GPU to 72 status...

[nucleon]

Originally Posted by kladner

I started to put something in to an earlier comment regarding a sudden increase in GPU power in one quarter or another. It started to get too convoluted, so I dropped that part. In that context, such an increase would only be significant for us more plebeian workers if it came from someone other than the stratospheric crowd. A big boost from Xyzzy would only crowd nucleon. I realize that the Xyzzy mountain range has developed a long plateau, but the nucleon ski slope continues apace, and relative to the average slope of Xyzzy seems to be pulling ahead. Of course, if someone like The Judger decided to join this distinctly limited field it would be felt even the stratosphere.

As previously mentioned, Jerry Hallett seems to be set to upset more apple carts before he reaches his natural level in the ranks.

My TF output has been greater than Oliver's for some time. Check out the TF scores for the last 365 days on the main website.

As tempting as holding back for a record is, it's a _lot_ of work for me. I have 16 instances to manage. It's also a risk (though very minor) of losing results. I like submitting ASAP so I (and the project) doesn't lose any progress.

-- Craig

[kladner]

Originally Posted by nucleon

My TF output has been greater than Oliver's for some time. Check out the TF scores for the last 365 days on the main website.

As tempting as holding back for a record is, it's a _lot_ of work for me. I have 16 instances to manage. It's also a risk (though very minor) of losing results. I like submitting ASAP so I (and the project) doesn't lose any progress.

-- Craig

I had not checked the relative output levels. But now that you mention it I do remember noticing that Oliver was not at the top of the list anymore.

@Dubslow: I hadn't made the connection between Jerry and flashjh. On setting a one day record, I can see Craig's point, especially in his situation. It would be interesting, though.

[Dubslow]

Hehe. It helped that he signed one of his posts with "Jerry"

Even if it was 5 people (including Xyzzy) that read this thread (not including nucleon) that still has a pretty decent chance to break >200 TFLOPS. I think our goal should be a PFLOP.

According to Today's Numbers, 100 TFLOPS = 50K GD/D -- so a PFLOP = 500K GD/D (and 1K GD/D = 2 TFLOPS). (It's exact enough to make me think that that's how George defines/calculates FLOPS.)

Lets see -- Xyzzy gets 1275 GD/D... so that's... wow. 2.5 TFLOPS. WTH. And nucleon is somewhere around 2.8-2.9 TFLOPS. Just think about that. 2.8 trillion floating point ops per second!

So, you and I together get roughly 1/6th of Xyzzy's throughput... if we get flash and LaurV to participate, I think we might be able to get... maybe 3/4ths of Xyzzy's output? Maybe we could double it. So if we could hit 5 TFLOPS, and reported a week's worth of that all at once... That would add 30 TFLOPS or so to the daily output, at the cost of the previous week. Either we need more people or a longer period.

(Out of curiosity, anybody know how George got 2 TFLOPS = 1K GD/D ? That would imply that 1 GD/D is 2 GFLOPS... sounds about right. How close is that to being an actual FLOP count? Does the -2 in the LL count as its own FLOP? Now I'm just rambling. [I don't know about the FFT, but the convolution is one FLOP per bit, right?])

[chalsall]

Originally Posted by Dubslow

Also, what about that suggestion where everyone drops a bomb all at the same time? That might be felt all the way out to the moon

It might be amusing, but from my perspective all it would do is make the graphs difficult to read. The G72 backend (and I suspect PrimeNet) would not have any problems dealing with several thousand results all at once.

BTW, I've exposed another graph from the Admin area: Completions per Hour over the past Two Days.

[KyleAskine]

I would participate as long as we are supposed to upload on a weekday (MUCH easier for me).

[Dubslow]

Originally Posted by chalsall

BTW, I've exposed another graph from the Admin area: Completions per Hour over the past Two Days.

Could you please elaborate? What's the diamonds represent? Complete in the sense of 72+P-1? What's the difference between the red and blue? Is "Completions per Day" just TF? Assignment completion or bit level completion?

