Sambanova Systems, Inc.

Kin Hing Leung

Gregory Frederick Grohoski

Sumti Jairath

Manish K Shah

Raghu Prabhakar

Ravinder Kumar

Martin Russell Raumann

Bandish B Shah

Ram Sivaramakrishnan

Arnav Goel

Ranen Chatterjee

Dawei Huang

Emre Ali Burhan

Manoj Unnikrishnan

Qi Zheng

Matthew Thomas Grimm

Sitanshu Gupta

Yuan Lin

Luca Boasso

David Brian Jackson

Arvind Krishna Sujeeth

Pramod Nataraja

Conrad Alexander Turlik

Andrew Deng

Maran Wilson

Greg Dykema

Milad Sharif

Guoyao Feng

Kuan Zhou

Taylor Lee

Tianyu Sun

Yue Fu

Mark Luttrell

David Alan Koeplinger

Ajit Punj

Joshua Brot

Feng Sheng

Likun Hao

A system for a data-parallel execution of at least two implementations of an application on reconfigurable processors with different layouts is presented. The system comprises a pool of reconfigurable data flow resources with data transfer resources that interconnect first and second reconfigurable processors having first and second layouts that impose respective first and second constraints for the data-parallel execution of the application. The system further comprises an archive of configuration files and a host system that is operatively coupled to the first and second reconfigurable processors. The host system comprises first and second compilers that generate for the application, based on the respective first and second constraints, first and second configuration files that are stored in the archive of configuration files and adapted to be executed data-parallel compatible on respective first and second reconfigurable processors.

System of heterogeneous reconfigurable processors for the data-parallel execution of applications

A system with a cost estimation tool for estimating a realized bandwidth consumption of a logical edge between a logical producer unit and a logical consumer unit of an operation unit graph during placement and routing of the logical producer unit, the logical consumer unit, and the logical edge onto a reconfigurable processor is presented as well as a method of operating such a cost estimation tool and a non-transitory computer-readable storage medium including instructions that, when executed by a processing unit, cause the processing unit to operate such a cost estimation tool The cost estimation tool may be configured to determine the realized bandwidth consumption of the tentative assignment based on an upper bandwidth limit of the logical edge, an end-to-end bandwidth, a scaling factor of a realized bandwidth, and a congestion estimation of the physical link.

Operating a cost estimation tool for placing and routing an operation unit graph on a reconfigurable processor

A cost estimation tool in a system for implementing an operation unit graph on a reconfigurable processor is presented as well as a method of operating a cost estimation tool for determining scaled logical edge bandwidths in an operation unit graph in preparation of placing and routing the operation unit graph onto a reconfigurable processor. The cost estimation tool may be configured to receive the operation unit graph, divide the operation unit graph in first and second subgraphs, determine maximum latencies of the first and second subgraphs, and determine a scaled logical edge bandwidth of a logical edge that couples a first logical unit of M logical units in the first subgraph with a second logical unit of N logical units in the first subgraph based on M, N, and scaled bandwidth limits of the M and N logical units.

Estimating a scaled cost of implementing an operation unit graph on a reconfigurable processor

A method for placing, routing and using compute units and memory units in a reconfigurable computing grid includes receiving a placement graph for a computing task that defines a set of unplaced memory units, a set of unplaced compute units and data connections between the unplaced memory units and the unplaced compute units, the data connections comprising primary connections corresponding to the primary ports of the unplaced compute units and secondary connections corresponding to the secondary ports of the unplaced compute units. The method also includes forming a subgraph for each unplaced memory unit having a primary connection, each subgraph comprising the unplaced memory unit and each unplaced compute unit connected to the unplaced memory unit via a primary connection. The method also includes placing each formed subgraph as a cluster on the reconfigurable computing grid. A corresponding computer program product and system are also disclosed herein.

Cluster-based placement and routing of memory units and compute units in a reconfigurable computing grid

A computing system includes an array of configurable units made up of sub-arrays of configurable units. Each sub-array has a first number of configurable compute units and a second number of configurable memory units with a first spatial arrangement. Each configurable unit includes a configuration data store. The system also includes a statically configurable bus system coupled to the configurable units and a tag indicating a sub-array of configurable units having a defect. A defect-aware configuration controller sends configuration data to the configuration data stores to implement a data processing operation using the array of configurable units by generating static route control signals for the statically configurable bus system, based on the tag and without support of a host processor, to send a portion of the configuration data targeted to the sub-array having the defect to a configuration data store of an alternative sub-array of configurable units in the array.

Avoiding use of a subarray of configurable units having a defect

A system for executing an application on a pool of reconfigurable processors with first and second reconfigurable processors having first and second architectures that are different from each other is presented. The system comprises an archive of configuration files with first and second configuration files for executing the application on the first and second reconfigurable processors, respectively, and a host system that is operatively coupled to the first and second reconfigurable processors. The host system comprises a runtime processor that allocates reconfigurable processors for executing the application and an auto-discovery module that is configured to perform discovery of whether the reconfigurable processors include at least one of the first reconfigurable processors and whether the reconfigurable processors include at least one of the second reconfigurable processors. The runtime processor starts execution of the first and second configuration files in dependence upon the discovery of the auto-discovery module.

