[ Home | Articles | Account | People | Projects | FAQ ] got a new video card, a nvidia 8600 gt. the overclocked model from club3d (it's in sweden i think). it's not really to get a better world of warcraft from linux (works like a charm by the way, on debian etch 4.0 32 bits + cedega 6.0.5). i had a 7900 gt previously and could not try to play with Cuda. at work (cnrs) we got a cluster. and its admin bought two tesla cards from nvidia which are dedicated to mathematics (128 threads par cpu, two per card). the good thing about it and compared to the ati solution (this is gonna change very soon i think under's amd guidance) is we use nvcc which are wrappers (not very clean if you dive your nose in) around gcc. it's moving fast. first, each card offers us 500 Gflops so we have a theorical 1 Tflop with just two cards loaded on two machines. it's not very portable and you really need to know the architecture behind it to have it work properly. the first week we got the cards was funny. once the card was plugged the machine was not even booting. it only worked on a very very few specific dell machines. after a nvidia developer gave us some daily code, it finally worked. well the machine was booting but once the nvidia's cuda driver was installed it did not see the card. it took over two weeks to have the cards available from debian on the cluster ;) and of course all pre-packaged stuff is only for redhat, suse.. well. i am a bsd monk so i really dont like linux. but god i do love debian (i never thought i would ever say that). the nice thing about the whole thing is you buy a cheap video card, and you can from linux write cuda code. then, we just move that code to the cluster where the tesla cards run it with their raw computing power. that's cool to be able to compile and run it on small scales and then just load it on the cluster for full range calculations. the only thing that really annoys me if that most Cuda functions are UPPERCASE like this and i got a happy hacking keyboard without capslock key... :/
well tonight i'm going to install and try at home for the first time to use cuda on my 8600 and i'd like to implement something easy like a md5 hash or so. and then upload it to the cluster but it's not really something you can get scalar execution from (well, it's a stream you work upon so it's gonna be disk-bound in performance).
another problem are mathematicians. most of them if not all know how to use matlab or libraries to their fields of work (petsc, and so on) and for cuda to really work it has to be as transparent as possible. most know how to write C code, but we're not really seeing them wanna learn the inners of those tesla cards to get real fast, optimized scalar code to run on those gpus. we need either fortran libraries that use as much cuda as possible or to have as much as possible matlab functions to be redirected to the gpu through some addon/lib for matlab. most of our mathematicians do code, but their work is mathematics. not spending days the nose deep in C with the tesla architeture in mind for best performance. hope nvidia will improve this on those parts :)
FOAF updates: Trust rankings are now exported, making the data available to other users and websites. An external FOAF URI has been added, allowing users to link to an additional FOAF file.
Keep up with the latest Advogato features by reading the Advogato status blog.
If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!