Both rigs are still cooking, with 9 WUs complete on the new 3.7 gig baseline. Right now the 920 is winning by about 2.8 min/WU, but we'll see, 15 each to go yet.
Emu had an even bigger delta in his test. I wonder if these WUs are different from my first batch somehow? This is indeed getting more interesting....
We may need some more "clock-for-clock", "setting-for-setting" data before we go off on any other tangents? Let's see how my new baseline comes in...
We're trying to make this test useful, without a lot of side tracks, particularly since they cost Emu and I about 12 to 16 hours of dead production for each run, depending on clock......Check my thinking here..... The initial goal of the thread was to verify that the 860 and by extension, it's Lynnfield brothers, were up to what the socket 1366 chips had to offer in terms of crunch power. A second goal was to show which socket had the best power efficiency for the same production.
This is to offer choices to our teammates to buy with. Some of you are already going the Lynnfield route. DAK and INFRNL have joined us. (BTW, I'm still drooling over an X3440 rig....I just might have to have one...
) Based on our earlier testing, I don't think you will be at all dissapointed. My 860 cranks out the points. It's the only way RSMura is not running me over while my other rigs are gone....
I would ask you, the audience, are our goals as they were above? Anyone want to see something different? I took the short side track to get a feel for how the exact same parameters, on the same rig vary over different runs. While this showed some variance with only one trial, it is hardly statistically valid, but it costs production time.... This test method is new to all of us. Perhaps more runs like that are of value?
The other factor to remember is that Emu and I are running different sets of test WUs. The only thing each of us can do is run a 920 against an 860. Any value to us getting together and figuring out who sends test WUs to whom? Remember the problem with this is in the various account keys. Those go with the data. I'd be willing to send Emu mine, it's just a lot of data, about 130meg.
I've come to trust Emu. Just don't let those WUs escape into the wild and bring down the planet's mainframes.... The issue here is that Emu's have more run time since I managed to corrupt my first ones....
One thing I mentioned before that may be of use is to refine jcool's formula for PPD estimating. I think we could do that with this test method, but we would need to decide on what we are looking for. We all know straight bclk makes a good difference. We have some indication that QPI makes a pretty good difference too, Uncore/NB makes some difference, and RAM timings may make a yet-to-be-determined difference. This is with HFCC WUs though....
What I will say is that I'm not willing to burn an infinite amount of CPU time here. I'm about producing valid, useful WUs. That's why I farm. On the other hand, I'm willing to maximize my production too. Where is the balance point is what I'm asking the team?
Bob
Bookmarks