US 2026/0228044 A1Published application

Hardware resource reservation in a reconfigurable dataflow architecture

AssigneeSambaNova Systems, Inc.

InventorsPushkar Shridhar NANDKAR, Raghunath SHENBAGAM

Life of this patent

Estimating the term…

Loading patent…

Abstract

A system includes a first RDU including a local interconnect usable for coupling to a second RDU, a host coupled to the first RDU using a system interconnect coupled to the local interconnect, and a local memory on the host. The system interconnect is accessed by an RDRT driver running on the host. The RDRT driver may record a workload reservation specifying an RDU resource usable to execute a workload, selected from: an RDU tile, a local host memory, an HBM, a DDR memory, or a peripheral port. The reservation specifies a condition of shared accessibility, exclusive accessibility, or allowed substitution of the RDU resource with a second RDU resource. The workload can include an AI/ML application for execution by the system.

Claims20

Dependent claims

A system comprising:

a first reconfigurable dataflow unit (RDU) including a local interconnect usable for coupling to a second RDU;
a host coupled to the first RDU using a system interconnect coupled to the local interconnect, wherein the system interconnect is accessed by a reconfigurable dataflow runtime (RDRT) driver executing under an operating system running on the host;
a local memory on the host;
wherein the RDRT driver is configured to:
record a workload reservation specifying an RDU resource usable to execute a workload by the system, the RDU resource selected from:
an RDU tile including multiple pattern compute units (PCUs), a local memory at the host, a high-bandwidth memory (HBM), a dual data rate (DDR) memory module pair, or a peripheral port, the workload reservation specifying at least one condition selected from:
shared accessibility of the RDU resource,
exclusive accessibility of the RDU resource, or
allowed substitution of the RDU resource with a second RDU resource included in the system,
wherein the workload includes an artificial intelligence (AI)/machine learning (ML) application for execution using the system.

7 more elements

A method comprising:

receiving, at a reconfigurable dataflow runtime (RDRT) driver, a first indication of a first workload for at least partial execution using a reconfigurable dataflow unit (RDU) system including a first RDU, wherein the RDRT driver executes on an operating system executing on a host coupled to the first RDU using a system interconnect;
and
responsive to receiving the first indication, recording a workload reservation specifying an RDU resource in the RDU system for execution of the first workload, the RDU resource selected from an RDU tile including multiple pattern compute units (PCUs), a local memory at the host, a high-bandwidth memory (HBM), a dual data rate (DDR) memory module pair, or a peripheral port, the workload reservation specifying at least one condition selected from:
shared accessibility of the RDU resource,
exclusive accessibility of the RDU resource, or
allowed substitution of the RDU resource with a second RDU resource,
wherein the first workload includes an artificial intelligence (AI)/machine learning (ML) application for execution using the RDU system.

4 more elements

Tangible computer-readable media comprising instructions executable by a computer system to:

receive, at a reconfigurable dataflow runtime (RDRT) driver, a first indication of a first workload for at least partial execution using a reconfigurable dataflow unit (RDU) system including a first RDU, wherein the RDRT driver executes on an operating system executing on a host coupled to the first RDU using a system interconnect;
responsive to receiving the first indication, record a workload reservation specifying an RDU resource in the RDU system for execution of the first workload, the RDU resource selected from an RDU tile including multiple pattern compute units (PCUs), a local memory at the host, a high-bandwidth memory (HBM), a dual data rate (DDR) memory module pair, or a peripheral port, the workload reservation specifying at least one condition selected from:
shared accessibility of the RDU resource,
exclusive accessibility of the RDU resource, or
allowed substitution of the RDU resource with a second RDU resource,
wherein the first workload includes an artificial intelligence (AI)/machine learning (ML) application for execution using the RDU system.

3 more elements

Dependencies are inferred from the claim text.

Facts

Application: 19/042,467Filed 31 Jan 2025
Published: 6 Aug 2026
Term ends: Not estimatedThe source records don’t support an estimate.

Classification12 codes

G06F 9/5016
The resources being hardware resources other than CPUs, Servers and Terminals: the resource being the memory
G06F 9/5027
To service a request: the resource being a machine, e.g. CPUs, Servers, Terminals
G06F 9/4881
By program, e.g. task dispatcher, supervisor, operating system: scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
G06F 9/5005
Allocation of resources, e.g. of the central processing unit [CPU]: to service a request