As you may know, our gpu server is called gupiter, voted by the research members in our last formal meeting.
I just installed the gputools package in R (with a lots of difficulties). Now you can compute on the the Tesla K40 from R. It is enough to load the gputools library and use the cuda wrapped functions.
Moreover, our gpuiter now can give you access to RStudio through RStudioServer. Since we have only two seats over the gupiter, I thought it is much better if I provide access to the other team members. Now you can remotely can have access (with graphical support) through RStudioServer.
Sajjad already made a post on comparison of GPU with CPU in MATLAB over gupiter. I just run a toy example. Let's generate a 1024X1024 random matrix and compute the Euclidean distance over the matrix. Seems for this example gpu is about 30 times faster! I am sure we can get much more efficiency over larger data.
library(gputools)
set.seed(5446)
X <- matrix(rnorm(2^20),ncol = 2^(10))
cputime <- system.time(d <- dist(X))
gputime <- system.time(gpud <- gpuDist(X))
cputime/gputime[1]## user system elapsed
## 31.11688 0.00000 31.12338
Great! so from now on, the competition has started to see which framework (programming language, or API) could beat the other ones. Moreover, it got me wanna implement my algorithms on GPU directly through the Cuda interface :))
ReplyDeleteSuggestion; it could be wonderful to open a contest with an award to implement one algorithm on GPU with any tool (GPUtools, Matlab, Cuda, ...) to see whose algorithm is the most efficient one, then in the formal/informal meeting give the prize :D
ReplyDeleteThe purpose of the contest is to compare the performance of each library/algorithm to make the best out of this powerful GPU server.
Parfait! Juste une petite remarque à propos de son nom. Puisqu’on a décidé de remplacer la première lettre de ‘Jupiter’ par G afin que son nom soit plus proche de GPU, alors il se considère comme un nom propre. Donc, il vaut mieux écrire Gupiter au lieu de gupiter.
ReplyDelete