The NVIDIA A10 is an entry-level Ampere architecture datacenter GPU optimized for AI inference and graphics workloads, announced in April 2021. It uses the GA102 die in a single-slot, low-power configuration designed for dense inference server deployments and cloud graphics instances.
The A10 features 24GB of GDDR6 memory with 600 GB/s bandwidth, providing ample capacity for most inference models. The GA102 die includes 9,216 CUDA cores, 288 Tensor Cores, and 72 RT cores for hardware-accelerated ray tracing, sharing its architecture with the consumer RTX 3090 and professional RTX A6000.
A distinguishing feature is the single-slot PCIe form factor with 150W TDP, enabling deployment of up to 4-6 GPUs per server depending on chassis configuration. This density makes the A10 cost-effective for inference at scale where multiple lower-power GPUs are preferred over fewer high-power units.
The A10 supports all Ampere inference optimizations including INT8 and INT4 Tensor Core operations, structural sparsity acceleration, and the Multi-Process Service (MPS) for efficient GPU sharing across multiple inference processes. Primary deployment scenarios include cloud AI inference instances (AWS G5, Azure NC A10 v4).