Average memory access time (AMAT)

Average memory access time is the expected mean latency of a memory hierarchy. Every access first pays cache hit time; the miss penalty is then weighted by the probability of a miss. The metric combines a fast common case with rarer slow cases. It does not replace full runtime measurement, but helps compare cache configurations. The simple AMAT model assumes every access first pays hit time and a miss then incurs an additional penalty. With multiple cache levels, local miss rates and penalties must be combined recursively. Miss rates are conditional on accesses that reach each level. These values describe latency per access; throughput and overlapping requests require additional models.

Use a 4 ns hit time, rmiss=0.05 and a 40 ns additional miss penalty: AMAT=4+0.05·40=6 ns. Enter the miss fraction as 0.05, not 5. Five percent of accesses incur the extra penalty, while every access still pays hit time. The simple AMAT model assumes every access first pays hit time and a miss then incurs an additional penalty. With multiple cache levels, local miss rates and penalties must be combined recursively. Miss rates are conditional on accesses that reach each level. These values describe latency per access; throughput and overlapping requests require additional models.

Here, miss penalty is the extra time after checking the cache. A miss therefore takes thit+tpenalty. Some sources instead call the entire miss latency miss time and use hit_rate·hit_time+miss_rate·miss_time. Check the source’s definition to avoid counting hit time twice. The simple AMAT model assumes every access first pays hit time and a miss then incurs an additional penalty. With multiple cache levels, local miss rates and penalties must be combined recursively. Miss rates are conditional on accesses that reach each level. These values describe latency per access; throughput and overlapping requests require additional models.

With hit time and penalty fixed, reducing miss rate lowers mean time linearly. If rmiss falls from 0.05 to 0.025 with a 40 ns penalty, the extra term drops from 2 ns to 1 ns. A larger cache may also increase hit time or power. Do not optimize only one factor in isolation. The simple AMAT model assumes every access first pays hit time and a miss then incurs an additional penalty. With multiple cache levels, local miss rates and penalties must be combined recursively. Miss rates are conditional on accesses that reach each level. These values describe latency per access; throughput and overlapping requests require additional models.

With serial levels, every access pays the L1 hit time. An L1 miss proceeds to L2 at its local miss rate; that access and a possible memory miss form the extra penalty. Use miss rates conditional on reaching each level, not unrelated global rates. A recursive calculation makes these paths explicit. The simple AMAT model assumes every access first pays hit time and a miss then incurs an additional penalty. With multiple cache levels, local miss rates and penalties must be combined recursively. Miss rates are conditional on accesses that reach each level. These values describe latency per access; throughput and overlapping requests require additional models.