Powerful new floating-point unit supporting single-cycle, double-precision calculations
The G3 has a low-end FPU intended for consumer machines before the 3D gaming boom. This design decision was made back when Motorola and IBM thought the G3 was going to be the low-end PowerPC CPU and the PowerPC 604 series was going to be the high-end. They were surprised to find that the backside cache of the G3 increased its general performance over that of the PowerPC 604 series by a large degree. The G4 has a heavy duty FPU similar to the PowerPC 604 series along with the speed boosting backside cache of the G3. This should cause a speed increase for 3D games.
Full 128-bit internal memory data paths
Data stream prefetching operations supporting four simultaneous 32-bit data streams
The wide data paths of the G4 allow it to sometimes process four different pieces of data at once. The 128-bit internal memory data paths give a wide enough bus to allow four 32-bit data streams to be processed at once. In some cases, this can translate to two to four times the performance of other CPUs per clock cycle. This is partly where the G4's gigaflop performance comes from.
1MB backside level 2 cache running at half the processor speed
The G4 PowerMacs Apple has announced all come with 1MB of level 2 cache. The G4 processor is capable of taking 2MB of L2 cache. We consider 2MB of L2 cache a likely option in future G4 servers. The performance gains of 2MB over 1MB are small for most operations. The performance gains of 1MB over 512k are also small for most operations. As for cache speed, while a full speed L2 cache is possible, the cost of such a cache is very high but in most situations, the performance gains are not.