# NVIDIA/gdrcopy

A fast GPU memory copy library based on NVIDIA GPUDirect RDMA technology

Repository: https://github.com/NVIDIA/gdrcopy
Canonical: https://ross.abutalabs.com/products/gdrcopy
Language: C
License: MIT
License Family: permissive
Topics: gpu-memory, nvidia, kernel-mode-driver, gpudirect-rdma, libraries, linux
Last push: 2026-07-14T23:16:03+00:00

## Health v2 (maintenance only)
Score: 84/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 92, release rhythm 64, longevity 100
- inputs: {"age_days": 4286, "days_push": 50, "days_rel": 82, "gap_med": 96.0, "n_releases_24m": 7}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1412, forks 194 (observed 2026-08-28T04:04:39.423465+00:00)

## What it is
GDRCopy is a low-latency GPU memory copy library built on NVIDIA GPUDirect RDMA technology, allowing CPU-driven copies between host and GPU memory with minimal overhead. It includes a kernel-mode driver, a user-space API for mapping GPU memory as host-accessible memory, and benchmark/test applications for bandwidth, latency, and API performance.

## Use cases
- copy data between CPU and GPU with very low latency
- map GPU memory into the host address space for direct CPU access
- reduce cudaMemcpy overhead for small transfers
- measure GPU memory copy bandwidth and latency
- run ping-pong latency tests between GPU and CPU
- sanity-test GPUDirect RDMA driver and library installation
- build HPC or networking applications that touch GPU memory from the host

## When to choose
- You need low-latency host-to-device transfers and cudaMemcpy's 6-7us overhead is too high
- You run Linux with an NVIDIA Data Center or RTX GPU (Kepler or newer) and can install a DKMS kernel module
- You want CPU mappings of GPU memory for direct reads/writes or for wiring GPU memory into RDMA-capable NICs
- You need to benchmark or validate GPU memory copy performance

## When to avoid
- You need maximum device-to-host bandwidth - GPU BAR mappings cannot be prefetched, so D-H reads are slow; use cudaMemcpy for bulk transfers
- You are on Windows or macOS - GDRCopy is Linux-only
- You cannot or do not want to install kernel modules (DKMS/kmod) on your system
- You use consumer GeForce GPUs, which do not support GPUDirect RDMA

## Facets
- artifact type: library
- maturity: active
- function: gpu-computing, benchmarking
- domain: gpu-computing, hardware, performance, developer-tools
- platform: c, cli
- tags: gpudirect-rdma, gpu-memory, cuda, nvidia, kernel-module, dma, low-latency, memory-copy, hpc, host-device-transfer, dkms, linux, gpu

## Member repositories
- NVIDIA/gdrcopy (main) score 84

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:39.423465+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:38:19.250655+00:00, confidence not recorded.
  - readme: https://github.com/NVIDIA/gdrcopy (fetched 2026-08-28T04:04:39.423465+00:00, sha 210a5cb7807a)
- Data as of 2026-08-30T08:39:29.467469+00:00.