What is a GPU?

What is AI?

HPC Basics Series

Graphical Processing Units (GPUs) are special processors that are optimized for highly parallelized calculations.  In my opinion, GPU is really a misnomer, as the GPU chips are really just optimized for vector processing.  Vector processing has been around for a very long time and optimized for very large-scale matrix calculations and manipulations.  Certain classes of problems lend themselves to acceleration through vectorization, for example, Eigenvalue and Eigenvector problems, as well as other matrix decomposition problems. What this means is that any problem that can be reduced to a series of matrix or vector operations can be accelerated with a GPU.  

We are seeing GPU usage in Artificial Intelligence (AI), specifically in the area of model training for deep learning and neural networks as these training exercises can be described as matrix operations.  The acceleration seen with GPUs over CPU-only applications can be on the order of a 10x increase in speed or more.  Put in perspective, jobs that took overnight to run can now be done over lunch.  When dealing with large amounts of data, keeping the GPUs filled becomes an issue and interestingly, network speed and latency begin to limit performance – not the processors.  At Nimbix, we suggest that when doing large scale jobs that require GPU resources, that a high iOPS data source be used to maximize the GPU performance.

GPUs function as co-processors in that specific jobs are passed to them for processing by the CPU.  A programmer instructs a set of processes to be performed on the GPU by invoking an API call in the program.  What makes this a bit tricky is that every GPU maker has their own API, however, CUDA from NVIDIA, seems to be becoming the industry standard.

GPUs, are becoming part of the computing landscape and software is beginning to take advantage of this type of specialized hardware.  The order of magnitude of accelerations seen is essential to coping with the tremendous amount of data that is being generated and leveraged in fields such as AI, image processing, video processing, bio-medical computing, data mining, and analytics.  A million data points are no longer an anomaly but the rule- billions of data points are becoming commonplace.  The only way to cope with this volume of data is through hardware acceleration.

Here at Nimbix, our environment is complete with the newest GPUs from Nvidia including:

  • V100
  • P100
  • K80
  • and others models

These state of the art GPUs help complete your job faster reducing your compute time.  Get started on the Nimbix Cloud today.