[tonybolger]

Agree. Also, the bottleneck will always be I/O.

Not sure what kind of processing you're doing, but for most sequencing stuff, it would appear that IO shouldn't be the bottleneck, at least until you're into large scale clusters / grids.

If you're disk throughput bound, add spindles. If it's disk IOPS, SSDs make a massive difference (and IOPS-bound data tends to be relatively small). If it's networking, go 10Gbit.

[lh3]

I suspect the 80:20 rule (or more like 99:1 rule) will be the saviour here - you don't need to optimize the performance of all the code, just the key nasty part that takes up the vast majority of the execution time.

For all BWT-based mappers, "the key nasty part" is to count the occurrences in BWT. This may take 80% CPU time. For BWA, another nasty part is to maintain a priority queue used by BFS.

[rskr]

I think the problem is that GPU's don't scale very well, and consume too much electricity. IE most people who do bioinformatics have access to a cluster, and seldom is it cost effective to enable GPU's on a cluster even when the cluster motherboards come with them installed, since the GPU will use much more power than the CPU itself, and for relatively little benefit, and for the most part alignment jobs are embarrassingly parallel, so they scale with well with the number of processors that are added.

GPU's are for kids and noobs, IMO.

[gregg.tehennepe]

BGI is reporting substantial speed ups from their GPU-enabled aligner and SNP detector.

[rskr]

BGI is reporting substantial speed ups from their GPU-enabled aligner and SNP detector.

See there it takes a whole compute farm of GPU's and it still takes six hours, and probably uses more watts too.

[adaptivegenome]

Thanks Nils. I thought there was a standard such as OpenGL or CUDA that would be cross platform/GPU cards so that such code may be portable across GPUs.

@Richard, does this mean than increasing RAM has more realized improvements than optimized/parallelized code?

There is a standard. It is OpenCL and most hardware and operating systems support it. In fact MacOSX is built using OpenCL. While universal, the trouble with OpenCL is that it is not as fast as hardware optimized technologies such as CUDA.

I personally am a fan of OpenCL and I believe even the new processors such as Knights Corner will utilize OpenCL to farm threads to the coprocessor.

[ymc]

icic. Looks like it is not yet the right time to buy a 3GB GTX 580

[colindaven]

Those interested could also have a look at

MummerGPU

http://sourceforge.net/apps/mediawiki/mummergpu/index.php?title=MUMmerGPU

SARUMAN short read mapper

http://www.cebitec.uni-bielefeld.de/brf/saruman/saruman.html

[brianlamx]

You can use a 3GB GTX 580 with BarraCUDA, other than that you just need a standard workstation with 4GB RAM and some disk space.

http://seqbarracuda.sf.net