I want to know what is the difference between a normal chip or processor and a processor designed for AI?
@beta I agree that all questions should be allowed, my reasoning for this is because everyone must start somewhere, a database with all the answers (all the final compositions of complex structures) would fail to answer the simplest of questions, the reason for this is because they were assumed to be already known. a database of solutions must account for all possible problems thus a base is very important when building a structured form of knowledge. The base ( metaphorically the simplest question) will eventually bring the user of the database to the top ( most complex ) form of their question. In some situations a user may not know the right question to ask. By answering simple forms of a question we are able to form compound structures ( questions built from multiple questions being answered to discover the correct question to ask ) the benefit of this is that anyone can ask a complicated question ( one that they do not comprehend or know how to ask yet ) and fully understand the answer by learning the correct question to ask ( the answer to the original question can only be achieved if the answers to the remaining parts of the compound question are known) using this logic one can see the importance of being able to give an adequate answer to even the simplest or most vague of questions, if a similar question is asked again the proper answer can be given after you know what part the user does not understand.
As to the answer to the OP's question, I believe that the difference between processors is the ability to hold possible answers in a suspended state ( committed to neither true or false ) until more of the equation has been solved) think of this as the ability to answer a complicated question made up of multiple yes or no questions. An ai processor will allow for the equation to be answered as though both possibilities are true even though only a single answer can exist. That processor must answer ( solve for that part of the equation ) without committing to a single answer until the final pieces of information have been made known. At that point the AI can narrow a certain question down to only a few possible answers compared to an infinite number of answers. Once narrowed down the equation can then be solved backwards using the known values and referencing them to the working values of other possible answers. What is given as a result is an event based calculation as to the probability of an outcome, dependent on unknown variables that are solvable only when the answer to their state is known. AI chips simply allow for quantum processing in which a single piece of information is true and false while being solved and will choose a single value as you solve the equation by adding more known values to the equation that pertain to the unknown values calculation. values can be added manually or by the AI successfully providing proof that a certain value has to be at a certain state for the entire equation to be solved.
Specialized AI hardware takes advantage of highly parallelizable nature of many neural network designs. GPUs are designed for pixel crunching - coincidentally very paralelisable too. This is why they often offer orders of magnitude better performance (in NN tasks) than CPUs. That being said, GPUs have shortcomings to. First, many of GPU's built in features are not utilized in neural network applications- in many cases accurate floating point operations are not necessary, data for simple calculations involved in NNs. If you remove all the unused features from GPUs and use that extra space for more compute units, you get something like Googles' https://cloud.google.com/blog/big-data/2017/05/an-in-depth-look-at-googles-first-tensor-processing-unit-tpu TPU.
Parallel processing is very important when computing ANNs. Though they are not designed for AIs, GPUs are widely used in machine learning because of their higher capability of paralel processing and it is due to their higher number of cores comparing CPUs.
All in all, the dedicated hardware should have GPU-like architecture.