[Bioclusters] FPGA in bioinformatics clusters (again?)

[Christopher Dwan]

It's excellent to see this discussion pop up again.  Kudos to Joe  
(Larry?) for keeping things close to the road.

The key thing about FPGAs from my perspective is that FPGA boards are  
not general purpose compute engines.  It's not a one to one  
substitution from FPGAs to compute nodes.  They are a component that  
can accelerate a particular sort of computation.  They no more solve  
the general purpose computing problem than "faster disk drives"  
solves data bottlenecks.

Therefore, the question is almost never "should I get an FPGA OR a  
cluster."  It's usually "Would a couple of FPGA boards make my  
cluster more useful?"

The core question is "is my cluster spending most of its time  
crunching on some algorithm for which an accelerator board exists?"   
If so, FPGA systems may be a decent investment.  TimeLogic's  
marketing material, back when I last saw it, focused on the fact that  
the rest of your  cluster would be free do do other stuff, if only  
you were able to move all the BLAST, HMM, and Smith Waterman searches  
off to a special purpose resource.

There is an ancillary question:  "Am I willing to make my computing  
system slightly more complex, and possibly dedicate resources to  
maintaining and exploiting this complexity?"  Depending on how much  
local resource a group is willing to dedicate to their custom  
systems, lots of interesting tools become quite appealing.

My opinion on truly "general purpose" FPGAs is that they're not well  
suited for *direct* purchase by most life sciences cluster users.   
This is because *most* of the bio-cluster people aren't all that  
interested in spending money developing or optimizing custom codes.   
Most of the folks I work with are not that interested in profiling,  
re-compiling, optimizing, tuning, etc.  Some certainly are, and many  
of those people read this list.

So, there is certainly an opportunity for companies to make  
optimized / accelerated versions of some standard applications and re- 
sell them, which brings us back to where the thread began.

-Chris Dwan