GR wus getting longer?
20 Mar 2021 2:22:25 UTC
Topic 225083
(moderation:
I haven't been running E@H much until recently. I spent some time dialing in my setups and thought I had things running pretty consistently but now I'm seeing like 20-40% longer run times on multiple different hosts/gpus. Do they generally vary that much?
The crunch times have now
)
The crunch times have now gone back to normal with the latest LATeah4000 series data file.
The previous Lateah3000 series were unusually fast. They lasted for perhaps close to 2 months but the current times are more typical for tasks delivered for several years prior to that. The RAC bonus party is over - as was pointed out recently :-).
Cheers,
Gary.
4000 series are still about
)
4000 series are still about 10% faster than the historical average speed for GR from my observations. but yes 3000 series were the "fast" ones, and also had slightly different run behavior (less GPU utilization, more memory use)
my RTX2070's on the "old" style GRs always ran them in about 10minutes flat. now it's about 9 mins with 4000 series. 3000 series were running them in about 6:40
_________________________________________________________________________
Gary Roberts wrote: The
)
I had no idea credit was better for a while, I guess I switched to einstein at a good time. is GR still the best credit? Or which app relies most on double precision, as that's the only thing keeping these 7970s going.
MindCrime wrote: Gary
)
neither app relies very heavily on DP. only a small portion of the computation is DP from what I recall, the bulk is SP.
if you're interested in DP, you should look into Milkyway@home. their application relies almost entirely on DP.
_________________________________________________________________________
The new set of GR tasks is
)
The new set of GR tasks is short ones again.