US 12,717,592 B2Granted

Hardware layer-by-layer program control in a reconfigurable dataflow processor

AssigneeSambaNova Systems, Inc.

InventorsManish K. Shah, Denis Sokolov, Raghu Prabhakar, Arjun Sabnis, Joshua Earle Polzin, Arnav Goel

Life of this patent

Estimating the term…

Loading patent…

Abstract

A reconfigurable data processor includes a bus system, an array of configurable units, and a configuration load controller connected to the bus system and coupled to a memory. The configuration load controller incorporates a first set of registers accessible from a host processor for storing addresses of a first configuration file, a second set of registers loaded by loading a configuration file for storing addresses of a second configuration file, and an address generation unit with working address registers. The processor is configured to load a first configuration file from the memory and initiate execution based on a request from runtime software. Additional configuration files are automatically loaded upon completion of a previous configuration file based on information stored in the previous configuration file.

Claims19

Dependent claims

A system for sequentially loading configuration files into a reconfigurable dataflow processor (RDP) for executing sections of a computation graph, the system comprising:

a host computer configured to pre-load a plurality of configuration files and associated argument load files for different sections of a computation graph into a memory, including a first configuration file and associated first argument load file and a second configuration file and associated second argument load file;
and
an RDP comprising:
a set of dataflow processing units to execute the sections of the computation graph;
at least one control/status register (CSR);
configuration load circuitry configured to load a configuration file and associated argument load file of the plurality of configuration files and the associated argument load files from the memory into the set of dataflow processing units of the RDP;
and
a CLAA register and an ALAA register;
wherein the host computer is further configured to store first memory addresses of the first configuration file and the associated first argument load file into the at least one CSR and initiate a configuration load and execute process in the RDP;
and
wherein the configuration load circuitry is further configured to:
load the first configuration file and associated first argument load file from the memory into the set of dataflow processing units and initiate execution of a first section of the computation graph by the set of dataflow processing units of the RDP based on the first configuration file and associated first argument load file;
upon completion of the execution of the first section of the computation graph, load a second address based on first information in the first configuration file into the CLAA register and load the second configuration file from the memory into the set of dataflow processing units using the second address stored in the CLAA register;
upon completion of loading the second configuration file into the set of dataflow processing units, load a third address based on second information in the second configuration file into the ALAA register and load the associated second argument load file from the memory into the set of dataflow processing units using the third address stored in the ALAA register;
and
upon completion of loading the associated second argument load file into the set of dataflow processing units, initiate execution of a second section of the computation graph by the set of dataflow processing units of the RDP based on the second configuration file and associated second argument load file.

13 more elements

A reconfigurable dataflow processor comprising:

a set of dataflow processing units:
a CLA register to hold a first memory address;
a CLAA register to hold a third memory address;
and
configuration load circuitry coupled to the set of dataflow processing units and a memory to load a first configuration file from the first memory address in the memory into the set of dataflow processing units in response to a first signal, and to load a second configuration file from the third memory address in the memory into the set of dataflow processing units in response to a second signal;
wherein the first configuration file is executed by the set of dataflow processing units;
the CLAA register is loaded with the third memory address, calculated based on first information included in the first configuration file, in response to completion of execution of the first configuration file by the set of dataflow processing units;
the second signal is generated upon completion of loading the CLAA register with the third memory address;
and
the second configuration file is executed by the set of dataflow processing units.

7 more elements

A method for sequentially loading and executing sections of a computation graph on a reconfigurable dataflow processor (RDP), the method comprising:

storing a plurality of configuration files for different sections of a computation graph, including a first configuration file for a first section and a second configuration file for a second section, into memory by a host computer;
storing, by the host computer, a first memory address of the first configuration file into a CLA register of the RDP;
initiating, by the host computer, a configuration load process in the RDP;
loading, by configuration load circuitry of the RDP using the first memory address, the first configuration file from the memory into a set of dataflow processing units of the RDP;
executing, by the set of dataflow processing units, the first section of the computation graph based on the first configuration file;
sending a first signal from a dataflow processing unit in the set of dataflow processing units to the configuration load circuitry upon completion of the execution of the first section;
storing a third memory address, based on first information included in the first configuration file, into a CLAA register of the RDP in response to receiving the first signal;
loading, by the configuration load circuitry, after the third memory address has been stored into the CLAA register, the second configuration file from the memory into the set of dataflow processing units using the third memory address from the CLAA register;
and
executing, by the set of dataflow processing units, the second section of the computation graph based on the second configuration file.

7 more elements

Dependencies are inferred from the claim text.

Facts

Application: 18/824,846Filed 4 Sep 2024 · us provisional application
Priority: 5 Sep 2023
Published: US 2025/0077239 A16 Mar 2025
Granted: 25 Aug 2026
Term ends: Not estimatedThe source records don’t support an estimate.

Classification4 codes

G06F 9/44505
Program loading or initiating: configuring for program initiating, e.g. using registry, configuration files
G06F 9/5038
The resource being a machine, e.g. CPUs, Servers, Terminals: considering the execution order of a plurality of tasks, e.g. taking priority or time…
G06F 2209/5017
Indexing scheme relating to G06F9/50: task decomposition
G06F
Computing and data processing