AlphaZero beats AlphaGo Zero, Stockfish, and Elmo

[jhellis3]

So what is so new?

Sigh.... I will try to put it terms you can understand.

AlphaZero is not like SF.

AlphaZero evaluated at 80 thousand nodes/sec while SF was at 80 million.

But the eval AlphaZero is using is a self constructed network. So the end result is more like Magnus Carlsen evaluating 80 thousand nodes per second with 0 mistakes. But actually it is worse than that because its "understanding" of the game is even better than Magnus's, it is beyond the human realm, and it is ever improving.

There are no gaps in its understanding for you to repeatedly exploit, and should a very large miracle occur and you were to find one. It would learn, on its own. It does not play chess as humans understand it, it plays chess as it understands it...

[abulmo2]

It actually is, instead of 4TPUs required to run Alpha0 so far, on x64 hardware one would need around 2000 Haswell cores to achieve the same speed of NN

Details about Google's TPU are available here.
https://drive.google.com/file/d/0Bx4haf ... xtcEk/view

For NN inference, 1 TPU is around 15-30 times faster than an Haswell (with multiple cores, so your figure of 2000x faster for 4 TPUs vs 1 core is right), but other comparable properties favor the CPU:
TPU vs CPU
Die size: TPU is < half size of an Haswell
Frequency: 700 Mhz vs 2300
Power: 40W vs 145W
Memory Bandwitdth: 51 vs 34

The TPU is much more efficient than a CPU (for NN inference), but it does not eat more power, occupies more size, etc.
So using 4 TPU vs 64 Haswell cores, is fair I think.

[abulmo2]

With what is this different from a self-tuning software, as widely used in autotuning engines, applied on a very large scale/involving tremendous hardware?

Most of current chess eval can be modelised as a single neuron (or perceptron). AlphaGo uses many neurons. That's a big difference.

[Milos]

It actually is, instead of 4TPUs required to run Alpha0 so far, on x64 hardware one would need around 2000 Haswell cores to achieve the same speed of NN

Details about Google's TPU are available here.
https://drive.google.com/file/d/0Bx4haf ... xtcEk/view

For NN inference, 1 TPU is around 15-30 times faster than an Haswell (with multiple cores, so your figure of 2000x faster for 4 TPUs vs 1 core is right), but other comparable properties favor the CPU:
TPU vs CPU
Die size: TPU is < half size of an Haswell
Frequency: 700 Mhz vs 2300
Power: 40W vs 145W
Memory Bandwitdth: 51 vs 34

The TPU is much more efficient than a CPU (for NN inference), but it does not eat more power, occupies more size, etc.
So using 4 TPU vs 64 Haswell cores, is fair I think.

As fair as using SF running on a single CPU Haswell and 100 Ultrascale+ FPGA chips used only for evaluation. Sorry, but there is no f from fair in this comparison. If we miraculously got Alpha0 source code for x64, we would need around 2000 Haswell cores for the similar performance, or at least 100 top of the range 1080Ti GPUs.

[Lyudmil Tsvetkov]

It actually is, instead of 4TPUs required to run Alpha0 so far, on x64 hardware one would need around 2000 Haswell cores to achieve the same speed of NN

Details about Google's TPU are available here.
https://drive.google.com/file/d/0Bx4haf ... xtcEk/view

For NN inference, 1 TPU is around 15-30 times faster than an Haswell (with multiple cores, so your figure of 2000x faster for 4 TPUs vs 1 core is right), but other comparable properties favor the CPU:
TPU vs CPU
Die size: TPU is < half size of an Haswell
Frequency: 700 Mhz vs 2300
Power: 40W vs 145W
Memory Bandwitdth: 51 vs 34

The TPU is much more efficient than a CPU (for NN inference), but it does not eat more power, occupies more size, etc.
So using 4 TPU vs 64 Haswell cores, is fair I think.

Apart from frequency, how would the other parameters affect speed?
Is not frequency already factored in the '20 times faster assessment'?

2000/1, with added SMP inefficiencies, will make it 3000/1, 3000/64,
in what way could that be fair?

[jhellis3]

Or 4 V100s which will be available in 2018, or you can order in a workstation now for 70k.

You seem to think GPU progress and process node shrinks are just going to stop. V100 is on 12nm and 7nm chips will have mass availability in 2018.

It is over. Deal with it. Or not.

[Lion]

I agree with you.
Also what people who claim the HW was much faster..... what they don’t understand is that the thing learned from itself in a very short time!

What if we now give it 1 Year to further learn?


Side note, I looked at the games and they are really impressive!

[Lyudmil Tsvetkov]

With what is this different from a self-tuning software, as widely used in autotuning engines, applied on a very large scale/involving tremendous hardware?

Most of current chess eval can be modelised as a single neuron (or perceptron). AlphaGo uses many neurons. That's a big difference.

Many neurons, but of what quality?
It should be very bad, in case with 1000 times more neurons than SF, it will achieve 400 elo lower strength on single core.

[jhellis3]

How many Elo do you have with 1 brain cell?

[Lyudmil Tsvetkov]

Or 4 V100s which will be available in 2018, or you can order in a workstation now for 70k.

You seem to think GPU progress and process node shrinks are just going to stop. V100 is on 12nm and 7nm chips will have mass availability in 2018.

It is over. Deal with it. Or not.

Hardware has no bearing on human life.