Home > AI Solutions > Artificial Intelligence > White Papers > Dell Automotive Reference Architecture > Dell Validated Design for Generative AI with AMD
The Dell Validated Design for Generative AI with AMD Accelerators describes a layered software stack of software components and technologies that are used to build, train, verify, deploy, and manage AI models. This stack encompasses various layers, each serving a specific purpose in the AI development life cycle.
The Dell Reference Design (DRD) for generative AI with AMD accelerators (referred to as the AMD DRD) brings not only generative AI to the world’s enterprise data centers but also general AI workloads. This DRD is a full-stack solution that enables enterprises to create, run, and deploy custom AI models. Dell Technologies has designed a scalable, modular, and high-performance infrastructure that enables enterprises to create a wave of generative and general AI solutions that help reinvent their industries to accomplish tasks that were previously unimaginable. The DRD software stack focuses on making AI model creation and development easier for the data scientist.
The following figure shows the layers of the AMD DRD’s AI software stack. Dell Technologies supports the components shown in blue, and AMD supports the components shown in black. Partner applications and open-source components are shown on a white background.
Figure 3. DRD for AI with AMD software stack
The following table summarizes each of the layers of the DRD for AI with AMD software stack:
Description | Supported | Comments |
Foundational models | Partners | Autonomous driving companies do not typically use LLMs. A few innovative and forward-thinking autonomous driving companies use advanced neural network models called VLMs which are similar to LLMs. This advancement in AI model development is called Autonomous Vehicle Technology 2.0 (AV2.0). |
Supported libraries | AMD | Libraries include MIVisionX for vision, whisper.ai for speech, vLLM for LLMs, and Recommenders. |
AIOps and MLOps platform | Partners | This layer is a subset of the many AI-compatible framework possibilities. |
Workload manager | Dell Technologies | Kubernetes or Slurm |
Omnia | Dell Technologies | The software infrastructure layer consists of host provisioning, Kubernetes deployment, and monitoring using Graphana panels to view telemetry data. |
ROCm libraries | AMD | ML and Computer Vision: CK, MIGraphX, MIOpen, MIVisionX, RPP, rocDecode, rocAL Communication: RCCL Math: hipBLAS, hipBLASLt, rocBLAS, hipSolver, rocFFT, rocSPARCE, rocWMMA |
ROCm Compiler and Tools | AMD | LLVM, hipCC, HIPify, ROCgdb, ROCProfiler, ROCTracer |
ROCm Runtime | AMD | AMD CLR, HIP, ROCr, Kubernetes device plugin, Open Containers Initiative (OCI) compatible with Docker, containerd, apptainer, and so on |
Kubernetes runtime environment | Kubernetes | Kubernetes device plugin, Open Containers Initiative (OCI) compatible with Docker, CRIO-O, containerd, Apptainer etc |
Operating system | Operating system provider | Enterprise Linux |
Hardware Layer | Dell Technologies and AMD | Compute, networking, and storage hardware infrastructure |