NVIDIA cuObject Server Release Notes#

Release information for NVIDIA® cuObject server.

Version	Date	cuObject server libraries with CUDA Release
v2.1.0	Nov 4, 2026	13.5
v2.0.0	Aug 13, 2026	13.4
v1.2.0	May 26, 2026	13.3
v1.0.0	Jan 12, 2026	13.1.1

Release information for NVIDIA® cuObject server for developers and users.

cuObject is a high-performance suite of libraries designed to enable direct data transfers between GPU memory or system memory and object storage (S3-compatible) solution via RDMA. By relying on RDMA operations, rather than TCP transfer methods, cuObject both avoids using CPU kernel code for TCP processing and can bypass the CPU for data payloads. cuObject eliminates the traditional bottleneck of staging data in local scratch file systems, enabling high-throughput data ingestion for AI training and inference at scale.

Architecture Overview#

cuObject consists of two primary components:

cuObjClient: Provides client-side APIs for GET/PUT operations with user-defined callbacks for control path communication, while data path operations leverage RDMA for high-performance transfers.
cuObjServer: Implements the RDMA-accelerated server side for S3-compatible object storage services, supporting multi-threaded concurrency, automatic connection management, and Dynamically Connected (DC) transport.

Refer to the following guides for more information about cuObject: