Asia Pte, Ltd. Architecture for table-based mathematical operations for inference acceleration in machine learning
US10997510B1 (en) 2018-05-22 2021-05-04 Marvell Asia Pte, Ltd. Architecture to support tanh and sigmoid operations for inference acceleration in machine learning
US10929778B1 (en) * 2018-05-22 2021-02-23 Marvell Asia Pte, Ltd. Address interleaving for machine learning
KR102559581B1 (en) 2018-05-23 2023-07-25 삼성전자주식회사 Storage device including reconfigurable logic and method of operating the storage device
US11500674B2 (en) 2018-06-26 2022-11-15 Intel Corporation Circuitry with adaptive memory assistance capabilities
US11710029B2 (en) * 2018-09-28 2023-07-25 Intel Corporation Methods and apparatus to improve data training of a machine learning model using a field programmable gate array
US11422954B2 (en) 2018-09-28 2022-08-23 Intel Corporation Techniques for accelerating memory access operations
CN109379262B (en) * 2018-09-30 2021-03-23 天津市英贝特航天科技有限公司 Dual-redundancy CAN bus communication card and method for PCIe interface
CN109446135A (en) * 2018-10-24 2019-03-08 盛科网络(苏州)有限公司 A kind of device and method for realizing the downloading of more large capacity FPGA remote high-speeds
CN109933369B (en) * 2019-02-01 2021-09-21 京微齐力(北京)科技有限公司 System chip of artificial intelligence module integrated with single instruction multiple data flow architecture
WO2020215124A1 (en) * 2019-04-26 2020-10-29 The University Of Sydney An improved hardware primitive for implementations of deep neural networks
CN110708513B (en) * 2019-10-18 2021-06-01 中国科学院长春光学精密机械与物理研究所 8K video multi-core heterogeneous processing device
US20220058468A1 (en) * 2020-08-21 2022-02-24 U.S. Army Research Laboratory Field Programmable Neural Array
CN112637080B (en) * 2020-12-14 2022-11-01 中国科学院声学研究所 Load balancing processing system based on FPGA

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040111388A1 (en) * 2002-12-06 2004-06-10 Frederic Boiscuvier Evaluating relevance of results in a semi-structured data-base system
US20050234953A1 (en) * 2004-04-15 2005-10-20 Microsoft Corporation Verifying relevance between keywords and Web site contents
US20050246328A1 (en) * 2004-04-30 2005-11-03 Microsoft Corporation Method and system for ranking documents of a search result to improve diversity and information richness
US7197497B2 (en) * 2003-04-25 2007-03-27 Overture Services, Inc. Method and apparatus for machine learning a document relevance function

