The SiFive Intelligence™ Family is a highly scalable, integrated hardware-plus-software AI platform designed to accelerate edge-to-cloud AI inference, large language models (LLMs), and modern machine learning workloads. Built on the open RISC-V standard with full RVA23 compliance, the platform provides an agile alternative to proprietary AI architectures. It spans from the ultra-low-power X100 Series for far-edge IoT, to the high-throughput X200/X300 vector processors, up to the XM Series Gen 2, which integrates scalar, vector, and matrix engines into a unified compute subsystem delivering massive TOPS with leading energy efficiency.
The SiFive Intelligence platform introduces groundbreaking architectural innovations:
Unified Matrix & Vector Synergy: The XM Series Gen 2 natively fuses a powerful Matrix Math Engine with wide Vector Units (up to 1024-bit) and 64-bit CPUs, creating a "Fat Outer Product" design heavily tuned for LLMs and deep learning bottlenecks.
Advanced Co-Processor Interfaces: To act as the ultimate Accelerator Control Unit (ACU), the platform introduces Scalar Coprocessor Interfaces (SSCI) and Vector Coprocessor Interfaces (VCIX). These provide direct-to-core, high-bandwidth register access, enabling seamless, ultra-low-latency communication with custom hardware accelerators.
Custom AI Extensions: Features proprietary single-cycle AI instructions (e.g., exponential acceleration) to eliminate critical NN bottlenecks, alongside hardware support for modern AI data types (INT8, FP16, BF16).
The SiFive AI Software Stack: A complete framework incorporating the open-source SiFive Kernel Library, fully optimized for MLIR, TensorFlow Lite, ONNX, and PyTorch, allowing developers to seamlessly deploy AI models from simulation to production silicon.