[Bioclusters] Use of FPGA based BLAST accelerators?

[Chris Dwan (CCGB)]

> Does anyone on the list have any experience with using some of the
> commercially available FPGA accelerators for Bioinformatics applications
> (e.g. from Timelogic)?

We've got a system from TimeLogic with two of their 'G' series
DeCypher boards.  Depending on the application, each board gives about
the same throughput as 100 CPUs in our 1.4Ghz P-III cluster.  HMM
based apps are better than that, pairwise alignment based searches are
slightly worse.  The most common objection from my users is that it's
not plug-compatible  with NCBI's BLAST.  I.e: You have to write
parsers to take advantage of the system.

That's not much of an objection, in my book.

Our usual model is to prototype a processing pipeline on the cluster,
and then migrate the CPU intensive portions to the TimeLogic.  This
combines the flexibility of the cluster with the power of the
specialized machine.  

TimeLogic has made some pretty significant improvements in the past
year or so in terms of software quality and reliability of their
systems.  It was downright painful to work with these devices when we
got them in late 2000.  Today, they're really quite stable.

> Based upon some of the performance claims I've seen, it seems to me that
> augmenting a standard Linux cluster with one of these devices should
> provide pretty compelling price/performance.  Why is this not commonly
> done?

The buy-in price is quite high.  Most academic centers start small and
build up their compute farm over several funding cycles.  You can't do
that when buying one of these devices.   Also (sad as it is), a 200
CPU cluster has way more "blinkin' lights" appeal than a single
machine that quietly outperforms that cluster on some applications.  

-Chris Dwan
 Center for Computational Genomics and Bioinformatics
 University of Minnesota