ARGOMENTO:

Brunone ha scritto:

Per noi poveri utilizzatori di Mac questa notizia è decisamente gradita dato che non possiamo utilizzare le nostre GPU (ATI o NVIDIA che siano)

Odio insistere, ma... perchè? Cosa cambierebbe?

forse quello che ho scritto è leggermente fuorviante
sta portando "fuori strada" quasi tutti

Ducati 749 ha scritto:

in pratica tutti i crediti verranno normalizzati:
* la stessa applicazione su CPU o GPU darà gli stessi crediti
* progetti diversi sullo stesso pc daranno gli stessi crediti
* pc diversi sullo stesso progetto avranno gli stessi crediti

a quanto ho capito i "tassi di conversione" saranno gestiti a livello centrale da boinc e aggiornati periodicamente

in ogni caso direi che l'articolo va tradotto e pubblicato

In effetti oggettivamente quei 3 punti sono... sbagliati

gabi.2437 ha scritto:

In effetti oggettivamente quei 3 punti sono... sbagliati

eh addirittura! come sei drastico


different projects should grant about the same amount of credit per CPU hour, averaged over hosts. Projects with GPU apps should grant credit in proportion to the efficiency of the apps.
If a given application has multiple versions (e.g., CPU and GPU versions) the granted credit per job is adjusted so that the average is the same for each version.
if an app has only GPU versions, then for each version V we let S(V) be the average scaling factor for that GPU type among projects that do have both CPU and GPU versions. This factor is obtained from a central BOINC server.

Il metodo che proporrei io è quello usato da Folding

Si prende il progetto, si prende un hardware e si dice "questo hardware fa 1500 RAC se acceso 24/24"

E si danno i crediti basandosi su queste prestazioni, se il mio hardware macina il doppio, avrò 3000

Interessante questo messaggio di Gipsel

Proposed new BOINC credit system:

boinc.berkeley.edu/trac/wiki/CreditNew

Comments? Please, well thought out interpretations of how this is going to affect things here...

The very first sentences for the motivation start already from the wrong premises:

Peak FLOPS and efficiency
GPUs typically have a much higher (50-100X) peak FLOPS than CPUs. However, application efficiency is typically lower (very roughly, 10% for GPUs, 50% for CPUs).

That is completely bogus. It entirely depends on the problem and can also be easily the other way. In fact, MW runs with a lot higher efficiency on GPUs than on CPUs. The reason is the ridicilous amount of register space on GPUs (a HD5870 has 5 [five] MegaBytes of registers plus 640kB of shared memory, much more than cache). That allows the massive multithreading happening on GPUs (the GPU doesn't have to swap out the registers each time it changes to another thread) which enables to hide all kind of latencies. That is simply not possible on CPUs so far (only in a rudimentary way with hyperthreading).

Just to throw out some numbers, on MW a typical GPU reaches about 85% of efficiency (which is not exactly the ratio of peak flops, that's only ~70% , nvidia GPUs maybe even slightly more (the efficiency is not the problem, peak performance is). This compares to about 50% on a Core i7 (hyperthreading saves the day) and just about 35% on a typical CPU (all running the vectorized SSE3 version). The good thing with GPUs is that they offer you an effective instruction latency of just a single cycle for all operations when dealing with heavily compute bound parallel problems.