Bitfusion Compatibility and Interoperability

Bitfusion 2.5.0 and Bitfusion 2.5.1

Bitfusion works with the hardware and software stack of AI/ML environments and applications. In bullet form, this stack might look like the following:

  • Application (often Python)
  • Framework (examples include TensorFlow and PyTorch)
  • AI/ML and other libraries (examples include CuDNN, CUBLAS, CUDA runtime driver)
  • CUDA driver (libcuda.so)
  • Hardware driver (NVIDIA driver, nvidia.ko. provides an interface, on Linux, called Resource Manager)
  • OS (a Linux distribution)
  • GPU (example: NVIDIA T4)

This page discusses the versions, models, and products that are compatible with Bitfusion 2.5.1 and 2.5.1.

 

Frameworks

Bitfusion is not designed for specific frameworks. Below, however, we do discuss the frameworks we use to test Bitfusion.

TensorFlow

In general, you can choose any version of TensorFlow as long as it works with a supported version of CUDA.
The most heavily tested versions are  1.13.1, 1.15, and 2.2 

TensorRT

In general, you can choose any version of TensorRT as long as it works with a supported version of CUDA.
The version used in testing is 7.1.3

PyTorch

In general, you can choose any version of PyTorch as long as it works with a supported version of CUDA.
The most heavily tested versions are 1.1, 1.4, and 1.6.
 

Libraries

CuDNN

In general, you can choose any version of CuDNN as long as it works with a supported version of CUDA.
Most heavily tested are various 7.x versions and version 8.0.1

Other libraries

For the most part, the versions of all other libraries, are dictated by the version of the framework and CUDA.  The list is lengthy, dynamic, and cannot be maintained here. Please contact support if you experience any issues.

 

CUDA Driver

CUDA Toolkit

Bitfusion supports CUDA 9.2, 10.0, 10.2 and 11.0
 

NVIDIA Driver

In general, NVIDIA drivers support concurrent and previous versions of CUDA
Bitfusion 2.5.x requires NVIDIA driver 450 or later

GPUs

Bitfusion has general support for datacenter class (qualify for DirectPath I/O on ESXi), CUDA-supporting, NVIDIA GPUs.
The GPUs’ actively tested with Bitfusion:

  • P40
  • V100
  • T4
  • A100

Operating Systems

Bitfusion before 2.5.0

  • Ubuntu 16.04
  • Ubuntu 18.04
  • CentOS 7.0
  • RHEL 7.4 and higher minor versions

Bitfusion 2.5.0

  • Ubuntu 16.04
  • Ubuntu 18.04
  • Ubuntu 20.04
  • CentOS 7.0
  • CentOS 8.0
  • RHEL 7.4 and higher minor versions

Filter Tags

AI/ML Application Acceleration Modern Applications vSphere vSphere 6.5 vSphere 6.7 vSphere 7 Document Deployment Considerations Intermediate Design Deploy