3 ms·

64 bit computing became dominant because of the need to address memory larger than 32 bits. They have the capability to scale quite 'easily' to 128 bits or furt

by bArray 5y ago

64 bit computing became dominant because of the need to address memory larger than 32 bits. They have the capability to scale quite 'easily' to 128 bits or further if they wanted to, there just wouldn't be a benefit. The Play Station 2 for example has 128 bit SIMD [1], 128 bit computing has been possible for a long time. Once you have the ability to create a processor at 7nm with some bits, scaling is not so tough. Even if you cannot reliably create larger pieces of silicon, you just do something like AMD did with multiple dies connected by a fabric to mitigate risk. Absolutely worst case, you have a motherboard with multiple processors, or even computers in different buildings. In terms of qubits, it is very likely that the problem can be distributed over multiple quantum computers. A significantly incentivized actor could definitely pull it off. If you can reliably manufacture ~100 qubit quantum computers, it's just a matter of scale. [1] https://en.wikipedia.org/wiki/PlayStation_2_technical_specifications#Central_processing_unit https://en.wikipedia.org/wiki/PlayStation_2_technical_specif...

tsimionescu 5y ago
> In terms of qubits, it is very likely that the problem can be distributed over multiple quantum computers. No, this is very wrong. Qubits are only different from classical bits of they can communicate before becoming entangled with the environment (decoherence). You can't run some kind of "quantum cable" between two separate QCs in a rack and get twice the qubits - the interactions with the wire will break the entanglement between the qubits, and you will just have an unreliable classical computer with 100 bits of memory. To perform a quantum computation, ALL the qubits (all your memory) must be in an entangled state together - this is the massive problem. Even worse, this state must be maintained while applying different transformations on the qubits from the outside.
- bArray 5y ago
  In the paper they suggest that spending ~24 hours more doing computation means you can divide the number of qubits by 24, indicating that at least in one sense it is linearly scalable. From my limited understanding, you process for a given amount of time, after which you can classically pull out an answer with some given probability, with some trade off with time and noise. I imagine it would be somewhat possible to have several quantum computers running in parallel which end early, each correctly deducing the answer with some given probability. If each of the N^x machines has a 1/N chance of having the correct answer, you could simply test each solution classically. And that assumes there is not some way to seed the search effort classically during the setup of the quantum circuit.
  - tsimionescu 5y ago
    There are two aspects here. In order to implement a quantum algorithm, you need a minimum number of interacting qubits, which the article calls a qubit depth. All of these qubits must be entangled with each other in order to achieve any quantum speed-up. Below this number, your quantum algorithm simply can't be represented on the machine - it would be like trying to multiply two 128-bit numbers on a processor with 2 8-bit registers: you simply don't have enough working memory to do the calculations you need. An extra complication for QCs is that you also need error correcting calculations in addition to your base calculations. So, if you want to multiply 2 128-bit numbers, not only do you need at least 256 (q)bits of working memory, you need some additional number to correct for errors in the calculation - and with currently known error correcting methods, you need A LOT more. That's why the article is giving a minimum number of qbits for the Bitcoin calculation: 10^7 physical bits, which represent a measly ~2000 logical (perfect) qubits. This is the minimum number you would need to keep entangled for your your minimum clock period. We are currently at 10^2, and even getting to 200 is a research-level task; 10^3 is far away. Once we get to something like 10^7, we may be able to start thinking of parallelizing at the whole machine level. Even still, it's important to understand that quantum algorithms have, as far as we know for now, an exponential advantage over classical algorithms (note: this only applies to certain algorithms, NOT any algorithm). This only applies as long as you are running in the quantum regime. That is, if a particular quantum computer can resolve a problem for N components in 1 minute, and a quantum computer of double the qubit number can finish it in 30s, 2 QCs of the first type will finish it in something like 59s, since they will not benefit from the exponential quantum speedup.
ninkendo 5y ago
I’m beginning to worry that the 100% sarcasm in my post has gone un-noticed…

by bArray 5y ago

tsimionescu 5y ago

> In terms of qubits, it is very likely that the problem can be distributed over multiple quantum computers. No, this is very wrong. Qubits are only different from classical bits of they can communicate before becoming entangled with the environment (decoherence). You can't run some kind of "quantum cable" between two separate QCs in a rack and get twice the qubits - the interactions with the wire will break the entanglement between the qubits, and you will just have an unreliable classical computer with 100 bits of memory. To perform a quantum computation, ALL the qubits (all your memory) must be in an entangled state together - this is the massive problem. Even worse, this state must be maintained while applying different transformations on the qubits from the outside.

bArray 5y ago
In the paper they suggest that spending ~24 hours more doing computation means you can divide the number of qubits by 24, indicating that at least in one sense it is linearly scalable. From my limited understanding, you process for a given amount of time, after which you can classically pull out an answer with some given probability, with some trade off with time and noise. I imagine it would be somewhat possible to have several quantum computers running in parallel which end early, each correctly deducing the answer with some given probability. If each of the N^x machines has a 1/N chance of having the correct answer, you could simply test each solution classically. And that assumes there is not some way to seed the search effort classically during the setup of the quantum circuit.
- tsimionescu 5y ago
  There are two aspects here. In order to implement a quantum algorithm, you need a minimum number of interacting qubits, which the article calls a qubit depth. All of these qubits must be entangled with each other in order to achieve any quantum speed-up. Below this number, your quantum algorithm simply can't be represented on the machine - it would be like trying to multiply two 128-bit numbers on a processor with 2 8-bit registers: you simply don't have enough working memory to do the calculations you need. An extra complication for QCs is that you also need error correcting calculations in addition to your base calculations. So, if you want to multiply 2 128-bit numbers, not only do you need at least 256 (q)bits of working memory, you need some additional number to correct for errors in the calculation - and with currently known error correcting methods, you need A LOT more. That's why the article is giving a minimum number of qbits for the Bitcoin calculation: 10^7 physical bits, which represent a measly ~2000 logical (perfect) qubits. This is the minimum number you would need to keep entangled for your your minimum clock period. We are currently at 10^2, and even getting to 200 is a research-level task; 10^3 is far away. Once we get to something like 10^7, we may be able to start thinking of parallelizing at the whole machine level. Even still, it's important to understand that quantum algorithms have, as far as we know for now, an exponential advantage over classical algorithms (note: this only applies to certain algorithms, NOT any algorithm). This only applies as long as you are running in the quantum regime. That is, if a particular quantum computer can resolve a problem for N components in 1 minute, and a quantum computer of double the qubit number can finish it in 30s, 2 QCs of the first type will finish it in something like 59s, since they will not benefit from the exponential quantum speedup.

ninkendo 5y ago

I’m beginning to worry that the 100% sarcasm in my post has gone un-noticed…