Human-like brain computer model

Already built supercomputers

Best computing systems

Date		System	Cores	TFlop/s	Power (kW)
1997	20 years ago	ASCI Red, Intel	478	1	≈500
2007	10 years ago	BlueGene/L, IBM	212 992	478	2 329
2017	Today	Sunway TaihuLight, NRCPC	10 649 600	93 000	15 371

Neuron simulation

Goal	Computing needed	Can simulate	Power needed
Simplified neuron	1 000 TFlop/s	93 neurons	165 kW per neuron
Electrophysiologically realistic Hodgkin-Huxley neuron	1 200 000 TFlop/s	0,077 neurons	199 623 kW per neuron

Human brain simulation

Goal	Computing needed	Can simulate %	Power needed	Of the world's yearly energy ^[1]
Simplified brain	1 × 10¹² TFlop/s	9,3 × 10^-7	1,65 × 10¹⁵ kW	≈ 151 %
Realistic brain ^[2]	1 × 10¹⁶ TFlop/s	9,3 × 10^-10	2,00 × 10¹⁸ kW	≈ 182 116 %

Possible to build with today's technology

Best AI computing components

Date		Component	TFlop/s	Power (kW)	Price
2017		NVIDIA Tesla V100	120	0,3	$6999

Neuron simulation

Goal	Computing needed	Components needed	Power needed	Cost
Simplified neuron	1 000 TFlop/s	8	2 kW per neuron	$55 992
Electrophysiologically realistic Hodgkin-Huxley neuron	1 200 000 TFlop/s	10 000	3 000 kW per neuron	$69 990 000

Human brain simulation

Goal	Computing needed	Components needed	Power needed	Of the world's yearly energy ^[1]
Simplified brain	1 × 10¹² TFlop/s	8 × 10¹⁰	3,0 × 10¹² kW	≈ 0,3 %
Realistic brain ^[2]	1,0 × 10¹⁶ TFlop/s	8 × 10¹³	3,0 × 10¹⁵ kW	≈ 273,7 %

Notes

[1]	World energy consumption 1,10 × 10¹⁵ kWh [Wikipedia].
[2]	Whole human brain processing rate estimated as 10²⁸ FLOPS (or 32 x 10²⁸ bits/sec) [Tuszynski, 2006]. Assumes classical computation on microtubules. Tuszunski explains that each microtubule dimer can have 32 states (four states per dimer, four states per electron inside the tubulin dimer, with at least two computational states), and since there are 13 dimers per ring and 1 250 rings per midsize microtubule, then each microtubule would have 100 kilobytes of information. The known number of microtubules per neuron suggests 10⁹ bytes per neuron, yielding 10¹⁹ bytes per brain assuming 10¹⁰ neurons per brain. If the electrons oscillate or make transitions in this state on the order of nanoseconds, then the processing power is 10²⁸ FLOPS (or 32 x 10²⁸ bits/sec).

[1]

World energy consumption 1,10 × 10¹⁵ kWh [Wikipedia].

[2]

Whole human brain processing rate estimated as 10²⁸ FLOPS (or 32 x 10²⁸ bits/sec) [Tuszynski, 2006].
Assumes classical computation on microtubules. Tuszunski explains that each microtubule dimer can have 32 states (four states per dimer, four states per electron inside the tubulin dimer, with at least two computational states), and since there are 13 dimers per ring and 1 250 rings per midsize microtubule, then each microtubule would have 100 kilobytes of information. The known number of microtubules per neuron suggests 10⁹ bytes per neuron, yielding 10¹⁹ bytes per brain assuming 10¹⁰ neurons per brain. If the electrons oscillate or make transitions in this state on the order of nanoseconds, then the processing power is 10²⁸ FLOPS (or 32 x 10²⁸ bits/sec).

How much computer power do we need for human-like brain computer model?

Already built supercomputers

Best computing systems

Neuron simulation

Human brain simulation

Possible to build with today's technology

Best AI computing components

Neuron simulation

Human brain simulation

Notes