IBM announces system-on-chip AI hardware (venturebeat.com)

IBM has unveiled its Artificial Intelligence Unit (AIU), a purpose-built system-on-chip AI accelerator designed to enhance AI workloads in data centers and enterprise servers. Unlike CPUs or GPUs, the AIU is an application-specific integrated circuit (ASIC) specifically architected for AI tasks, delivering higher efficiency and performance. The AIU features 32 cores fabricated using a cutting-edge 5nm process—smaller and more advanced than the 7nm cores used in IBM’s Tellum chip found in their z16 mainframes. It connects to servers via a standard PCIe interface, enabling flexible deployment across IBM’s x86 and Power server lines, as well as partner-built infrastructure. A standout aspect of IBM’s AIU is its built-in virtualization capability, allowing multiple AI workloads to securely share the accelerator in multi-tenant cloud and enterprise environments. This contrasts with traditional GPUs, which historically lacked robust virtualization support and required complex retrofitting. The AIU also supports widespread AI software frameworks like PyTorch and TensorFlow, easing integration for data scientists. Furthermore, IBM integrates approximate computing techniques, which strategically trade off minimal precision to boost speed and efficiency—leveraging the inherent tolerance for small errors in AI inference without significantly impacting accuracy. This innovation exemplifies IBM’s decade-long commitment to creating specialized, high-performance AI hardware designed from the ground up to meet the unique demands of modern AI and deep learning applications.