Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
User avatar
towforce
Posts: 10336
Joined: Wed Mar 08, 2006 11:57 pm
Location: Birmingham UK

Re: Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Post by towforce » Tue May 05, 2020 11:18 am

yurikvelo wrote:
Tue May 05, 2020 7:22 am
you cannot increase number of qubits, since on large (macro-world) scale, wave function collapses (decoherence happens). only a few (in order of 1...100) qubits can work in a quantum state. systems with more than few (currently 8...10, possibly dozens) qubits work as a cluster of independent nodes, connected with extremely slow external networks.

Do you happen to know whether this is a fundamental problem, or whether improving technology might be able to solve it? If it's fundamental, it's a very big limitation. :(
The future is more important than the past.

User avatar
yurikvelo
Posts: 570
Joined: Sat Dec 06, 2014 12:53 pm

Re: Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Post by yurikvelo » Tue May 05, 2020 12:42 pm

towforce wrote:
Tue May 05, 2020 11:18 am
Do you happen to know whether this is a fundamental problem, or whether improving technology might be able to solve it? If it's fundamental, it's a very big limitation. :(
limitation on speed of light works at all scales.

Though quantum computing can exploit different "working bodies", all current implementations are electron or ion (atom with removed electron) manipulation (cooled down to 0.01...0.02 Kelvin and holded inside some crystalline semiconductor structures).

Though electron "speed" is fast, ~1/150 speed of light - it is not infinite (and even much less than speed of light). It's not actually speed in terms of macro mechanics, rather dependencies of its wave function, probabilities and interaction time. It's equivalent is like wave propagating with 1/150c.

In hydrogen atom, length of atom orbite is ~1/3 * 10^-10 (on it's average Bohr orbite)

Ratio of equivalent electron wave length (сompton wavelength) and orbite distance is known as "fine-structure constant" and is 1/137.036

2 * 10^6 "speed" at 1/3*10^-10 distances gives "frequencies" like 6*10^16 (7 orders higher than fastest conventional transistors)

But this happens so fast only at small (sub-atomic) scale. Fastest interactions (photon/EM-wave) travel only at speed of light. When you go far beyond atom sizes - interaction times increases linearly (at least linearly, when only distance limit).

"Reading/writing" a state to single electron is not possible (Uncertainty principle), so real-world implementations rely on a statistically big number of charge carriers (electrons or ions) - charged "islands/dots".

Real-world qubits are built as Josephson junction - a micrometer sized loops of superconducting metal interrupted by a number of Josephson junctions (very thin insulator between 2 superconductors). Due to small (micrometer-range) size and zero or low resistance (1 junction = 1/20208 V voltage drop or about 49 microVolts at >= critical current and zero voltage drop if <=critical current) - it works fast. Not 10^16 as single electron, but still many orders of magnitude faster than conventional transistors.

In conventional transistors your speed is limited by amount of charge you need to accumulate or dissipate when going from open to close state. When you decrease dimensions and increase current (it is limited by amount of free charge carriers, so you cannot easy decrease size and increase current) - you can change state faster. Practical limit is about 10^11 Hz.

In josephson "transistor" difference between states is whether you reach critical current or not. You don't need to pump energy for a long time. After you reach critical current, tunnel-junction become to radiate EM waves. You can put EMI detectors (magnetic loop antennas) and detect this radiation.
These detectors detect either intensity, or phase or current direction (clockwise or counterclockwise)

This detectors are slow as turtles. Your computation (qubit manipulation) is really fast (sub-atomic time scale), but reading experiment outcome takes ages (many orders of magnitude slower than conventional transistor).

At 0.01....0.02K temperatures, longest experiments run only dozen of nanoseconds (due to many instability factors quantum state decoherates fast: you still apply external energy source and emit EM waves, this is very unstable and bulky installation). For some exotic implementations (nitrogen atoms inside diamond cystal) up to dozen seconds coherence time achieved (carbon atoms in the diamond are replaced by a nitrogen atom and an empty site with a floating loose electron)

A number of qubits taken together is a qubit register. Quantum computers perform calculations by manipulating qubits within a register.