Family Cites Families (97)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6675499B2 (en) 1989-08-30 2004-01-13 Anatomic Research, Inc. Shoe sole structures
EP0566015A3 (en) 1992-04-14 1994-07-06 Eastman Kodak Co Neural network optical character recognition system and method for classifying characters in amoving web
US5320538A (en) 1992-09-23 1994-06-14 Hughes Training, Inc. Interactive aircraft training system and method
US5432895A (en) 1992-10-01 1995-07-11 University Corporation For Atmospheric Research Virtual reality imaging system
US5423554A (en) 1993-09-24 1995-06-13 Metamedia Ventures, Inc. Virtual reality game method and apparatus
US5490784A (en) 1993-10-29 1996-02-13 Carmein; David E. E. Virtual reality system with enhanced sensory apparatus
US5577981A (en) 1994-01-19 1996-11-26 Jarvik; Robert Virtual reality exercise machine and computer controlled video system
US5892961A (en) 1995-02-17 1999-04-06 Xilinx, Inc. Field programmable gate array having programming instructions in the configuration bitstream
US5913727A (en) 1995-06-02 1999-06-22 Ahdoot; Ned Interactive movement and contact simulation game
JPH11515097A (en) 1995-09-19 1999-12-21 モルフォメトリックス テクノロジーズ インク. Neural network assisted multispectral segmentation system
US6141034A (en) 1995-12-15 2000-10-31 Immersive Media Co. Immersive imaging method and apparatus
US6601049B1 (en) 1996-05-02 2003-07-29 David L. Cooper Self-adjusting multi-layer neural network architectures and methods therefor
US6084979A (en) 1996-06-20 2000-07-04 Carnegie Mellon University Method for creating virtual reality
US6064749A (en) 1996-08-02 2000-05-16 Hirota; Gentaro Hybrid tracking for augmented reality using both camera motion detection and landmark tracking
JP3170638B2 (en) 1996-08-08 2001-05-28 谷 白糸 Virtual reality experience device
US5892962A (en) 1996-11-12 1999-04-06 Lucent Technologies Inc. FPGA-based processor
US6215898B1 (en) 1997-04-15 2001-04-10 Interval Research Corporation Data processing system and method
US6122627A (en) 1997-05-09 2000-09-19 International Business Machines Corporation System, method, and program for object building in queries over object views
US6477527B2 (en) 1997-05-09 2002-11-05 International Business Machines Corporation System, method, and program for object building in queries over object views
US6134540A (en) 1997-05-09 2000-10-17 International Business Machines Corporation System, method, and program for applying query rewrite technology to object building
US6011407A (en) 1997-06-13 2000-01-04 Xilinx, Inc. Field programmable gate array with dedicated computer bus interface and method for configuring both
US5883628A (en) 1997-07-03 1999-03-16 International Business Machines Corporation Climability: property for objects in 3-D virtual environments
US6037914A (en) 1997-08-25 2000-03-14 Hewlett-Packard Company Method and apparatus for augmented reality using a see-through head-mounted display
DE59811987D1 (en) * 1997-10-29 2004-10-28 Siemens Ag Method and arrangement for determining distribution information
JP2003526129A (en) 1997-12-17 2003-09-02 エリクセントリミティド Implementation of a multiplier in a programmable array
US6222757B1 (en) 1998-02-25 2001-04-24 Xilinx, Inc. Configuration memory architecture for FPGA
US6226237B1 (en) 1998-03-26 2001-05-01 O2 Micro International Ltd. Low power CD-ROM player for portable computer
US6625299B1 (en) 1998-04-08 2003-09-23 Jeffrey Meisner Augmented reality technology
WO2002095534A2 (en) 2001-05-18 2002-11-28 Biowulf Technologies, Llc Methods for feature selection in a learning machine
US7921068B2 (en) 1998-05-01 2011-04-05 Health Discovery Corporation Data mining platform for knowledge discovery from heterogeneous data types and/or heterogeneous data sources
US6199008B1 (en) 1998-09-17 2001-03-06 Noegenesis, Inc. Aviation, terrain and weather display system
US6356637B1 (en) 1998-09-18 2002-03-12 Sun Microsystems, Inc. Field programmable gate arrays
GB9902115D0 (en) 1999-02-01 1999-03-24 Axeon Limited Neural networks
JP2000350865A (en) 1999-06-11 2000-12-19 Mr System Kenkyusho:Kk Game device for composite real space, image processing method therefor and program storage medium
US6408257B1 (en) 1999-08-31 2002-06-18 Xerox Corporation Augmented-reality display method and system
US20020010734A1 (en) 2000-02-03 2002-01-24 Ebersole John Franklin Internetworked augmented reality system and method
US8095508B2 (en) 2000-04-07 2012-01-10 Washington University Intelligent data storage and processing using FPGA devices
US6611823B1 (en) 2000-04-20 2003-08-26 Board Of Regents, The University Of Texas System Backlash compensation using neural network
US6751600B1 (en) 2000-05-30 2004-06-15 Commerce One Operations, Inc. Method for automatic categorization of items
EP1170868B1 (en) 2000-07-04 2008-08-27 Sun Microsystems, Inc. Field programmable gate arrays (FPGA) and method for processing FPGA configuration data
US7099860B1 (en) * 2000-10-30 2006-08-29 Microsoft Corporation Image retrieval systems and methods with semantic and feature based relevance feedback
JP4009433B2 (en) 2001-03-29 2007-11-14 株式会社コナミデジタルエンタテインメント GAME DEVICE, GAME PROGRAM, AND GAME SYSTEM
JP2004534963A (en) 2001-03-30 2004-11-18 コーニンクレッカ フィリップス エレクトロニクス エヌ ヴィ Methods, systems and devices for augmented reality
EP1262907B1 (en) 2001-05-28 2007-10-03 Honda Research Institute Europe GmbH Pattern recognition with hierarchical networks
US6999083B2 (en) 2001-08-22 2006-02-14 Microsoft Corporation System and method to provide a spectator experience for networked gaming
US7024033B2 (en) 2001-12-08 2006-04-04 Microsoft Corp. Method for boosting the performance of machine-learning classifiers
US7138963B2 (en) 2002-07-18 2006-11-21 Metamersion, Llc Method for automatically tracking objects in augmented reality
JP4065507B2 (en) 2002-07-31 2008-03-26 キヤノン株式会社 Information presentation apparatus and information processing method
GB0218188D0 (en) 2002-08-06 2002-09-11 Hewlett Packard Co Methods and arrangements applicable to exhibition spaces
