7im wrote:Funny thing is the Linux TFLOPS have not gone down, nor has the active client count. The first set of numbers is from November 1, 2013. The second set of numbers was taken last week. Same TFLOPS, active clients have gone up, active cores have gone down. Decide for yourself.
Code: Select all
Native x86 CPUs Cores Total CPUs
14 14 4512 20226 45204
14 14 4586 17426 48451
AFAIK multiple teams has some kind of FAH-Christmas-race, having an extra spike in "normal" 2-core to 12-core computers at the same time some other users shuts-down their BA-computers are a possibility that can explain the small increase in active cpu's but a large decrease in #cores. Since where's a 50-day delay from shutting-down until this has any effect on the stats-page, the numbers for 10.02.2014 was around 22.12.2013 so likely during any kind of Christmas-race. Todays numbers on the other hand means around 31.12.2013 and depending on race-end this can be afterwards. For the last two stats-updates, the numbers was:
Code: Select all
Linux 9 9 2878 10030 46946
Linux 9 9 2875 10019 46943
Well, to me a halving of active cores since 01.11.2013 and a 36% drop in active cpu's is a marked decrease for Linux.
The possible reasons for this drop are:
1: There's a problem with the Google-worksheet so the old numbers are incorrect.
2: The stats for 10.02.2014 was incorrect meaning can't really compare to old numbers.
3: The bug-fix for Amd-GPU's broke the stats for Linux (and from the look of things Windows & Mac).
4: Users are turning-off their FAH cpu-clients.
Seeing how the "Total CPUs" for Linux decreased with 3 between the two latest stats-updates, the stats are atleast partially broken but this doesn't neccessarily means the other numbers also are broken.
For the other stats, Amd has a marked increase after the bug-fix. The two Nvidia-types also has a marked increase, maybe these also has been under-counted, or it's possible all GPU-numbers are basically "pre-fix" + "clients returned anything after the fix", so can continue to increase until hit the 50-day-limit and would afterwards get a large drop for the Nvidia-types (No idea if GPU also uses a 50-day limit or not, previous it was shorter). Amd-numbers was so low before the fix so any drop due to a few clients being counted twice will be minimal.
For CPU-clients on the other hand it looks more like "pre-fix" + "fails to count clients returned anything after the fix" meaning these stats are broken. Or, users are choosing #4 instead, and is turning-off their FAH cpu-clients.