Bigadv points change

[P5-133XL]

What is needed is an explanation from the researchers confirming the exponential time value of returns, and this thread would be ended.

"We set up the QRB system with a reasonable plan for how the science connects to time. To first order, that's a good estimate of the value of science vs time."

Obviously not, since this thread did not end or perhaps Dr. Pande is not considered a researcher and that's why.

[bruce]

Obviously not, since this thread did not end or perhaps Dr. Pande is not considered a researcher and that's why.

As you point out, this topic has now reached the point that people are repeating things that have already been said and arguing over trivia rather than solving the problem. It's either time to lock the topic or focus the discussion more closely on constructive ideas.

[mdk777]

Obviously not, since this thread did not end or perhaps Dr. Pande is not considered a researcher and that's why.

I was just hoping that a more detailed explanation would help remove any lingering doubt.

[7im]

Read that FAQ I linked, and then tell me how any doubt about form fitting the points curve to more than 8 cores can be called "lingering"

Substantial is a word I would use.

[ElectricVehicle]

Each league has a separate ranking, though they largely share the same points system. The minor and major league sports share most of the same rules and scoring, but have independent rankings.

I think you forgot about those of us in The Geezer League

I always knew you were EVIL Bruce! There is NO WAY I want compete with forum member Grandpa_01 - he's going to wipe me all over the shuffleboard!!!

Seriously more power to him and never underestimate the Geezers! FOLD ON!

[Grandpa_01]

Each league has a separate ranking, though they largely share the same points system. The minor and major league sports share most of the same rules and scoring, but have independent rankings.

I think you forgot about those of us in The Geezer League

I always knew you were EVIL Bruce! There is NO WAY I want compete with forum member Grandpa_01 - he's going to wipe me all over the shuffleboard!!!

Seriously more power to him and never underestimate the Geezers! FOLD ON!

[mdk777]

Read that FAQ I linked, and then tell me how any doubt about form fitting the points curve to more than 8 cores can be called "lingering"

Reading this FAQ I don't see much discussion of the QRB, or advbig or beta big advbig at all.

I think your background causes you to read in much more information into this FAQ than is really there.
You are correct, it doesn't talk about 48 core machines, but it really doesn't discuss the time value of the QRB either.

Other than saying it is important and will continue.

Bonuses have played a key role in aligning points with science and we will continue to use them. For example, returning work units (WUs) promptly can be very important for the science we’re doing, so we provide bonuses for this, especially with the high performance clients.

[MtM]

Read that FAQ I linked, and then tell me how any doubt about form fitting the points curve to more than 8 cores can be called "lingering"

Reading this FAQ I don't see much discussion of the QRB, or advbig or beta big advbig at all.

I think your background causes you to read in much more information into this FAQ than is really there.
You are correct, it doesn't talk about 48 core machines, but it really doesn't discuss the time value of the QRB either.

Other than saying it is important and will continue.

Bonuses have played a key role in aligning points with science and we will continue to use them. For example, returning work units (WUs) promptly can be very important for the science we’re doing, so we provide bonuses for this, especially with the high performance clients.

7im is the same person who been saying faq's need to be brief and concise, stating a faq not covering something being equal to meaning it didn't consider it therefore seems odd

All we seen so far is VP saying: this is as close as it gets.

[7im]

Let's put some numbers in to the current forumula according to PG's example (the FAQ the mdk777 linked).

Q:How are you awarding bonuses?
A: Bonuses are awarded according to the following formula:
Total points = base points * bonus factor

The bonus factor is computed based on the time from when our server issues the work unit to when it receives the work unit (WU_time), the time from work issuance to when the deadline would expire (deadline_time), the time from issuance to when the work unit times out and is marked for reissuance (timeout_time), and a constant factor k.
If WU_time > timeout time, bonus factor = 1.
If WU_time <= timeout time, bonus factor = sqrt(deadline_time * k / WU_time)

IMPORTANT:
Bonuses are only given *if* a client has a passkey and *after* a client has completed 10 Core A2 work units. Also, to qualify for a bonus, a client must have returned >80% of its work units within the deadline. Otherwise bonus factor = 1.

Example:
For project 2681, we will initially set k=2. We may adjust k as necessary. Again for project 2681, the current deadline time is 6 days and the current timeout time is 4 days. Most users' 8-core machines, clocked at 2.8 GHz or higher, complete these work units in slightly under 3 days, so they would receive a 100% bonus.

