GPU VS CPU

This section compares the complementary architecture of GPU and CPU. Because GPUs and CPUs are orthogonally optimized with each other, their combination into a heterogeneous GPGPU system provides cost and performance advantages for certain applications, compared to a pure CPU approach.

Basic Differences between CPU and GPU Architecture

GPU and CPU are designed and optimized for two very different types of applications, the architecture is significantly different. This can be seen by comparing the relative number of print areas (number of transistors) dedicated to cache, control logic, and processing logic for the two types of processor technology. In CPU control logic and memory cache constitute the majority of CPU real estate. This is as expected for an architecture that is set to process sequential code as fast as possible. On the GPU side it uses a massive parallel SIMD (one instruction multiple data) architecture to perform mathematical operations primarily. Thus, the GPU does not require the same complex capabilities of CPU control logic (e.g. faulty) execution, branch prediction, data hazard, etc.). Also does not require a large amount of cache memory.

Performance and Performance Comparison per Watt

The video game market has driven the need for real-time graphic realism which continues to increase. This translates to a more parallel GPU processor core with greater floating-point capabilities. As a result, the GPU is designed to maximize the number of floating-point operations per second (FLOPs) it can perform. In addition, newer NVIDIA architectures, such as Kepler and Maxwell, have focused on increasing the performance per watt ratio (FLOPs / watt) than before GPU architecture by reducing the power needed by each GPU core processor. This was achieved by Kepler by reducing its processor core clock, while increasing the number of on-chip transistors (following Moore's Law) allows for a positive net gain of 3x the performance per watt compared to the Fermi architecture. Other than that, Maxwell's architecture has increased execution efficiency. This trend increases the FLOP that GPUs can do versus multicore CPUs that deviate exponentially, thus creating a large performance gap. Similar can be said about the performance per watt gap between these two different processing technologies.