The Implementation of High- Performance GPU Computing to ... › webassets › cms › library ›...
Transcript of The Implementation of High- Performance GPU Computing to ... › webassets › cms › library ›...
GOALThe use of the multiplexed data acquisition
strategy, LC/MSE provides the most comprehen-
sive and unbiased method of analyzing complex
biological systems. However, it is computation-
ally intensive and places a heavy demand on
informatics post-processing. Here we describe a
novel solution to significantly speed up associ-
ated peak detection and data processing times.
BAckGrOundLC/MSE is a key method of data acquisition that
utilizes parallel sequencing of all detectable
peptides as they elute from the UPLC® Column
with the use of high mass resolution and accuracy,
which helps maximize coverage and confidence in
results. When analyzing and utilizing this data, it
is imperative to uniquely identify each ion in both
the low and elevated energy functions, and then to
correlate the eluting peptides with their associated
fragments, using retention time alignment. This
provides highly specific and sensitive information
for database searching and quantitative profiling.
The processing of these large-scale datasets
requires significant computational power and is
often a bottleneck in determining the constituent
peptides and proteins. In this technical brief, we
describe the implementation of peak detection
and data processing algorithms on a Graphics
Processing Unit (GPU) that resulted in significant
time savings for processing LC/MSE data.
The Implementation of High-Performance GPU Computing to Revolutionize LC/MSE Data Processing
Implementation of Lc/MSE on a GPu reduces computation time and speeds up overall processing time by a minimum of 5x.
0
200
400
600
800
1000
1200
2003 2004 2005 2006 2007 2008 2009 2010
GPU
CPU
Figure 1. Historical comparison of the computation power difference between the CPU and GPU in GFLOPS.
The GPU processes all graphics and images shown in a computer display. In this
sense, it is a dedicated processor, whereas the CPU (Central Processing Unit) is
a general processor that handles all other tasks in a computer. In recent years,
the GPU’s computation power has increased at a higher rate than the CPU, mostly
driven by the need for higher quality graphics in today’s consumer and professional
applications. Because of its greater computation power, the GPU is now being used
for tasks other than graphics. For example, in the last three years, it has been used
Waters Corporation 34 Maple Street Milford, MA 01757 U.S.A. T: 1 508 478 2000 F: 1 508 872 1990 www.waters.com
Waters is a registered trademark of Waters Corporation. The Science of What’s Possible is a treademark of Waters Corporation. All other trademarks are the property of their respective owners.
©2010 Waters Corporation. Produced in the U.S.A.May 2010 720003488EN AO-PDF
for computationally intensive applications in various
scientific and engineering disciplines.
Unlike the CPU, which is a series processor that can
only process one program instruction at a time, the
GPU is a parallel processor that can process many
program instructions in parallel. Another major
difference is the number of cores that can be placed in
a chip. Today, the CPU packs up to six cores in a chip,
whereas the GPU packs up to 448.
The GPU achieves significant reductions in computa-
tion time in compute-intense applications that can
be parallelized, such as applications that do the same
repetitive computation on different data.
ThE SOLuTIOnIn LC/MSE data processing, peak detection from the raw
data is performed by Apex 3D, an algorithm which has
been previously described in detail1. By running this
algorithm on the GPU, we obtained an 50 x increase
in speed in the computational sections of the program
for a typical data size. With this reduced computation
time, the data access sections of the program became
dominant in the overall processing time, which resulted
in a net speedup of approx 5x.
This effect is less important as data become larger and
more complex, as shown in Figures 2 and 3. In these
cases, the time in data access is less relevant and the
net speedup is closer to the computational speedup.
SuMMAryHere we have described the implementation of LC/MSE
data algorithms on a GPU that significantly enhances
the speed of data processing. This will be crucial in
processing multi-dimensional LC-IMS-MSE data in the
near future.
Reference
0 10 20 30 40 50 60 70 80
60 min
90 min
120 min
180 min
1D UPLC _ MS
6.1x
5.8x
5.5x
5.1x
time (minutes)ac
quisi
tion
time
E
GPU
CPU
Figure 2. Improvement in performance using a 240-core GPU compared with a quad-core CPU for processing LC/MSE data files of varying file size from different chromatographic separations.
0 50 100 150 200 250 300
3 fraction
5 fraction
10 fraction
2D UPLC MS E
GPU
CPU
8.1x
6.9x
6.2x
time (minutes)
acqu
isitio
n tim
e
_
Figure 3. Comparison of data processing times for LC/MSE data files from 3-, 5-, and 10-step 2D-LC experiments using a 240-core GPU compared with a quad-core CPU.