suppose humans have no computer but get a book with instruction how to calculate the move that lc0 play at 1 node per move.
How much time do you think humans need to spend and how many calculation they need to calculate.
Can a team of humans perform the same task faster and how much faster(I guess that generally team of x humans can do it faster but less than x times faster)
time to calculate 1 node of lc0 for humans question
Moderator: Ras
-
- Posts: 10903
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
-
- Posts: 28395
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: time to calculate 1 node of lc0 for humans question
Running the NN requires multiplication, so 99.9% of humans could not do it at all.
LC0 nets come in different sizes, but what I vaguely recall from AlphaZero is that the NN was 40 layers of 8x8 'cells', each cell containing 256 neurons, each connected to each of the 256 neurons in the 3x3 cell area of the previous layer. So that would be 40x8x8x256x256x9 = 1.44G weights. Suppose the human could do 1 multiplication per second, then it would take him 45 years (if they never sleep). But I suppose the time to multiply two numbers under 256 would take most people more like a minute, so it would be more like 6000 years.
Fortunately the task is highly parallelizable; you could put tens of thousands of people on it with virtually no loss of efficiency.
This is just a rough order-of-magnitude estimate; I could be off by a fator 10.

LC0 nets come in different sizes, but what I vaguely recall from AlphaZero is that the NN was 40 layers of 8x8 'cells', each cell containing 256 neurons, each connected to each of the 256 neurons in the 3x3 cell area of the previous layer. So that would be 40x8x8x256x256x9 = 1.44G weights. Suppose the human could do 1 multiplication per second, then it would take him 45 years (if they never sleep). But I suppose the time to multiply two numbers under 256 would take most people more like a minute, so it would be more like 6000 years.
Fortunately the task is highly parallelizable; you could put tens of thousands of people on it with virtually no loss of efficiency.
This is just a rough order-of-magnitude estimate; I could be off by a fator 10.
-
- Posts: 12534
- Joined: Thu Mar 09, 2006 12:57 am
- Location: Birmingham UK
- Full name: Graham Laight
Re: time to calculate 1 node of lc0 for humans question
What have I missed here? You said the network has 40 layers: aren't the calculations from one layer needed for the calculations in the next layer?
Other important factors if humans were doing the work:
1. Error checking and correction
2. Fatigue: the humans calculating these meaningless numbers would get bored out of their tiny minds! Maybe you'd periodically kill some to motivate the others?
What technology would the humans be using to assist them in their work?
1. Smart speaker - a 2010s technology
2. Electronic calculator - a 1960s technology
3. Paper - a 2000 year old technology (modern paper - older types exist)
4. Abacus - a 2500 year old technology
5. They have to do all calculations in their heads
Human chess is partly about tactics and strategy, but mostly about memory
-
- Posts: 28395
- Joined: Fri Mar 10, 2006 10:06 am
- Location: Amsterdam
- Full name: H G Muller
Re: time to calculate 1 node of lc0 for humans question
Indeed, they are. But since a single layer requires 36M multiplications and additions, and is independent of anything other than the output of the previous layer, the work can be easily divided. Each layer has 8x8x256 outputs, so 16K humans could simply each calculate a single output. Which requires 9x256 multiplications and additions. If they are all done they can start with the next layer.
-
- Posts: 10903
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: time to calculate 1 node of lc0 for humans question
I believe that memorization of all multiplications below 256 is a task that some humans can do with enough training and even if it takes a full year to do it then after this full year they can do 1 multiplication per second and I guess the main time is going to be used to follow the instruction what to multiply.hgm wrote: ↑Fri Apr 12, 2024 11:50 am Running the NN requires multiplication, so 99.9% of humans could not do it at all.![]()
LC0 nets come in different sizes, but what I vaguely recall from AlphaZero is that the NN was 40 layers of 8x8 'cells', each cell containing 256 neurons, each connected to each of the 256 neurons in the 3x3 cell area of the previous layer. So that would be 40x8x8x256x256x9 = 1.44G weights. Suppose the human could do 1 multiplication per second, then it would take him 45 years (if they never sleep). But I suppose the time to multiply two numbers under 256 would take most people more like a minute, so it would be more like 6000 years.
Fortunately the task is highly parallelizable; you could put tens of thousands of people on it with virtually no loss of efficiency.
This is just a rough order-of-magnitude estimate; I could be off by a fator 10.
-
- Posts: 12534
- Joined: Thu Mar 09, 2006 12:57 am
- Location: Birmingham UK
- Full name: Graham Laight
Re: time to calculate 1 node of lc0 for humans question
towforce wrote: ↑Fri Apr 12, 2024 3:26 pmWhat technology would the humans be using to assist them in their work?
1. Smart speaker - a 2010s technology
2. Electronic calculator - a 1960s technology
3. Paper - a 2000 year old technology (modern paper - older types exist)
4. Abacus - a 2500 year old technology
5. They have to do all calculations in their heads
Quick amendment to the list - I'm adding slide rules (because NN's don't require 8-digit accuracy) and pinwheel calculators (the best type of calculator prior to electronic ones).
1. Smart speaker - a 2010s technology
2. Electronic calculator - a 1960s technology
3. Pinwheel calculator - an 1800s technology
4. Slide rule - a 1600s technology
5. Paper - a 2000 year old technology (modern paper - older types exist)
6. Abacus - a 2500 year old technology
7. They have to do all calculations in their heads
Based on the previous post in this thread, option 7 seems to be Uri's preferred option.
Human chess is partly about tactics and strategy, but mostly about memory