AI Appliances
[01] ai-station-essential
AI Station Essential – Single GPU AI Workstation
Pre-validated. Pre-loaded. Ready for training. Designed for data scientists, university laboratory deep learning education, and small teams for lightweight inference. The ideal starting point for AI development – no complex configuration required to begin model training.
- $ Ubuntu 22.04 LTS operating system
- $ NVIDIA Driver 550+ GPU drivers
- $ CUDA 12.4 computing platform
- $ Docker container runtime
- $ PyTorch 2.3 deep learning framework
- $ TensorFlow 2.15 machine learning library
Use Cases
Data scientists conducting new model prototype development, university laboratory deep learning courses, small teams running lightweight inference tasks. Suitable for budget-sensitive startups and research projects.
Optional Upgrades
Second GPU acceleration card, liquid cooling system, 5-year extended warranty, additional NVMe storage expansion, professional workstation graphics card replacement options.
[02] ai-station-advanced
AI Station Advanced – Multi-GPU Rack System
Pre-validated. Pre-loaded. Built for distributed training. A multi-GPU training platform designed for medium-sized teams, supporting PyTorch DDP and Horovod distributed training frameworks. This system can serve as a single-node high-performance workstation or the starting point for a multi-node cluster.
Software & Performance
Pre-installed with Ubuntu 22.04 LTS, NVIDIA Driver 550+, CUDA 12.4/cuDNN, Docker and NVIDIA Container Toolkit, Kubernetes single-node configuration, PyTorch 2.3 and TensorFlow 2.15 distributed training support.
Capable of training ResNet-152 in hours, BERT-base models in days. Supports multi-GPU parallelization (DDP, Horovod), suitable for teams scaling from single workstations to clusters.
Ideal Users
Medium-sized teams conducting distributed model training, multi-GPU research projects (computer vision, natural language processing), enterprise AI development teams, university laboratories and corporate research groups.
Expansion Options
Upgrade to NVIDIA H100 or L40S GPUs, add second node to build 8-GPU cluster, InfiniBand interconnect for high-speed multi-node training, extend warranty to 5 years.
[03] ai-pod-inference
AI Pod Inference – Edge Inference Server
Compact. Rugged. Pre-loaded with inference software stack. A real-time inference server designed for edge AI deployment, suitable for manufacturing quality control, retail analysis, security monitoring, and other low-latency edge scenarios. Small form factor design enables flexible deployment in space-constrained edge environments.
Performance Metrics
Configured with 4 GPUs: ResNet-50 inference > 100+ inferences/second, latency < 10ms. Configured with 2 GPUs: balanced cost-performance, supports 4-6 video streams, suitable for medium-scale deployments. Full TensorRT optimization ensures minimum latency.
Application Scenarios
Real-time computer vision (manufacturing, retail, security), edge AI deployments (low latency requirements), remote sites (where cloud centralization is not feasible), edge video analytics (single server supports up to 6 streams).
Pre-loaded Software
Ubuntu 20.04/22.04 LTS, NVIDIA AI production drivers, CUDA/TensorRT, Triton Inference Server (optional), Docker and NVIDIA Container Toolkit, common model optimization scripts.
[04] ai-pod-training
AI Pod Training – 8-GPU Cluster Node
High-density. Pre-wired. Cluster-ready. An enterprise-grade GPU server designed for large-scale AI training, supporting 8 high-end GPUs per node and scalable to 4-64 node clusters via InfiniBand. This is the core computing unit for building AI factories and HPC clusters.
Software & Clustering Capabilities
Pre-loaded with Ubuntu 22.04 LTS (HPC/AI optimized), NVIDIA Driver/CUDA/cuDNN/NCCL, Docker/Enroot/Singularity container support, Slurm/Kubernetes cluster orchestration (optional), PyTorch/TensorFlow/JAX multi-node support.
Configured with 8 H100 GPUs: training speed is 20x faster than single-GPU workstations, suitable for large models. Designed for scaling to multi-node via InfiniBand, ideal cluster size 4-64 nodes. Each node can process tens of billions of parameter models (using model parallelism).
LLM Fine-tuning
Enterprise custom training for Llama 3, GPT-like models and other large language models.
Deep Learning Research
Large-scale research projects and cutting-edge algorithm exploration.
Enterprise AI Factory
Production training platform requiring maximum throughput.