The quick answer is that the more fragmentation you have on a network, the less efficient, and thus thus the slower the network will operate. If you consider IP packets (layer 3) as pieces of freight and the frames (Layer 2) as trucks carrying that freight, then you can get an idea of what I mean.
Trunks have a maximum capacity of X let’s say, and each piece of freight that comes in and is ready to be loaded on the truck must be of a size of X or less. If it is, it will be loaded and sent away. If it is larger than X, it must be broken in two or more and placed on two or more trucks. The trucks carrying the pieces of this freight must be identified somehow with documentation stating that when the trucks arrive at the next stop, the freight must be removed and reassembled. This whole procedure takes time, additional overhead information and more trucks, and is thus less efficient.
Similarly, if an IP packet is larger than the allowable frame size, during encapsulation, it will be fragmented into two or more pieces, and each piece will be placed in a different frame. Each IP fragment gets its own header (more overhead) and Information about the fragmentation is placed within that header. Each IP fragment is placed in a separate Ethernet frame each with its own header as well (more overhead). CPU resources are required by network devices to decapsualte, reorder and reassemble the IP packet.
Also keep in mind that an efficient network will carry frames that are close to the maximum allowed MTU. This ensures the largest data-to-overhead ratio, allowing for the highest throughput for the lowest amount of overhead. Fragmentation often results in frames much smaller than this maximum.
If you imagine that a network device processes hundreds of thousands or millions of packets per second, you can quickly realize how damaging IP packet fragmentation may get.
I hope this has been helpful!