WO2004017590A1 (en) 2002-08-19 2004-02-26 The Regents Of The University Of California Fpga implementation of a digital qam modulator
US20070035562A1 (en) 2002-09-25 2007-02-15 Azuma Ronald T Method and apparatus for image enhancement
US8458028B2 (en) 2002-10-16 2013-06-04 Barbaro Technologies System and method for integrating business-related content into an electronic game
KR100490726B1 (en) 2002-10-17 2005-05-24 한국전자통신연구원 Apparatus and method for video based shooting game
US7050078B2 (en) 2002-12-19 2006-05-23 Accenture Global Services Gmbh Arbitrary object tracking augmented reality applications
US20040225483A1 (en) 2003-02-24 2004-11-11 Michal Okoniewski Fdtd hardware acceleration system
EP1654692A1 (en) 2003-07-11 2006-05-10 Computer Associates Think, Inc. Method and apparatus for automated feature selection
US7099745B2 (en) 2003-10-24 2006-08-29 Sap Aktiengesellschaft Robot system using virtual world
JP4642757B2 (en) 2004-07-23 2011-03-02 パナソニック株式会社 Image processing apparatus and image processing method
US7369869B2 (en) 2004-07-26 2008-05-06 Motorola, Inc. Method and system of scanning a TDMA channel
US8547401B2 (en) 2004-08-19 2013-10-01 Sony Computer Entertainment Inc. Portable augmented reality device and method
US20060047704A1 (en) 2004-08-31 2006-03-02 Kumar Chitra Gopalakrishnan Method and system for providing information services relevant to visual imagery
US8585476B2 (en) 2004-11-16 2013-11-19 Jeffrey D Mullen Location-based games and augmented reality systems
US7519217B2 (en) 2004-11-23 2009-04-14 Microsoft Corporation Method and system for generating a classifier using inter-sample relationships
US20060122834A1 (en) * 2004-12-03 2006-06-08 Bennett Ian M Emotion detection device & method for use in distributed systems
US7430238B2 (en) 2004-12-10 2008-09-30 Micronas Usa, Inc. Shared pipeline architecture for motion vector prediction and residual decoding
US7716198B2 (en) 2004-12-21 2010-05-11 Microsoft Corporation Ranking search results using feature extraction
US7454550B2 (en) 2005-01-05 2008-11-18 Xtremedata, Inc. Systems and methods for providing co-processors to computing systems
US7526101B2 (en) 2005-01-24 2009-04-28 Mitsubishi Electric Research Laboratories, Inc. Tracking objects in videos with adaptive classifiers
JP2008530642A (en) 2005-02-07 2008-08-07 ペーアーツェーテー イクスペーペー テクノロジーズ アクチエンゲゼルシャフト Low latency mass parallel data processor
US20060224532A1 (en) 2005-03-09 2006-10-05 Case Western Reserve University Iterative feature weighting with neural networks
US20070132785A1 (en) 2005-03-29 2007-06-14 Ebersole John F Jr Platform for immersive gaming
DE602005013752D1 (en) 2005-05-03 2009-05-20 Seac02 S R L Augmented reality system with identification of the real marking of the object
US20060262140A1 (en) 2005-05-18 2006-11-23 Kujawa Gregory A Method and apparatus to facilitate visual augmentation of perceived reality
US7707163B2 (en) 2005-05-25 2010-04-27 Experian Marketing Solutions, Inc. Software and metadata structures for distributed and interactive database architecture for parallel and asynchronous data processing of complex data and for real-time query processing
US7865492B2 (en) 2005-09-28 2011-01-04 Nokia Corporation Semantic visual search engine
US8094928B2 (en) 2005-11-14 2012-01-10 Microsoft Corporation Stereo video for gaming
US20070162448A1 (en) 2006-01-10 2007-07-12 Ashish Jain Adaptive hierarchy structure ranking algorithm
US9182228B2 (en) 2006-02-13 2015-11-10 Sony Corporation Multi-lens array system and method
US20070233679A1 (en) 2006-04-03 2007-10-04 Microsoft Corporation Learning a document ranking function using query-level error measurements
WO2007123919A2 (en) * 2006-04-18 2007-11-01 Gemini Design Technology, Inc. Method for ranking webpages via circuit simulation
US7840482B2 (en) 2006-06-19 2010-11-23 Exegy Incorporated Method and system for high speed options pricing
US8019594B2 (en) 2006-06-30 2011-09-13 Robert Bosch Corporation Method and apparatus for progressively selecting features from a large feature space in statistical modeling
US7593934B2 (en) 2006-07-28 2009-09-22 Microsoft Corporation Learning a document ranking using a loss function with a rank pair or a query parameter
US7805438B2 (en) 2006-07-31 2010-09-28 Microsoft Corporation Learning a document ranking function using fidelity-based error measurements
US7804984B2 (en) * 2006-07-31 2010-09-28 Lumidigm, Inc. Spatial-spectral fingerprint spoof detection
US8510298B2 (en) 2006-08-04 2013-08-13 Thefind, Inc. Method for relevancy ranking of products in online shopping
US20080126275A1 (en) 2006-09-27 2008-05-29 Crnojevic Vladimir S Method of developing a classifier using adaboost-over-genetic programming
US20080104531A1 (en) 2006-09-29 2008-05-01 Stambaugh Thomas M Spatial organization and display of enterprise operational integration information
US7660793B2 (en) 2006-11-13 2010-02-09 Exegy Incorporated Method and system for high performance integration, processing and searching of structured and unstructured data using coprocessors
US7650317B2 (en) 2006-12-06 2010-01-19 Microsoft Corporation Active learning framework for automatic field extraction from network traffic
US8117137B2 (en) 2007-04-19 2012-02-14 Microsoft Corporation Field-programmable gate array based accelerator system
US7844085B2 (en) 2007-06-07 2010-11-30 Seiko Epson Corporation Pairwise feature learning with boosting for use in face detection
US7958068B2 (en) 2007-12-12 2011-06-07 International Business Machines Corporation Method and apparatus for model-shared subspace boosting for multi-label classification
US7974974B2 (en) 2008-03-20 2011-07-05 Microsoft Corporation Techniques to perform relative ranking for search results
US20090265290A1 (en) 2008-04-18 2009-10-22 Yahoo! Inc. Optimizing ranking functions using click data
US8108324B2 (en) 2008-05-15 2012-01-31 Intel Corporation Forward feature selection for support vector machines
US8015132B2 (en) 2008-05-16 2011-09-06 Samsung Electronics Co., Ltd. System and method for object detection and classification with multiple threshold adaptive boosting

