CPU Vs GPU Vs NPU

The Core

AI is the buzzword of the modern digital age. And to further the cause,  hardware needs to keep pace with the software demand. We now have high end laptops and desktop PC which provide unmatched performance and scalability.

In computation of machine learning algorithms,  three common factors comes up frequently when accessing the viability of a device.

CPU

Also known as the Central Processing Unit, this forms the brain of the system. It handles multiple operations from different peripherals and components like Ethernet, WiFi, Sound and Graphics card. However this is done in a linear order. This can affect the speed and bandwidth capability. Nowadays CPU has multiple cores to perform multiple subsets of the same task .They can be dual core (2), quad core (4), hexa core (6) or quad core (8) . Higher core strength relates to better performance, though leading to higher power consumption. Most of the devices are connected through PCI Express slot (Peripheral Component Interface ). They use serial topology with different routes for data, address and control instructions. This increases the bandwidth and transfer speed rate. These slots also uses lower number of pins, reducing overall component space. Examples of CPU include Intel i3, i5, i7, i9, Lunar lake and AMD Ryzen 3, 5, 7, 9, ThreadRipper.

GPU

Stands for Graphics Processing Unit. Same as CPU, they perform complex computational operations. They have multiple cores, separate RAM (Random Access Memory), storage for data and command instructions. However they are specifically used for Graphical and Visual operations such as video rendering, machine learning and AI operations. They execute parallel processing where a complex task is separated into subsets and distributed among different cores. This increases efficiency and higher output rate. There are three subdivisions of GPU. They include

• Discrete GPU : They act as a standalone component separate from the CPU and connected through data buses and PCI Express slots. They have dedicated memory and data storage with separate chip sets. Higher power consumption, but with more operational viability.

• Integrated GPU : Here the CPU is integrated with the GPU acting as one single interface. This allows for faster data transfer and lower space requirement. Can be considered as a system on chip concept. Popular in gaming and medium video editing operations.

• Virtual GPU : This GPU has no hardware and exists only in the cyber space or virtual environment. Mostly found in Cloud services.

NPU

Neural Processing Unit, a recent addition to the consumer electronics like laptops and mobile phones, are designed on the human brain concept. They perform neuron computation, by interlinking a mesh of nodes. This allows for complex mathematical computation and is specially designed for deep learning and machine learning. Operation is confined to 8 bit to reduce complexity and faster data processing. 

Also integrate hardware acceleration such as systolic array architecture or improved tensor processing.  NPU have specialized modules for vector addition, and matrix multiplication commonly used in AI.

Final Note

Advanced machine learning is performed on a symphony of these three chip sets working in  unison. They break complex algorithmic operations into smaller packets/bits to produce intended outputs.