Register with 10 and even 20 qubits were created (non-Josephson). E.g. 20-qubit register was 20 charged calcium atoms (ions) arranged in a line served as the qubits, and they were entangled using a series of lasers. Researchers were able to get the calcium ions to entangle with two, three, or occasionally even four other calcium ions in the system (out of 20-qubit not a single one is controllable, and entaglement happens with at most 2-3 qubits).

There are quantum systems such as ultra-cold gases in which entanglement between a large number of particles has been detected (but not manipulated - we just say "hey, see: entaglement occur in nature"

IBM's 50-qubit machine and Google's 72-qubit Bristlecone - individual quantum states of the qubits aren't controllable, and nor can the system read out the individual qubits.

Working prototypes that can compute at least anything are all Josephson-junctions, acting as individual qubits. Each individual qubit is controlled using bulky and enormously slow probes with conventional transistors and conventional DSP to process signals.

Computing 6 = 2 * 3 is similar to hunting on Higgs bozon: it takes years of time, megawatts of energy, megatons of steel, concrete; hundreds of scientists, tons of coffee and good luck, couple Noble prizes, kilometers of wires, clusters of Intel Xeon and nVidia Tegra.

If no one can reproduce your results - claim it as "quantum superiority": "you bastards, don't even try"

User avatar
towforce
Posts: 10336
Joined: Wed Mar 08, 2006 11:57 pm
Location: Birmingham UK

Re: Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Post by towforce » Tue May 05, 2020 1:25 pm

yurikvelo wrote:
Tue May 05, 2020 12:42 pm
towforce wrote:
Tue May 05, 2020 11:18 am
Do you happen to know whether this is a fundamental problem, or whether improving technology might be able to solve it? If it's fundamental, it's a very big limitation. :(
limitation on speed of light works at all scales.

Though quantum computing can exploit different "working bodies", all current implementations are electron or ion (atom with removed electron) manipulation (cooled down to 0.01...0.02 Kelvin and holded inside some crystalline semiconductor structures).

Though electron "speed" is fast, ~1/150 speed of light - it is not infinite (and even much less than speed of light). It's not actually speed in terms of macro mechanics, rather dependencies of its wave function, probabilities and interaction time. It's equivalent is like wave propagating with 1/150c.

In hydrogen atom, length of atom orbite is ~1/3 * 10^-10 (on it's average Bohr orbite)

Ratio of equivalent electron wave length (сompton wavelength) and orbite distance is known as "fine-structure constant" and is 1/137.036

2 * 10^6 "speed" at 1/3*10^-10 distances gives "frequencies" like 6*10^16 (7 orders higher than fastest conventional transistors)

But this happens so fast only at small (sub-atomic) scale. Fastest interactions (photon/EM-wave) travel only at speed of light. When you go far beyond atom sizes - interaction times increases linearly (at least linearly, when only distance limit).

"Reading/writing" a state to single electron is not possible (Uncertainty principle), so real-world implementations rely on a statistically big number of charge carriers (electrons or ions) - charged "islands/dots".

Real-world qubits are built as Josephson junction - a micrometer sized loops of superconducting metal interrupted by a number of Josephson junctions (very thin insulator between 2 superconductors). Due to small (micrometer-range) size and zero or low resistance (1 junction = 1/20208 V voltage drop or about 49 microVolts at >= critical current and zero voltage drop if <=critical current) - it works fast. Not 10^16 as single electron, but still many orders of magnitude faster than conventional transistors.

In conventional transistors your speed is limited by amount of charge you need to accumulate or dissipate when going from open to close state. When you decrease dimensions and increase current (it is limited by amount of free charge carriers, so you cannot easy decrease size and increase current) - you can change state faster. Practical limit is about 10^11 Hz.

In josephson "transistor" difference between states is whether you reach critical current or not. You don't need to pump energy for a long time. After you reach critical current, tunnel-junction become to radiate EM waves. You can put EMI detectors (magnetic loop antennas) and detect this radiation.
These detectors detect either intensity, or phase or current direction (clockwise or counterclockwise)

This detectors are slow as turtles. Your computation (qubit manipulation) is really fast (sub-atomic time scale), but reading experiment outcome takes ages (many orders of magnitude slower than conventional transistor).

