Why integer and floating point are separated in terms of processor?

Answer 1

Since 1985, most processors are "pipelined." What these means is that instead of having the entire processing pathway executing one instruction, the pathway is divided into multiple units. The units pass their intermediate results through the use of registers, saving their results at the end of a cycle to a "pipeline register." At the beginning of the next cycle, the next unit in the datapath will read the result from the pipeline register as its input.

Think of pipelines as an assembly line production system. Originally you employ 5 mechanics to work on a car. While the 1st one works on the chassis, the other four sit and watch. The chassis is then passed onto the 2nd mechanic, who builds the engine while the other four have a rest. This way, you are only having 1 person work at a time while the others wait.

Now you decide to have them all work in parallel. The 1st mechanic produces a chassis, and passes it on to the 2nd mechanic. While the 2nd mechanic works on the works on the engine, the 1st mechanic goes back to building a new chassis, and so on. This way you have effectively increased your output to 5 times within the same period of time, while keeping your resources same and without changing the total time required for a single car to be built.

Processors work the same way. Only in the computing world, this assembly line work is called pipelining. For instance, in the early MIPS microprocessors, there were 8 pipleine stages. The 1st and the 2nd were instruction fetch (get instruction from memory) and instruction decode (decode instruction and get required input values). The 3rd stage was the execution stage, where the results to the computation are actually computed.

Generally, the execution stage is further pipelined to keep clock frequencies high. Now, integers and floating point numbers are stored differently. While an integer if just stored as its binary representation, FP numbers are stored as an integer number with another number that marks where the decimal point is located.

For instance, the number 20 is stored (in binary) as 0001 0100. Whereas, 20.2 might be stored as 202 x 10^(-1). While the 10 part is not stored, the numbers 2021 may be stored, the 1st three for the number, and the last for the exponent.

In order to do arithmetic operations on FP numbers, the first step is to bring them to a common exponent number. This requires an additional step in processing, translating to an additional pipeline stage. Also, integer operations are much more simplified in terms of binary and logic processing, requiring less pipeline stages.

Generally this results in a different number of pipeline stages. In the Pentium 4 processor, integer operations required 11 pipeline stages, while floating point operations required 17 operations. This makes integer operations execute faster and FP operations execute slower.

Hence the difference in the 2 for processor related operations.