There are alternatives to FPGAs that provide similar speedup and scalability. If you can utilize your existing infrastructure, but increase performance by several orders of magnititude (extending Moore's Law) with software, that may provide the best price/performance ratio and lowest impact of MTBF. Better yet, if you just have one or two commonly used apps like BLAST and HMMer that are causing you problems on your existing hardware, why not outsource to a service that plugs into your existing workflow? Just a thought. K -----Original Message----- From: Joe Landman [mailto:landman at scalableinformatics.com] Sent: Saturday, February 11, 2006 6:42 AM To: Clustering, compute farming & distributed computing in life science informatics Subject: Re: [Bioclusters] FPGA in bioinformatics clusters (again?) Acceleration systems make sense in specific situations. They are rarely general purpose. Most FPGA implementations are non-trivial to program. FPGA's are not the only form of accelerator, but they have received the most attention in recent months. That said, when they are applicable, and you have intensive processing needs, there is nothing quite like a custom CPU or a custom processing circuit to speed through your calculation. Getting 20-100x the performance is possible and reasonable as an expectation, per CPU/FPGA. If your application needs this, then there is little that is comparible. You can put multiple boards into multiple elements in your cluster, and provided your application is parallelizable at the process level (blast, hmmer, much of bioinformatics, ...) you ought to be able to realize multiple orders of magnitude acceleration without paying huge sums of money. Basically, if you are building a cluster for speed, then likely you need something like this for at least one of your applications. If you don't need speed, and you just need lots of apps, there are solutions to that, and accelerators won't help much there. If you run a few critical apps which are performance bottlenecked, and the bottleneck are CPU cycles (and not data motion), then accelerators will likely help. Accelerators make lots of sense in analysis pipelines, where specific calculations may bottleneck the entire workflow. They can be built/sold for reasonable cost in a number of cases, and the performance in these cases can be excellent. George Magklaras wrote: > The Linux Journal issue 142 (February 2006) talks about FPGA's in an > article with title 'Heterogeneous Processing: a Strategy for > Augmenting Moore's Law', written by a chap from Cray. Apart from the > ehmm indirect > XD1 product marketing, the article makes the case for FPGA's outlining > alternative approaches to traditional commodity HPC clusters, as well > as the obstacles of turning scalar proc code to FPGA code. > > Best Regards, > GM > -- Joseph Landman, Ph.D Founder and CEO Scalable Informatics LLC, email: landman at scalableinformatics.com web : http://www.scalableinformatics.com phone: +1 734 786 8423 fax : +1 734 786 8452 cell : +1 734 612 4615 _______________________________________________ Bioclusters maillist - Bioclusters at bioinformatics.org https://bioinformatics.org/mailman/listinfo/bioclusters