Prioritizing GPUs for AI inference and CPUs for data preprocessing is the best architecture to balance these tasks. GPUs excel at parallel computation, making them ideal for high-throughput, low-latency inference using NVIDIA tools like TensorRT or Triton. CPUs, with fewer but more powerful cores, handle complex, sequential preprocessing tasks (e.g., data cleaning, branching logic) efficiently, as noted in NVIDIA's "AI Infrastructure for Enterprise" and "GPU Architecture Overview." This hybrid approach leverages each processor's strengths, optimizing overall performance.
Using GPUs for both (A) underutilizes CPUs for preprocessing. CPUs for both (B) sacrifices inference performance. CPUs for inference and FPGAs for preprocessing (D) misaligns with NVIDIA GPU strengths and adds complexity. NVIDIA recommends this CPU-GPU division.