Q: I'd like to try running these units to burn in our new ultra-secret 80-core chip at work - with permission, of course. Is there any upper limit to the bonus points you will award?
A: At current time, the maximum bonus factor is 10x (completion of a 2681 work unit in 1/50th the deadline = approx 2 hrs 50 min). If you're in a position to exceed that, talk to us.

Total points = base points * bonus factor (where the bonus factor is: sqrt(deadline_time * k / WU_time))

Total points = base points * sqrt(deadline_time * k / WU_time)

Using the example above...

Total points = 8955 * sqrt(6 * 3 / 3)

Total points = 8955 * 2

Total points = 17,910 (--> 3 days to complete)
PPD is 5,970

Total points = 26,865 (--> 2 days to complete)
PPD is 13,432

Total points = 31,021 (--> 1.5 days to complete)
PPD is 20,680

Total points = 37,992 (--> 1 days to complete)
PPD is 37,992

Total points = 53,730 (--> .5 days to complete) (12 hours)
PPD is 107,460

Total points = 75,985 (--> .25 days to complete) (6 hours)
PPD is 303,942

Total points = 107,460 (--> .125 days to complete) (3 hours) (their previous 10x cap was here... so lifting the cap hasn't applied to anyone just yet. Looks more like a 100x cap to me. )
PPD is 859,680

[Amaruk]

Let's put some numbers in to the current forumula according to PG's example (the FAQ the mdk777 linked).

Q:How are you awarding bonuses?
A: Bonuses are awarded according to the following formula:
Total points = base points * bonus factor

The bonus factor is computed based on the time from when our server issues the work unit to when it receives the work unit (WU_time), the time from work issuance to when the deadline would expire (deadline_time), the time from issuance to when the work unit times out and is marked for reissuance (timeout_time), and a constant factor k.
If WU_time > timeout time, bonus factor = 1.
If WU_time <= timeout time, bonus factor = sqrt(deadline_time * k / WU_time)

IMPORTANT:
Bonuses are only given *if* a client has a passkey and *after* a client has completed 10 Core A2 work units. Also, to qualify for a bonus, a client must have returned >80% of its work units within the deadline. Otherwise bonus factor = 1.

Example:
For project 2681, we will initially set k=2. We may adjust k as necessary. Again for project 2681, the current deadline time is 6 days and the current timeout time is 4 days. Most users' 8-core machines, clocked at 2.8 GHz or higher, complete these work units in slightly under 3 days, so they would receive a 100% bonus.

Q: I'd like to try running these units to burn in our new ultra-secret 80-core chip at work - with permission, of course. Is there any upper limit to the bonus points you will award?
A: At current time, the maximum bonus factor is 10x (completion of a 2681 work unit in 1/50th the deadline = approx 2 hrs 50 min). If you're in a position to exceed that, talk to us.

Total points = base points * bonus factor (where the bonus factor is: sqrt(deadline_time * k / WU_time))

Total points = base points * sqrt(deadline_time * k / WU_time)

Using the example above...

Total points = 8955 * sqrt(6 * 3 / 3)

Total points = 8955 * 2

Total points = 17,910 (--> 3 days to complete)
PPD is 5,970

Total points = 26,865 (--> 2 days to complete)
PPD is 13,432

Total points = 31,021 (--> 1.5 days to complete)
PPD is 20,680

Total points = 37,992 (--> 1 days to complete)
PPD is 37,992

Total points = 53,730 (--> .5 days to complete) (12 hours)
PPD is 107,460

Total points = 75,985 (--> .25 days to complete) (6 hours)
PPD is 303,942

Total points = 107,460 (--> .125 days to complete) (3 hours) (their previous 10x cap was here... so lifting the cap hasn't applied to anyone just yet.)
PPD is 859,680

AFAIK the 2681s are extinct, as is that formula.

Current formula is:

final_points = base_points * max(1,sqrt(k*deadline_length/elapsed_time))

SOURCE

8,955 point WUs have k factor of 26.4, thus:



3 days = 65,070.26 points, 21,690.09 PPD

2 days = 79,694.47 points, 39,847.23 PPD

1.5 days = 92,023.25 points, 61,348.83 PPD (note: 10X cap already applies here)

1 day = 112,705 points, 112,705 PPD

0.5 day = 159,388.94 points, 318,777.88 PPD

0.25 days = 225,410 points, 901640 PPD

0.125 days = 318,777.88 points, 2,550,223.01 PPD

The TPF at which a 10X cap takes affect is about 00:22:45. A 980X @4 GHz runs at @00:21:09 and would be capped.

