AI Accelerators hardware for AI applications

[stanford]

I just wanted to get some idea on what kind of hardware the AI accelerators are made up of.

Does it have a bunch of cores optimized for multiplication and add? What kind of operation will it need to do for AI applications?

Thanks!

 

[andre_luis]

Generally speaking, these are ALU's that are implemented in cores capable of storing and processing arrays with large data mass, performing such algebraic and trigonometric mathematical operations internally with parallel processing as much as possible; the ability to work with numbers in floating point notation is a bonus. If you want to go further, have a search on GPU architectures.

 

[stanford]

Generally speaking, these are ALU's that are implemented in cores capable of storing and processing arrays with large data mass, performing such algebraic and trigonometric mathematical operations internally with parallel processing as much as possible; the ability to work with numbers in floating point notation is a bonus. If you want to go further, have a search on GPU architectures.

What kind of operations do AI application require? What would these hardware be optimized for? How would these hardware look different vs. the GPU?

 

[andre_luis]

What kind of operations do AI application require?

A lot of general questions with no specification of the target application. You migh be aware that AI covers a broad range of problems/approaches/solutions.

 

[ThisIsNotSam]

I just wanted to get some idea on what kind of hardware the AI accelerators are made up of.

Does it have a bunch of cores optimized for multiplication and add? What kind of operation will it need to do for AI applications?

Thanks!

No and yes and everything in between. I have built cores that had zero multipliers, some that had the 'add' operation in analog domain, and so on. It's too broad.