GPU Architectures.pdf

GPU Architectures

1. GPU Background: GPUs are designed for drawing triangles fast, with basic processing including projecting triangles into 2D, finding pixels, determining color, and calculating lighting.
2. GPU Design: GPUs process pixels in parallel, focusing on throughput rather than latency, with many simple parallel processors and low clock speed.
3. CPU vs. GPU Philosophy: CPUs prioritize low latency and high clock speed, while GPUs prioritize high throughput and parallel processing.

GPU Trends

1. More of the same: Increasing number of processors and memory controllers.
2. More CPU features: Adding caches, function pointers, and double precision.
3. More flexible: Supporting multiple kernels and programming models.

GPU Memory Philosophy

1. Computational Intensity: Prioritizing math operations over memory access.
2. No cache: GPUs do not rely on caches to reduce memory latency.
3. Solution: Give Up: GPUs optimize for high throughput rather than low latency.

CPU Memory Philosophy

1. Big caches: CPUs use large caches to reduce memory latency.
2. Instruction window: CPUs use out-of-order and multiple-issue instructions to hide memory latency.
3. Approach: CPUs reduce memory latencies with caches and hide memory latencies with other instructions.

Key Differences

1. Latency vs. Throughput: CPUs prioritize low latency, while GPUs prioritize high throughput.
2. Cache hierarchy: CPUs have a complex cache hierarchy, while GPUs have a simpler memory architecture.
3. Parallel processing: GPUs are designed for massive parallel processing, while CPUs are designed for serial processing.