Breakout thoughts compiled with N. Carter:

1
Breakout thoughts (compiled with N. Carter)

• Where will RAMP be in 3-5 Years (What is RAMP,
  where is it going?)
• Is it still RAMP if it is mapping onto something
  other than FPGA?
• Can we revert back to software simulation?
• Is RAMP cost effective with current Parallel HW
  (different 5 years ago).
• Now possible to build a 1K system with blades
• ex Niagara has 64 threads - could simulate 1K
  cores with 16 blades?
• Probably can get a 50 CPU cluster for "industry
  cost" of BEE3, 50K.
• Maybe 1K per core.

2
Breakout thoughts

• On Clusters
• For thoughput limited runs (lots of independent
  simulation runs) - ex generating 3000 data
  points
• then a cluster with many CPUs each working
  separately
• For low latency single run - network latency will
  dominate and result in low emulation performance.
• In all of these software approaches
• There is a big issue of the mismatch between
  host and target. For good matches (ISA,
  network, memory system) can be efficient.
• FPGA because of fine grain configurability
  largely avoids this issue.
• But are these processor based emulations still
  "RAMP" systems?

3
Breakout thoughts

• Perhaps what makes it RAMP is the design
  representation not the host.
• There may be something to using RAMP methodology
  and models for software (and other)
  implementations
• As it is too easy to cheat in software only
  models - starting from a hardware model (with
  explicit interconnect, etc.) will help
• Can Re"host" the emulation later to FPGAs.
• Has extra benefits of
• avoid the bother or cost of FPGA HW initially,
  and
• Help in verification by having two
  implementations
• Simple Scalar has been successful party because
  it runs on anything.

4
Breakout thoughts

• In order to pull this off need a higher level of
  abstraction for specifying units
• 1. Library approach (users want parameterized
  high-level blocks)
• 2. modeling language (builders of blocks want a
  common language for both HW/SW implementations)
• Can C-gates help?
• 3-5 years RAMP will need to transition into
  "tools
• Right now project is still investigating the
  space of bag of tricks
• At some point must transition to being much
  easier to use
• (a tool not a PhD thesis)
• Liberty as an example for RDL! (but has a steep
  learning curve!)