As you know from the previously discussed material, a frame-based network uses random-length frames that are generated by the source system.  The only limitation is that these frames must fall between minimum and maximum lengths, specified by standards.  The actual frames that are generated depend on the amount of data transmitted. For example, an Ethernet frame has a lower limit of 64 bytes and an upper limit of 1,518 bytes.  In some Token Ring networks, the size of a frame can be as large as 18,000 bytes.  At the first glance these variable-sized frames are very efficient at moving large amounts of data.  This is because a single frame could contain a large portion of the data and only a few frames are needed perform large data transmissions.  This concept makes frame-based technologies efficient in pure data networks.

On the other hand, variable-size frames are very difficult to predict.  The delay between the frames is variable and depends on several factors including the size of the frame.  Networking devices have to read and retransmit the frame.  Some types of devices (for example bridges, switches, and routers), must read and examine a portion of the frame to determine how to respond to it.  The necessity to process all incoming frames and to make individual decisions about each frame produces delays in moving through a forwarding device.  This delay is often referred to as latency.  Latency is usually not a problem for most data applications.  Voice and video transmissions, however, could be very sensitive to latency.  These transmissions are ever more sensitive to latency changes.  (If latency is high but constant, it will result in low quality but consistent audio or video signals).  This makes variable length frame-based technologies difficult to implement in environments where low and predictable latency is required.  This is where cell switching technologies such as ATM become very attractive.