GPU to 72 status...

[kracker]

Originally Posted by davieddy

...
How much has the GPU firepower increased since then?
Should keeping ahead to 73 or 74 bits really be such a struggle?

David

It certainly will be easier by this one thing: Get a few GPU yourself, eh?

[chalsall]

Originally Posted by davieddy

After all, you named your effort at coordination "GPUto72" in acknowledgement.

Actually, I named it that because it rhymed....

[Aramis Wyler]

Looking at the year graph, the total throughput on gpu72 has increased to about 177% of what it was last year. Keeping ahead of the curve at bitlevel 74 is ~400% of the work of keeping ahead of it at 72.

Now those are hard numbers, and 'comfortably' is not, so it's hard to say that we should or should not be able to do gpu to 74 because we could comfortably do it to 72.

With our current avg throughput of approx 15k ghzdays per day (not counting p-1 because the time is the same at 72 or 74) and taking an exponent from 69 to 74 being about 56.5 ghz days, then we could probably chuck out about 265 a day if we were all doing that. But we have a massive backlog of work getting all the dc tf done, and I don't think 265 is 'well above' the number of LLs primenet is handing out every day.

Me, I want rid of the dc tf. If we could get 60 days ahead of the wave in the ll tf front, I'd love to spend 30 days working the dc tf. Then of course we'd have to switch back to the ll tf till we were back up to 60. But we don't work as a block and that's ok. I factor to 74. Most people factor to 73. A bunch of people just do DCs or do non-wave related work (LMH).

It's getting done. This time next year it will be a non-issue.

[chalsall]

Originally Posted by Aramis Wyler

With our current avg throughput of approx 15k ghzdays per day (not counting p-1 because the time is the same at 72 or 74) and taking an exponent from 69 to 74 being about 56.5 ghz days, then we could probably chuck out about 265 a day if we were all doing that.

Thanks for your immediate above Aramis.

So everyone knows, over the past 30 days ~284.27 LL candidates were LLed per day. Over the last week, this rate increased to ~305.57 candidates per day. My reading of this is the M58(?) announcement has resulted in a greater number of GIMPS participants.

Over the same 30 days, GPU72 has averaged ~600 candidates TFed to at least 73 bits. Over the same week, this rate has dropped to ~500.

Three important things to keep in mind:

1. There was an expected surge of assignments following the announcement.

2. Historically, only approximately 20 percent of all LL assignments actually complete.

3. Completion rates and assignment rates are decoupled.

While I expect a great many candidates to start to "expire" and be recycled by Primenet in a couple of weeks or so, until then we still have to keep ahead of the assignment wave-front. And even in the last week there were days where Primenet handed out more candidates for LLing than GPU72 completed to at least 73 and released.

In the worst case scenario, we are only ~14 days ahead -- going to 73.

In, lets say, three weeks or so we can come back and decide if we have the firepower to go to 74 before releasing any candidates. Until then, those who want to go to 74 (like you Aramis -- thank you for that as well) are free to do so.

[davieddy]

Originally Posted by Aramis Wyler

Looking at the year graph, the total throughput on gpu72 has increased to about 177% of what it was last year. Keeping ahead of the curve at bitlevel 74 is ~400% of the work of keeping ahead of it at 72.

Now those are hard numbers, and 'comfortably' is not, so it's hard to say that we should or should not be able to do gpu to 74 because we could comfortably do it to 72.

By "comfortably" I mean two things:

1) TF say ~4M in front of LL, allowing for fluctuation in the rates of both,
and plenty of time to do P-1.

2) Keeping ahead of the LLwave leaves enough firepower for folk to prat about in other regions of the spectrum, as many persist in doing.

Be careful when quoting "hard numbers" - they can be easily refuted.
Here's mine: the TF bitlevel should increment when the exponent increases by a factor of 2^1/3 = 26%.

David.

[ixfd64]

Originally Posted by chalsall

My reading of this is the M58(?) announcement has resulted in a greater number of GIMPS participants.

[smartass]I must have missed all of the discoveries since M#48, then...[/smartass]

[chalsall]

Originally Posted by davieddy

Here's mine: the TF bitlevel should increment when the exponent increases by a factor of 2^1/3 = 26%.

David... I must ask...

If you are so authoritative, why didn't you create GPU72?

Rather than complain about almost everything we choose (and chose) to do?

[davieddy]

Originally Posted by kracker

It certainly will be easier by this one thing: Get a few GPU yourself, eh?

If that would make a significant difference, it simply goes to prove my point.

Although it natural to be impressed by big iron,
mocking folk for lack of it should be a no-no.

Get a brain.

[chalsall]

Originally Posted by ixfd64

[smartass]I must have missed all of the discoveries since M#48, then...[/smartass]

Damn... I mis-typed M58(?). I meant M48(?)

Those who live by the sword die by the sword.

But we have a lot of fun, and get a great deal of work done, in the meantime....

[Aramis Wyler]

Originally Posted by davieddy

Be careful when quoting "hard numbers" - they can be easily refuted.
Here's mine: the TF bitlevel should increment when the exponent increases by a factor of 2^1/3 = 26%.

I don't know you, daveiddy, but I fear for your well being. Do you cover the clocks in your house lest you learn the time? Maybe red lights and green lights are just the man trying to hold you down, I don't know.

All I did was read a chart, look in my backlog at the ghz hours taken to move a number from 69 to 74 (which I did many times in the 65m range recently) and do some 3rd grade math on how many times the one divides into the other. I challenge you to easily refute any of them.

You returned with a comment about how you felt about bitlevels. Thrilling. You might even be right, but nothing about how you feel refutes anything I said. I simply read the dial on the clock, and did some basic math so you and others wouldn't have to.

[bcp19]

Originally Posted by Aramis Wyler

I don't know you, daveiddy, but I fear for your well being. Do you cover the clocks in your house lest you learn the time? Maybe red lights and green lights are just the man trying to hold you down, I don't know.

All I did was read a chart, look in my backlog at the ghz hours taken to move a number from 69 to 74 (which I did many times in the 65m range recently) and do some 3rd grade math on how many times the one divides into the other. I challenge you to easily refute any of them.

You returned with a comment about how you felt about bitlevels. Thrilling. You might even be right, but nothing about how you feel refutes anything I said. I simply read the dial on the clock, and did some basic math so you and others wouldn't have to.

For the most part, you should ignore davieddy. He has harped on for well over a year now about the work done by those of us who choose to use GPUs to try and keep ahead of the wavefront. Somewhere in his brain though is a cross-circuit between us being useful and his not being able to be snide about our minor problems. Simple logic shows we(GPU72) have done wonders in making sure the majority of exponents are TF'd to a point where people can expect that most reasonable factors have been found, but we are faulted by davieddy every time a single exponent gets by us, no matter what the reason.

As my dad often says: It's a thankless job (especially where davieddy is concerned) but someone has to do it.