GPU Computing
Created on: 2025-03-04
Tags: #GPU Computing
Course: HPC
Definition
- Computing using GPUs or Graphical User Interface
- GPU Server = CPU Server + GPU (Connected through the PCIe slots).
- Peripheral Component Interconnect Express (PCIe) Slots
- General Purpose, high speed interface standard
- Use to connect various components like GPUs, storage, network cards, etc.
- Universal Standard and Compatible with range of devices.
- High latency due to general purpose design
- NVLink
- High-speed interconnect technology developed by Nvidia.
- Provides connectivity for GPU to GPU and GPU to CPU in their GPU architecture.
- It was designed to overcome the issues related to PCIe.
- Higher bandwidth, lower latency and more power efficient.
- Removes the requirement of CPU for communication between GPUs
Evolution of GPU
- Initially designed to work in combination with a CPU as accelerators and render graphics.
- Developed to produce more realistic graphics rendering
- Later programmability of the GPUs were introduced.
- Now, serve and support various diverse fields such as gaming, ML, etc.
GPGPU v/s GPU
- GPGPU - General Purpose GPU can be used for purposes beyond graphical processing such as performing computations typically conducted by a CPU.
- GPU - A special-purpose device that supplements the main-purpose CPU to speed up certain operations.
Types of GPU
- Integrated GPU: A graphics processor engine which is contained on the CPU. It uses CPU memory.
- Dedicated GPU: A GPU on a separate peripheral card. It has dedicated memory.
Comparison between CPU v/s GPU
| CPU | GPU | |
|---|---|---|
| Clock Speed | High | Low |
| Cores and Threads | Few but faster | Many but slower |
| Function | Handles main processing task | Specialized for parallel computing |
| Processing | Serial instruction | Parallel Instruction |
| Suited for | General purpose computing | High-performance computing |
GPU FLOPs Calculation
- If the instructions are executed on FADD or FMUL then OPS/thread is 1 so FLOPs/cycle/SM is 64.
- If the instructions are executed on FMA then OPS/thread is 2 so FLOPs/cycle/SM is 128.
- The maximum FLOPs require all FP 32 instructions to be FMA. If only FADD or FMUL are executed then the code can at best achieve 50% of the maximum FLOPs. \(FLOPS_{Max} = \text{SMCount} \times \text{CUDACores/SM} \times 2 \text{FLOPS/CUDACore} \times \text{GPU Frequency}\)
Linked from
- No other published notes link here yet.