Why std::memcpy on sram is slower than writing to dram using 'noc_async_write' ?

[codernine-moreh]

Using Device Program Profiler, I'm testing cycle durations for single tile write to dram and sram in NCRISC processor.
The one is to use noc_async_write api to write a single tile from circular buffer to dram and the other is to copy a memory block from cb to cb using 'tt_l1_ptr' raw pointer type.

I expected the later one should be faster than using noc api, but actually the result was opposite.
I wanna get to know if there is something that i missed.

• code snippet used to measure duartion on noc_async_wrtie call

//....
{
      DeviceZoneScopedN("SINGLE_TILE_NOC");
      noc_async_write(
          l1_read_addr,
          noc_addr,
         dtype_bytes * TILE_HW); 
      noc_async_write_barrier();
  }
//...

// duration : 575 cycles
PCIe slot, core_x, core_y, RISC processor type, timer_id, time[cycles since reset]
0,18,18,NCRISC,1649,568096169526,0,0,6,SINGLE_TILE_NOC,ZONE_START,153
0,18,18,NCRISC,1649,568096170101,0,0,6,SINGLE_TILE_NOC,ZONE_END,153,

• code snippet used to measure duartion on std::memcpy call

//....
tt_l1_ptr uint32_t* l1_rd_ptr =
    reinterpret_cast<tt_l1_ptr uint32_t*>(l1_read_addr);
tt_l1_ptr uint32_t* l1_wt_ptr =
    reinterpret_cast<tt_l1_ptr uint32_t*>(get_write_ptr(cb_temp));
{
  DeviceZoneScopedN("SINGLE_TILE_SRAM");
  std::memcpy(l1_wt_ptr, l1_rd_ptr, dtype_bytes * TILE_HW);
}
//...

// duration : 7956 cycles
PCIe slot, core_x, core_y, RISC processor type, timer_id, time[cycles since reset]
0,18,18,NCRISC,51207,470063023238,0,0,6,SINGLE_TILE_SRAM,ZONE_START
0,18,18,NCRISC,51207,470063031194,0,0,6,SINGLE_TILE_SRAM,ZONE_END,