System for executing an application on heterogeneous reconfigurable processors

A reconfigurable data processor comprises an array of configurable units configurable to allocate a plurality of sets of configurable units in the array to implement respective execution fragments of the data processing operation. Quiesce logic is coupled to configurable units in the array, configurable to respond to a quiesce control signal to quiesce the sets of configurable units in the array on quiesce boundaries of the respective execution fragments, and to forward quiesce ready signals for the respective execution fragments when the corresponding sets of processing units are ready. An array quiesce controller distributes the quiesce control signal to configurable units in the array, and receives quiesce ready signals for the respective execution fragments from the quiesce logic.

Quiesce reconfigurable data processor

A system for training parameters of a neural network includes a processing node with a processor reconfigurable at a first level of configuration granularity and a controller reconfigurable at a finer level of configuration granularity. The processor is configured to execute a first dataflow segment of the neural network with training data to generate a predicted output value using a set of neural network parameters, calculate a first intermediate result for a parameter based on the predicted output value, a target output value, and a parameter gradient, and provide the first intermediate result to the controller. The controller is configured to receive a second intermediate result over a network, and execute a second dataflow segment, dependent upon the first intermediate result and the second intermediate result, to generate a third intermediate result indicative of an update of the parameter.

Training a neural network using a non-homogenous set of reconfigurable processors

The technology disclosed relates to inter-node execution of configuration files on reconfigurable processors using network interface controller (NIC) buffers. In particular, the technology disclosed relates to a runtime logic that is configured to execute configuration files that define applications and application data for applications using a first reconfigurable processor connected to a first host, and a second reconfigurable processor connected to a second host. The first reconfigurable processor is configured to push input data for the applications in a first plurality of buffers. The first host is configured to cause a first network interface controller (NIC) to stream the input data to a second plurality of buffers from the first plurality of buffers. The second host is configured to cause a second NIC to stream the input data to the second reconfigurable processor from the second plurality of buffers.

Inter-node execution of configuration files on reconfigurable processors using network interface controller (NIC) buffers

The technology disclosed relates to runtime execution of functions across reconfigurable processor. In particular, the technology disclosed relates to a runtime logic that is configured to execute a first set of functions in a plurality of functions and/or data therefor on a first reconfigurable processor, and a second set of functions in the plurality of functions and/or data therefor on additional reconfigurable processors. Functions in the second set of functions and/or the data therefor are transmitted to the additional reconfigurable processors using one or more of a first reconfigurable processor-to-additional reconfigurable processors buffers, and results of executing the functions and/or the data therefor on the additional reconfigurable processors are transmitted to the first reconfigurable processor using one or more of additional reconfigurable processors-to-first reconfigurable processor buffers.

Runtime execution of functions across reconfigurable processor

A system for executing a graph partitioned across a plurality of reconfigurable computing units includes a processing node that has a first computing unit reconfigurable at a first level of configuration granularity and a second computing unit reconfigurable at a second, finer, level of configuration granularity. The first computing unit is configured by a host system to execute a first dataflow segment of the graph using one or more dataflow pipelines to generate a first intermediate result and to provide the first intermediate result to the second computing unit without passing through the host system. The second computing unit is configured by the host system to execute a second dataflow segment of the graph, dependent upon the first intermediate result, to generate a second intermediate result and to send the second intermediate result to a third computing unit, without passing through the host system, to continue execution of the graph.

Executing a neural network graph using a non-homogenous set of reconfigurable processors

A reconfigurable processor comprises an array of processing units and an instrumentation network. The array of processing units is configured to execute runtime events to execute an application. The instrumentation network is operatively coupled to the array of processing units. The instrumentation network comprises a control bus configured to form control signal routes in the instrumentation network. The instrumentation network further comprises a plurality of instrumentation counters having inputs and outputs connected to the control bus and to the processing units. Instrumentation counters in the plurality instrumentation units are configurable to consume control signals on the inputs and produce counts of the runtime events on the outputs.

Instrumentation networks for data flow graphs

A data processing system comprises memory, compile time logic, runtime logic, and instrumentation profiling logic. The memory stores a dataflow graph for an application. The dataflow graph has a plurality of compute nodes that are configured to be producers to produce data for execution of the application, and to be consumers to consume the data for execution of the application. The compile time logic partitions execution of the dataflow graph into stages. Each of the stages has one or more compute nodes, one or more producers, and one or more consumers. The runtime logic determines a processing latency for each of the stages by calculating time elapsed between producers of a particular stage receiving input data and consumers of the particular stage receiving output data. The instrumentation profiling logic generates performance statistics for the dataflow graph based on the processing latency determined for each of the stages.

Compile time instrumentation of data flow graphs

A data processing system comprises a pool of reconfigurable data flow resources and a runtime processor. The pool of reconfigurable data flow resources includes arrays of physical configurable units and memory. The runtime processor includes logic to receive a plurality of configuration files for user applications. The configuration files include configurations of virtual data flow resources required to execute the user applications. The runtime processor also includes logic to allocate physical configurable units and memory in the pool of reconfigurable data flow resources to the virtual data flow resources and load the configuration files to the allocated physical configurable units. The runtime processor further includes logic to execute the user applications using the allocated physical configurable units and memory.

Runtime patching of configuration files

A system includes a multidimensional array of homogenous Functional Configurable Units (FCUs), coupled using a multidimensional array of switches, and a parameter store on the device which stores parameters that tag a subarray of FCUs as unusable. Technologies are described which change the pattern of placement of configuration data, in dependence on the tagged subarray, by changing the routing through the array of switches. As a result, a multidimensional array of FCUs having unusable elements can still be used.

Growing community of inventors

Cupertino, CA, United States of America

Kin Hing Leung