FPGAs for Financial Services

1
FPGAs for Financial Services

• Date 18 September 2006
• Produced by Chris Swan

This is a technology presentation, and does
necessarily reflect any projects at Credit Suisse
The materials may not be used or relied upon in
any way.
2
Agenda

• The challenge data centre growth
• FPGA a tempting solution?
• The challenges ahead, and why we may need to
  overcome them
• Questions

3
The challenge data centre growth

• HPC workloads presently account for around 20 of
  a typical financial services data centre
• These workloads are also encroaching into the
  client space
• Growth of existing business x per annum over
  Moores law
• Growth of new business x per annum over Moores
  law
• Compounded growth means that we need to build new
  data centres just to house the grids
• Cost transparency isnt modulating demand
• We havent yet reached the perfect risk
  management balance point

4
FPGA a tempting solution?

• By encoding straight to dedicated hardware rather
  than running on general purpose CPUs it is
  possible to get three orders of magnitude better
  performance.
• Applies more to performance/watt than
  performance/chip
• This could potentially collapse a 1000 node grid
  into an expansion card within a traders
  workstation
• Chips can be rapidly repurposed (new/different
  code)
• VLSI industry has made these parts commodity
  items, so no exotic hardware costs

5
FPGA the challenges

• Very low level coding model
• Tool support
• This isnt what quants learn at school
• Only so much can be optimised in hardware
• Need to have a clear understanding of algorithms
• Re-factoring, or indeed totally different
  approaches may be required
• Need to accomplish clean separation
• Need to manage the boundary conditions
• Interfaces, skills, support

6
Why we might be forced to overcome the issues

• Products that cant be managed without the extra
  horsepower
• There are non computational optimisations in many
  models
• These provide arbitrage opportunities if more
  complex models are used
• Trading that opportunity is computationally
  expensive for the buy side
• Sell side may be forced into an arms race

7
Questions?
8
(No Transcript)