Hi
May I know what is the peak IPC of 3080? I am talking about peak sm__inst_executed.sum.per_cycle_active value. Assuming 68 SMs and 4 subpartitions in each SM, will it be 272?
The calculation is correct for Volta - Ampere. The metrics to get this information are:
smsp__inst_executed.sum.peak_sustained # the sustained number of warp instructions per cycle (rolled up from SMSP)
== sm__inst_executed.sum.peak_sustained # the sustained number of warp instructions per cycle (rolled up from SM)
smsp__thread_inst_executed.sum.peak_sustained # the sustained number of thread instructions per cycle (rolled up from SMSP)
1 Like
@Greg
I saw your post here. You said that by considering limitations, for 2 warp schedulers per SM, the max IPC is 1.5. So, do you know the same number for 3080? As I said, the max IPC should be 68 * 4 * 1=272. However, if I assume that 4 scheduler issue up to 3 warp instructions, it will be 68 * 4 * 0.75=204.
Can you clarify that?
CC 3.x (Kepler) can sustain 7 warp instructions per cycle per SM. SM sub-partition supports dual-issue.
CC 5.x - 6.x (Maxwell-Pascal) can sustain 6 warp instructions per cycle per SM. SM sub-partition supports dual-issue.
CC 6.0 [GP100] can sustain 3 warp instructions per cycle per SM. SM sub-partition supports dual-issue.
CC 7.x - 8.x [Volta, Turing, Ampere) can sustain 4 warp instructions per cycle per SM. SM sub-partition supports single issue.
The original calculation you performed for the 3080 is correct.
1 Like