(Thank you of course )

Originally Posted by KyleAskine

I would participate as long as we are supposed to upload on a weekday (MUCH easier for me).

Bang! Zoom! Straight to the moon!
( ^^ Hehe, I made a funny)

[chalsall]

Originally Posted by Dubslow

Could you please elaborate? What's the diamonds represent? Complete in the sense of 72+P-1? What's the difference between the red and blue? Is "Completions per Day" just TF? Assignment completion or bit level completion?

It's just a quick graph I threw together to monitor the Observation Spider, and thought some might enjoy it.

The green bars are GHz Days of work completed per hour. The red line is a (simple) running average of this. Both of these are related to the left-hand scale.

The black diamonds are the number of completions per hour, inclusive of all work types (TFing to any bit depth (including multi-level (for example, 69-72)), and P-1, DC and LL completions). The blue line is the running average of this. Both of these are related to the right-hand scale.

I have to say I find it just a little scary that we're averaging over 300 GHz Days per hour (probably over 400 once Xyzzy submits his results)!!!

[Dubslow]

Okay, thanks.

Originally Posted by chalsall

The black diamonds are the number of completions per hour, inclusive of all work types (TFing to any bit depth (including multi-level (for example, 69-72)), and P-1, DC and LL completions).

So is 69-72 counted as one or three? (I'm guessing three, since that would be a more consistent measure...)

Edit: There appears to not be a link on your site. (This may be deliberate, or it may be I haven't given you enough time I don't mind either way, just pointing it out.)

[chalsall]

Originally Posted by Dubslow

So is 69-72 counted as one or three? (I'm guessing three, since that would be a more consistent measure...)

No, only one. But you're right, it should be three. I'll adjust the SQL to take that into account.

Originally Posted by Dubslow

Edit: There appears to not be a link on your site. (This may be deliberate, or it may be I haven't given you enough time I don't mind either way, just pointing it out.)

There is. It's the "Hourly Graph" under the "Overall System Progress" menu.

[Dubslow]

Originally Posted by chalsall

There is. It's the "Hourly Graph" under the "Overall System Progress" menu.

Oh, hey. Whoops.

What's on the x-axis? OHH that's the time of day UTC? kk.

[Dubslow]

Originally Posted by Dubslow

Hehe. It helped that he signed one of his posts with "Jerry"

Even if it was 5 people (including Xyzzy) that read this thread (not including nucleon) that still has a pretty decent chance to break >200 TFLOPS. I think our goal should be a PFLOP.

According to Today's Numbers, 100 TFLOPS = 50K GD/D -- so a PFLOP = 500K GD/D (and 1K GD/D = 2 TFLOPS). (It's exact enough to make me think that that's how George defines/calculates FLOPS.)

Lets see -- Xyzzy gets 1275 GD/D... so that's... wow. 2.5 TFLOPS. WTH. And nucleon is somewhere around 2.8-2.9 TFLOPS. Just think about that. 2.8 trillion floating point ops per second!

So, you and I together get roughly 1/6th of Xyzzy's throughput... if we get flash and LaurV to participate, I think we might be able to get... maybe 3/4ths of Xyzzy's output? Maybe we could double it. So if we could hit 5 TFLOPS, and reported a week's worth of that all at once... That would add 30 TFLOPS or so to the daily output, at the cost of the previous week. Either we need more people or a longer period.

(Out of curiosity, anybody know how George got 2 TFLOPS = 1K GD/D ? That would imply that 1 GD/D is 2 GFLOPS... sounds about right. How close is that to being an actual FLOP count? Does the -2 in the LL count as its own FLOP? Now I'm just rambling. [I don't know about the FFT, but the convolution is one FLOP per bit, right?])

The days tick by, and yet the bomb hasn't dropped...

Perhaps instead of organizing some other drop, we all just collect our results now and just report when Xyzzy does. Xyzzy, would it be possible for you to give 12 hours or so advance notice, for all those who want to contribute to your massive bomb?