The IO Subsystem for the Modern, GPU-Accelerated Data Center

The NVIDIA MAGNUM IO™ software development kit (SDK) enables developers to remove input/output (IO) bottlenecks in AI, high-performance computing (HPC), data science, and visualization applications, reducing the end-to-end time of their workflows. Magnum IO covers all aspects of data movement between CPUs, GPUs, DPUs, and storage subsystems in virtualized, containerized, and bare-metal environments.

Get Magnum IO Container

Magnum IO Ecosystem

magnum io
The Magnum IO stack contains the libraries developers need to create and optimize application IO across the whole stack: Networking across NVIDIA® NVLink®, Ethernet, and InfiniBand. Local and remote direct storage APIs. In-Networking Compute to accelerate multi-node operations. And IO management of networking hardware.

Flexible Abstractions

Magnum IO enables AI, data analytics, visualization, and HPC developers to innovate and accelerate applications built using common high-level abstractions and APIs.

Architected for Scale

Magnum IO technologies allow for scaling up computation to multiple GPUs via NVLink and PCIe and across multiple nodes on InfiniBand and Ethernet at data center scale.

Advanced IO Management

Advanced telemetry and monitoring built with NVIDIA NetQ™ and NVIDIA UFM® help users to configure, troubleshoot, and fine-tune the interconnect infrastructure for peak performance.

Magnum IO Components

Network IO

Storage IO

IO Management

Accelerating IO Across Applications

Deep Learning

Magnum IO networking provides both point-to-point functions like send and receive, and collectives like AllReduce for deep learning training at scale. The collective APIs hide low-level optimizations like topology detection, peer-to-peer copy, and multi-threading to simplify deep learning training. Send/receive can enable users to accelerate giant deep learning models too big to fit in one GPU’s memory. GPUDirect Storage can also help alleviate IO bottlenecks from local or remote storage by bypassing bounce buffers on the CPU host.

High-Performance Computing

To unlock next-generation discoveries, scientists rely on simulation to better understand complex molecules for drug discovery, physics for new sources of energy, and atmospheric data to better predict extreme weather patterns. Magnum IO exposes hardware-level acceleration engines and smart offloads, such as RDMA, GPUDirect, and NVIDIA SHARP, while bolstering the high bandwidth and ultra-low latency of high data rate (HDR) 200Gb/s InfiniBand. This delivers the highest performance and most efficient HPC and machine learning deployments at any scale.

Data Analytics

Data science and machine learning are the world's largest compute segments. Modest improvements in the accuracy of predictive machine learning models can translate into billions of dollars. To enhance accuracy, the RAPIDS™ Accelerator for Apache Spark library has a built-in shuffle based on NVIDIA Mellanox UCX® that can leverage GPU-to-GPU communication and RDMA capabilities. Combined with NVIDIA networking, Magnum IO, GPU-accelerated Spark 3.0, and RAPIDS, the NVIDIA data center platform can speed up these huge workloads at unprecedented levels of performance and efficiency.

Get Started Using the Magnum IO Developer Environment

The Magnum IO Developer Environment is available as a container with the latest versions of all libraries, development tools, and profiling tools needed to begin development and optimization. The optimized applications can then be run in virtualized, containerized, or bare-metal environments.

Download on Github Download on NGC