At 0.01....0.02K temperatures, longest experiments run only dozen of nanoseconds (due to many instability factors quantum state decoherates fast: you still apply external energy source and emit EM waves, this is very unstable and bulky installation). For some exotic implementations (nitrogen atoms inside diamond cystal) up to dozen seconds coherence time achieved (carbon atoms in the diamond are replaced by a nitrogen atom and an empty site with a floating loose electron)

A number of qubits taken together is a qubit register. Quantum computers perform calculations by manipulating qubits within a register.

Register with 10 and even 20 qubits were created (non-Josephson). E.g. 20-qubit register was 20 charged calcium atoms (ions) arranged in a line served as the qubits, and they were entangled using a series of lasers. Researchers were able to get the calcium ions to entangle with two, three, or occasionally even four other calcium ions in the system (out of 20-qubit not a single one is controllable, and entaglement happens with at most 2-3 qubits).

There are quantum systems such as ultra-cold gases in which entanglement between a large number of particles has been detected (but not manipulated - we just say "hey, see: entaglement occur in nature"

IBM's 50-qubit machine and Google's 72-qubit Bristlecone - individual quantum states of the qubits aren't controllable, and nor can the system read out the individual qubits.

Working prototypes that can compute at least anything are all Josephson-junctions, acting as individual qubits. Each individual qubit is controlled using bulky and enormously slow probes with conventional transistors and conventional DSP to process signals.

Computing 6 = 2 * 3 is similar to hunting on Higgs bozon: it takes years of time, megawatts of energy, megatons of steel, concrete; hundreds of scientists, tons of coffee and good luck, couple Noble prizes, kilometers of wires, clusters of Intel Xeon and nVidia Tegra.

If no one can reproduce your results - claim it as "quantum superiority": "you bastards, don't even try"

Thanks for the comprehensive answer. Google did claim to have achieved quantum superiority last year.
The future is more important than the past.

mehmet123
Posts: 201
Joined: Sun Jan 26, 2020 9:38 pm
Location: Turkey
Full name: Mehmet Karaman

Re: Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Post by mehmet123 » Tue May 05, 2020 4:01 pm

Dann Corbit wrote:
Tue May 05, 2020 5:00 am
I guess they will use a quantum computer and solve chess in a single machine cycle.
We don't know their computer technology is far ahead of us. USSR is the first country to sent satellite,animal,human,woman vs. to space. But their computer technology is very behind of western countries.

mwyoung
Posts: 2336
Joined: Wed May 12, 2010 8:00 pm

Re: Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Post by mwyoung » Wed May 06, 2020 4:00 am

mehmet123 wrote:
Sun May 03, 2020 6:58 pm
Leo wrote:
Sun May 03, 2020 4:55 pm
AlphaZero at 5000 TPU. Stockfish cant utilize all those cores.
The speed of Stockfish at 4096 cores is ~4.800.000.000/second. It'a very exciting speed. I don't think Stockfish has a serious scaling problem.

If Stockfish has ability of to use 4096 cores so I don't think Stockfish will has a problem using 30.000 cores.
Stockfish does have issues scaling. On Playchess I have a higher rating then 64 core hardware, 54 core hardware, and 32 core hardware vs 16 core 2950x. My guess is they are over driving Stockfish with too many threads. As I out search them in depth, not NPS. All the users are running 128 threads, 108 threads, or 64 threads. This is played in real world rated games. And my games are played on automatch, against all players. Not cherry picked.
Professing themselves to be wise, they became fools,
take on me. Foes 0.

User avatar
yurikvelo
Posts: 570
Joined: Sat Dec 06, 2014 12:53 pm

Re: Which is stronger AlphaZero( 5000 TPU) or Stockfish(30.000 Cores)

Post by yurikvelo » Wed May 06, 2020 5:06 am

Laskos wrote:
Mon May 04, 2020 7:57 am
With Lazy SMP fixed depth is not a correct measure of strength. Probably the 8 threaded search is wider to the same depth compared to 4 threaded search, and probably stronger to the same depth.
Stockfish_20050217_8T - Stockfish_20050217_4T
Syzygy 6-man, Hash=2048, TC = 60" + 1"
RepeatStartPosition=1
StartPGNEPD=C:\Inst\CHESS\noob_3moves.epd

99,0/175 57% +49
+30-7=138 0101======11====0==1======1========1==01=====11=========1=======1================1===1=1=1==1=====1=0=1=1==1===1===1=====11=====1==0======1============1============1==10======

Post Reply