[7im]

It's the exact same formula. The "Max(1," added to the front means you will never get less than 1x the base points, as long as the WU is completed before the final deadline.

Both of those are assumed to be true when completing the WU in 3 days or less on a WU with a 6 day deadline. So I left those out, to keep things simpler.

But thanks for the updated WU data, and PPD data. The much larger numbers tend to help my argument.

When you go from 4 days to 2 days, 2x as fast, the PPD goes from 11,502 to 32,535, or 2.8x the PPD.

When you go from 2 days to 1 day, 2x as fast, the PPD goes from 32,535 to 92,023, or 2.8x the PPD, but 8x the first jump.

When you go from 1 days to .5 days, 2x as fast, the PPD goes from 92,023 to 260,281, or 2.8x the PPD, but 23x the first jump.

When you go from .5 days to .25 days, 2x as fast, the PPD goes from 260,281 to 736,185, or 2.8x the PPD, but 64x the first jump.



Power is doubling, 2x, 4x, 8x, 16x, 32x, 64x, while points are tripling, 3, 9, 27, 81, 243, 729.

Moore's law is going to continue doubling power every 2 years, while the cost of that consumer system stays about the same. So with the QRB and tripling the points, that's the points inflation people are talking about. Points will escalate higher and higher while systems increase in speed at the same relative rate.

Why triple the points when the power is only doubled?

ChasR has some good insights. Raise the Base points. Surely that data section of the WU should be worth more than the bonus. Lower the K values (this comes naturally with larger base points). Lowering K decreases the slope of the curve when you get down to very short return times.

And use the 3rd root instead of the 2nd root!


32 core people are getting 250,000 PPD today, while most people have 8 cores, and they are getting 20,000 PPD.

Let's move the clock ahead a few years so that most people have 32 cores, and the top end people has 128 cores. That's 250,000 PPD versus 6,000,000 PPD.

And ahead again. 6,000,000 vs. 92,000,000 Can you imagine making a hundred million points per day?

Why triple the points when the power is only doubled?

[mdk777]

Total points = 107,460 (--> .125 days to complete) (3 hours) (their previous 10x cap was here... so lifting the cap hasn't applied to anyone just yet. Looks more like a 100x cap to me. )

I can't believe what I am reading.

You are confabulating ppd with the bonus, and not taking in to consideration that more WU are being done

In the first example, 1 WU is done in 3 days. 17,910 points or 5970 ppd.

In the 3 hour example the points go up to 107,460 but in one day 8 WU get turned in

IF THERE WERE NO INCREASE IN BONUS AFTER 3 days, this would generate 8x17910=143280 ppd

So the bonus from going from 3 days to 3 hours is the division of these numbers 859680/143280= 6X

YES, 6x bonus multiplication for going from 3 days to 3 hours.

THIS ONLY LOOKS LIKE A 100 X bonus, because you discount the insane 24x increase of work done.

In 3 days, one WU returned, or in the same time 24 WU returned. Which helps the project and science?


I am just amazed, completely dumbfounded.

[7im]

You're playing number games too, but I left out the extreme adjectives when I gave my examples.

Whatever dude. At least I can tell the difference between four hundred 8 core sytems and what you called a single massively cored 3200 core system.

[GreyWhiskers]

If the points really are capped, then the following table gives the REAL results of the calculations. Of course, the sub-day ppd is high just because of producing more than one of these puppies per day, as mdk77 pointed out.

Code: Select all

                                    points      k         deadline
                                      8955      26.4          6

                                      days  Raw pts  Capped pts capped ppd
3 days = 65,070.26 points, 21,690.       3  65,070.3   65,070.3   21,690.1

2 days = 79,694.47 points, 39,847.       2  79,694.5   79,694.5   39,847.2

1.5 days = 92,023.25 points, 61,34     1.5  92,023.2   89,550.0   59,700.0

1 day = 112,705 points, 112,705 PP       1 112,705.0   89,550.0   89,550.0

0.5 day = 159,388.94 points, 318,7     0.5 159,388.9   89,550.0  179,100.0

0.25 days = 225,410 points, 901640    0.25 225,410.0   89,550.0  358,200.0

0.125 days = 318,777.88 points, 2,   0.125 318,777.9   89,550.0  716,400.0

[7im]

The bonus factor was capped at 10x, not the bonus points.

And don't trying to mixing my original numbers, when the K value was 2, with mdk777's numbers, then the K value was 26.4.