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040111388A1 (en) * 2002-12-06 2004-06-10 Frederic Boiscuvier Evaluating relevance of results in a semi-structured data-base system
US7197497B2 (en) * 2003-04-25 2007-03-27 Overture Services, Inc. Method and apparatus for machine learning a document relevance function
US20050234953A1 (en) * 2004-04-15 2005-10-20 Microsoft Corporation Verifying relevance between keywords and Web site contents
US20050246328A1 (en) * 2004-04-30 2005-11-03 Microsoft Corporation Method and system for ranking documents of a search result to improve diversity and information richness

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103729320A (en) * 2013-12-20 2014-04-16 天津光电通信技术有限公司 Method for implementing CY7C68013 communication on basis of FPGA (field programmable gate array)
CN103813125A (en) * 2014-03-03 2014-05-21 安庆师范学院 Multi-branch digital image processing system

Also Published As

Publication number Publication date
US20080262984A1 (en) 2008-10-23
US8117137B2 (en) 2012-02-14
US20120092040A1 (en) 2012-04-19
US8583569B2 (en) 2013-11-12

Similar Documents

Publication Publication Date Title
US8117137B2 (en) Field-programmable gate array based accelerator system
US8131659B2 (en) Field-programmable gate array based accelerator system
Tomasic et al. Performance of inverted indices in shared-nothing distributed text document information retrieval systems
US20180260710A1 (en) Calculating device and method for a sparsely connected artificial neural network
US11775430B1 (en) Memory access for multiple circuit components
WO2020073211A1 (en) Operation accelerator, processing method, and related device
Ling et al. Design and implementation of a CUDA-compatible GPU-based core for gapped BLAST algorithm
EP3314464B1 (en) Storage and retrieval of data from a bit vector search index
CN104133780A (en) Cross-page prefetching method, device and system
JP6165955B1 (en) Method and system for matching images and content using whitelist and blacklist in response to search query
JP2020518068A (en) Graph matching for optimized deep network processing
Feng et al. Crescent: taming memory irregularities for accelerating deep point cloud analytics
Shahshahani et al. Memory optimization techniques for fpga based cnn implementations
Chang et al. A memory-optimized and energy-efficient CNN acceleration architecture based on FPGA
Imani et al. CAP: Configurable resistive associative processor for near-data computing
WO2015094721A2 (en) Apparatuses and methods for writing masked data to a buffer
Umuroglu et al. A vector caching scheme for streaming fpga spmv accelerators
CN112306951B (en) CNN-SVM resource efficient acceleration architecture based on FPGA
Ke et al. LIRS: Enabling efficient machine learning on NVM-based storage via a lightweight implementation of random shuffling
WO2023124304A1 (en) Chip cache system, data processing method, device, storage medium, and chip
Gonçalves et al. Exploring data size to run convolutional neural networks in low density fpgas
US6842750B2 (en) Symbolic simulation driven netlist simplification
CN101751356A (en) Method, system and apparatus for improving direct memory access transfer efficiency
Xu et al. Fpga-based accelerator design for rankboost in web search engines
WO2021120036A1 (en) Data processing apparatus and data processing method

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 08746374

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 08746374

Country of ref document: EP

Kind code of ref document: A1