GPU Concepts / Architectures

Kevin Stratford

kevin@epcc.ed.ac.uk

Material by: Alan Gray, Kevin Stratford

EPCC Logo

Outline

  • Factors influencing performance
  • How do CPU / GPU architectures differ?
  • Some trends in architecture
  • Programming scientific applications?

Factors influencing performance

Schematic diagram represeting data transfer between
                  processor and memory. The processor may have one or
                  more levels of cachme memory.
  1. Clock speed: processing rate per compute element
  2. Memory latency: time to retreive a data element
  3. Memory bandwidth: amount of data transferred in unit time
  4. Parallelism: replicating the basic units of computation

CPUs: Clock speed

  • Power consumed by a device $\propto fV^2$
  • Historically, devices got faster by increasing clock frequency $f$
    • Voltage $V$ decreased to keep power managable
  • Now, voltage cannot be decreased easily
    • Bits are represented as a voltage difference
    • Engineering constraints mean a finite voltage difference must be maintained for correct operation
  • Increased clock frequency no longer available
Clock frequency in MHz against year of manufacture

See http://cpudb.stanford.edu/visualize/clock_frequency

CPUs: Memory latency

  • Memory latency is a serious concern in design
    • Can cost $O(100)$ clock cycles to retrieve data from main memory
    • Steps often taken in hardware to reduce latencies
  • Cache hierarchies
    • Large on-chip caches to stage data
    • L1, L2, L3, ..., controllers, coherency mechanisms, ...
  • Other latency-hiding measures
    • Hardware multithreading ("hyperthreading")
    • Out-of-order execution
E.g., Intel Sandybridge die
A typical picture of a CPU design
                (here Intel Sandybridge) showing the amount of
                real estate set aside for cache memory and memory control

Image: http://www.theregister.co.uk/ (Source: Intel)

CPUs: Memory bandwidth

  • CPUs generally use commodity DDR RAM
    • "Double Data Rate"
    • Manufacturers often don't publish bandwidth data
    • A standard benchmark might give $O(100)$ GB / s
  • In practice, memory bandwidth can be important
    • Many real applications bandwidth limited

CPUs: Parallelism

  • Source of increased performance now to use more cores
  • Almost all commodity processors are now "many-core" or "multi-core"
  • Limited clock speed keeps power consumption per core managable
E.g., Intel Knights Landing die (72 cores)
Picture of Intel Knights Landing die showing a relatively
                large area devoted to 72 cores

Image: http://www.hardwareluxx.de/ (Source: Intel)

CPUs: Summary

  • CPUs: a highly successful design for general purpose computing
    • Used for operating systems, data bases, input/output, ...
    • Inevitable design compromises...
  • ...mean not specifically intended for HPC
    • Much functionality, e.g., branch prediction, may not be required
    • Hardware / power devoted to infrequent (HPC) operations
  • Huge commercial market
    • Huge investment involved in fabrication of new chips
    • Bespoke HPC chips not economically viable

... from the dismal science...

GPUs

  • Large lucrative market in another area: games
    • Require Graphical Processing Units
    • Two dominant manufacturers: AMD and NVIDIA
  • GPUs are designed to do rendering
    • An embarrassingly parallel problem
    • Favours a balance between floating point/memory bandwidth
  • How is design reflected in factors influencing performance?

GPUs: Clock speed

  • Underlying clock frequency relatively modest
    • Perhaps in high 100s MHz
  • Performance based on parallelism...

GPUs: Memory latency

  • Problem has not gone away
    • Strategy to hide it again related to parallelism
  • Schedule very many independent tasks
    • More than can be accommodated on hardware at one time
  • If one task encounters a long-latency operation
    • ... swapped out and schedule another task
    • Done in hardware, so fast
    • Multithreading "writ large"

GPUs: Memory Bandwidth

Graphs comparing the performance of x86 and NVIDIA harware.
                On the left is Peak floating point operations per second
                against year of manufacture, and on the right is peak memory
                bandwidth against year of manufactor.
                The Figure is from NVIDIA

Source: NVIDIA

Parallelism: Basic unit is streaming multiprocessor (SM)
NVIDIA Streaming Multiprocssor. There are 64 cores per SM
                 in this particular device (Pascal).

Source: https://devblogs.nvidia.com/parallelforall/inside-pascal/

Many SMs form a graphics processing cluster (GPC)

Source: https://devblogs.nvidia.com/parallelforall/inside-pascal/

Trends in Architecture

  • Certain convergence in architectures
    • CPUs / GPUs will probably become more integrated
    • Memory space may become more uniform
  • No doubt in need for parallelism
    • More cores / threads

Programming

  • Graphics processing languages were / are used
    • DirectX, OpenGL
    • One or two early "heroic efforts" in scientific applications
  • In 2007 NVIDIA developed CUDA
    • Compute Unified Device Architecture
    • Primarily to make graphics programming easier
    • As a by-product, scientific applications become more tractable

Programming

  • At the same time, OpenCL was developed
    • Important in mobile phones, games
    • Not so much traction in scientific applications
  • Directives-based approaches are available
    • Standards situation was relatively slow to become clear
    • Relatively pain-free
    • Sources of poor performance can be obscured

Big Data / Machine Learning

An XKCD cartoon lampooning 'AI'

Source: https://xkcd.com/1897/

Summary

  • GPUs offer the opportunity of cost and energy efficient computing
  • The holy grail for programming:
    performance, portability, productivity