IEVD Feasibility on SIMATIC Rack Servers Equipped with NVIDIA GPUs

0
Dear Colleagues,I am seeking clarification regarding a potential IEVD deployment architecture. As is well understood, leveraging GPU-accelerated capabilities effectively often requires the use of applications such as the AI Inference Server on a BX-59A IED platform. However, in a scenario where a BX-59A is not available, I am evaluating an alternative architecture based on an RS-828A Rack Server IPC equipped with an NVIDIA L4.. GPU.In this setup, the intention is to run multiple IEVD instances within separate containers (for example 4 independent containerized instances) on the same server. With this in mind, I would appreciate clarification on the following points:Can IEVD fully utilize the server's GPU resources in this architecture and benefit from GPU-accelerated inference capabilities when deployed in containerized environments on an SIMATIC IPC RS-828A or RS-717A with NVIDIA L4 or others?Is there any validated deployment, reference architecture, benchmark, or officially supported use case that follows a similar approach?Are there any known limitations, prerequisites, resource allocation requirements or best practices that should be taken into account when deploying multiple IEVD instances sharing the same GPU resources?Any guidance regarding the technical feasibility, supportability and recommended deployment approach for this scenario would be highly appreciated.Thank you in advance for your support.
asked
2 answers
1

Hi Atakan,

Thank you for your questions.

Could you please let us know which hypervisor you intend to use for running IEVD?

Regarding your questions:

  • GPU utilization is currently not supported in IEVD. Official GPU support will be introduced with IEVD version 1.28, which is planned for release within the next few weeks.
  • Once GPU support becomes officially available, it will only be supported via PCI passthrough, meaning the GPU must be assigned directly to a single IEVD instance.
  • Consequently, the supported configuration will be one GPU per IEVD instance. Shared GPU usage between multiple IEVDs will not be supported.

Please let us know if you require any additional information.

Best,

Pascal

answered
0

Hi Pascal,


Thank you for your clarification.


This means that an architecture with multiple containerized IEVD instances sharing the same GPU on a single server would not be a supported approach.


Regarding your question about the hypervisor, we are currently evaluating the deployment concept and have not yet finalized our hypervisor selection. Our intention was first to understand whether the target architecture would be technically feasible and officially supportable from an IEVD perspective.


So, after the official IEVD 1.28 release:


My understanding is that if a server is equipped with multiple discrete GPUs, it should be possible to run one IEVD instance per GPU by assigning each IEVD instance a dedicated GPU via PCI passthrough. Could you please confirm whether this understanding is correct?


In that case, we would be able to use a single server hosting multiple IEVD instances, each assigned to its own dedicated GPU.

answered