AW-A40
Multi-Model Serving
The AW-A40 is designed for multi-model AI serving. With 4 NVIDIA GPUs, it can run 4 different models simultaneously — one per GPU — or combine GPUs for larger model parallelism. The 256GB RAM and 4TB NVMe provide balanced capacity for model caching and data processing. Liquid cooling ensures thermal stability during continuous multi-GPU operation. Perfect for AI service providers who need to offer multiple model endpoints from a single server.
Technical Specifications
Detailed hardware configuration of AW-A40
Performance Benchmarks
Real-world performance data from our test lab
Key Features
What makes AW-A40 the right choice
Quad NVIDIA V100/A100 GPUs for multi-model serving
Run 4 different models simultaneously — one per GPU
GPU pooling for large model parallelism (up to 128GB VRAM)
3000W redundant PSU with liquid cooling
ALWork multi-model scheduler with per-GPU load monitoring
4U chassis with tool-less drive bays for easy maintenance
Ideal For
AI service providers running multiple model endpoints, teams needing both inference and training from one server, and workloads requiring GPU pooling for large models.
Frequently Asked Questions
Common questions about AW-A40
Can AW-A40 run training workloads?
Yes, with 4 GPUs the AW-A40 supports distributed training using frameworks like PyTorch DDP and DeepSpeed. It is suitable for fine-tuning medium-sized models (up to 13B parameters with LoRA/QLoRA).
How is GPU pooling configured?
ALWork Model Manager supports both model parallelism (splitting one model across multiple GPUs) and data parallelism (replicating a model across GPUs for higher throughput). Configuration is done through the ALWork dashboard.