The NVIDIA A30 is a mainstream Ampere architecture datacenter GPU designed for mixed AI and HPC workloads, announced in April 2021. It provides a cost-effective path to Ampere datacenter features with a smaller memory footprint than the A100, making it suitable for inference and moderate-scale training.
The A30 uses the GA100 die (same as A100) with 24GB of HBM2 memory providing 933 GB/s bandwidth. It includes 3,584 CUDA cores and 224 Tensor Cores - approximately half the A100's compute resources - in a power-efficient 165W TDP package that fits in a dual-slot PCIe form factor.
A key feature is Multi-Instance GPU (MIG) support, allowing the A30 to be partitioned into up to 4 isolated GPU instances. This enables efficient sharing of GPU resources for inference workloads, with each instance having dedicated memory, cache, and compute resources with hardware-level isolation.
The A30 includes third-generation Tensor Cores supporting TF32, BF16, FP16, and INT8 precision for AI workloads. It also supports FP64 Tensor Core operations for HPC applications requiring double-precision computing, making it versatile for scientific computing deployments.