AI Hardware Requirements

Technical details about Prisma AI hardware requirements

Supported compute devices by service

The Prisma AI anomaly detection system consists of three components: the prisma-ai-trainer, prisma-ai-tester, and prisma-ai-server. The AI Trainer supports both CPU and GPU training, while the AI Tester and AI Server currently only support CPU inference.

Service	Supported Device Types
Prisma AI Trainer	CPU, GPU (NVIDIA CUDA)
Prisma AI Tester	CPU
Prisma AI Server	CPU

Ensure that your NVIDIA GPU supports hardware accelerated bfloat16 operations. This is supported for all NVIDIA GPUs with compute capability 8.0 or higher (this is the standard starting with ampere, which introduced the RTX 3000 series of consumer GPUs and the A100 datacenter GPU).

Supported GPU driver versions

For services that support GPU acceleration, ensure you are using the correct Docker image. The main Docker image does not ship any GPU support. For GPU support, pull the image with the appropriate suffix (for example, for the Prisma AI Trainer you would pull images.intellitrend.de/prisma/prisma-ai-trainer:7.14.0-cuda for CUDA support).

Vendor	Support Status	CUDA Toolkit	Image Suffix
Nvidia	Supported	CUDA 12.5	-cuda
AMD	Unsupported (Work in progress)	ROCM	-
Intel	Unsupported (Not Planned)	oneAPI	-

CUDA 12.5 requires an NVIDIA GPU driver of at least version 555.x on Linux. You can check the installed driver version with nvidia-smi.

Container runtime

All Prisma AI services run in Docker containers, so a working Docker engine is required on the host. For GPU-accelerated images (-cuda suffix), the host must additionally have:

The NVIDIA GPU driver (which provides nvidia-smi). See the official NVIDIA driver downloads for your distribution.
The NVIDIA Container Toolkit, which exposes the host GPU to Docker containers. Follow NVIDIA’s official installation guide for your distribution.

Once both are installed, verify the container runtime can see the GPU:

docker run --rm --gpus all nvidia/cuda:12.5.0-base-ubuntu22.04 nvidia-smi

Docker Compose example

To expose all GPUs on the host to the AI Trainer, add a deploy.resources.reservations.devices block to the service:

services:
  prisma-ai-trainer:
    image: images.intellitrend.de/prisma/prisma-ai-trainer:7.14.0-cuda
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

To restrict the container to a specific GPU (for example GPU index 0), replace count: all with device_ids: ["0"]:

deploy:
  resources:
    reservations:
      devices:
        - driver: nvidia
          device_ids: ["0"]
          capabilities: [gpu]

Minimum System Specifications

Service	Compute Device	CPU Cores	CPU Features	RAM	VRAM	GPU Compute Capability
Prisma AI Trainer	CPU	4	AVX	8GB	-	-
Prisma AI Trainer	GPU	2		8GB	8GB	≥ 8.0
Prisma AI Tester	CPU	2	AVX	2GB	-	-
Prisma AI Server	CPU	2	AVX	2GB	-	-

Recommended System Specifications

Service	Compute Device	CPU Cores	CPU Features	RAM	VRAM	GPU Compute Capability
Prisma AI Trainer	CPU	16	AVX512	32GB	-	-
Prisma AI Trainer	GPU	4		16GB	16GB	≥ 8.0
Prisma AI Tester	CPU	4	AVX512	8GB	-	-
Prisma AI Server	CPU	4	AVX512	8GB	-	-

The services themselves only persist model artifacts and intermediate training state locally; datasets and trained models are stored in S3, configured at the deployment level in the Prisma installation overview.

For optimal training and inference performance on CPUs your CPU should ideally support the AVX512 instruction set extension. Otherwise ensure that your CPU supports at least AVX2. Both inference and training on CPU will utilize all available cores and will run faster with more CPU cores. Providing more RAM will not provide any speedup unless the system is swapping.