heterogeneous computing

**Heterogeneous Computing** — using multiple types of processors (CPU, GPU, FPGA, custom accelerators) within a single system, assigning each workload to the processor best suited for it. **Why Heterogeneous?** - No single processor is optimal for all workloads - CPU: Great for sequential, branch-heavy code. Latency-optimized - GPU: Great for massively parallel, data-parallel work. Throughput-optimized - FPGA: Great for custom dataflow, low-latency, bit-manipulation - Custom ASIC: Maximum efficiency for specific fixed algorithms **Common Heterogeneous Architectures** - **CPU + GPU**: Most common. Used in AI training/inference, HPC, graphics - **CPU + FPGA**: Network processing (SmartNICs), low-latency trading, genomics - **CPU + AI Accelerator**: Google TPU, Apple Neural Engine, Intel Gaudi - **SoC**: Mobile chips integrate CPU + GPU + NPU + ISP + DSP (Apple M-series, Qualcomm Snapdragon) **Programming Models** - **CUDA**: NVIDIA GPU programming (dominant for AI/HPC) - **OpenCL**: Cross-vendor GPU/FPGA/CPU programming (portable but less optimized) - **SYCL/oneAPI**: Intel's cross-architecture programming model - **ROCm/HIP**: AMD GPU programming (CUDA-compatible API) - **Vitis/Vivado HLS**: FPGA programming with C++ synthesis **Challenges** - Data movement: Transferring data between CPU and accelerator is expensive - Programming complexity: Different programming models for each device - Load balancing: Partitioning work optimally across different processors - Portability: Code written for one accelerator may not run on another **Heterogeneous computing** defines the future of computing — as Moore's Law slows, specialized accelerators are the primary path to continued performance improvement.