AW-A80
Training + Inference
The AW-A80 is a dual-purpose AI server optimized for both training and inference. With 4 GPUs and 512GB RAM, it handles distributed training of medium-scale models while simultaneously serving inference workloads. The 8TB NVMe storage provides ample space for large training datasets and model checkpoints. Liquid cooling keeps temperatures stable during intensive training runs that can last days or weeks.
Technical Specifications
Detailed hardware configuration of AW-A80
Performance Benchmarks
Real-world performance data from our test lab
Key Features
What makes AW-A80 the right choice
Quad NVIDIA V100/A100 GPUs for training + inference
512GB DDR4 ECC RAM — enough for large model training
8TB NVMe SSD for training datasets and checkpoints
Simultaneous training and inference on separate GPU partitions
Liquid cooling for sustained 7×24 training loads
ALWork with training scheduler and experiment tracking
Ideal For
Teams that need a single server for both model fine-tuning and production inference, research labs, and organizations with moderate training requirements.
Frequently Asked Questions
Common questions about AW-A80
Can I train and infer simultaneously?
Yes. ALWork allows GPU partitioning — dedicate 2 GPUs to training and 2 GPUs to inference, or dynamically reallocate based on workload priority. The liquid cooling ensures thermal stability regardless of the workload mix.
What is the maximum model size for fine-tuning?
With 4× V100 32GB GPUs (128GB total VRAM) and 512GB system RAM, the AW-A80 can fine-tune models up to 13B parameters using QLoRA, or up to 7B with full fine-tuning using data parallelism.