There is some tips to speed-up keyhunt-cuda (rotor-cuda):

Apply this then you need less grid size, like 4096x512 will be enough for 4090:

https://bitcointalk.org/index.php?topic=5244940.msg63526413#msg63526413

Also change this:

__device__ __noinline__ void CheckHashSEARCH_MODE_SA(uint64_t* px, uint64_t* py, int32_t incr, uint32_t* hash160, uint32_t* out)
{
   switch (mode) {
   case SEARCH_COMPRESSED:
      CheckHashCompSEARCH_MODE_SA(px, (uint8_t)(py[0] & 1), incr, hash160, out);
      break;
   case SEARCH_UNCOMPRESSED:
      CheckHashUnCompSEARCH_MODE_SA(px, py, incr, hash160, out);
      break;
   case SEARCH_BOTH:
      CheckHashCompSEARCH_MODE_SA(px, (uint8_t)(py[0] & 1), incr, hash160, out);
      CheckHashUnCompSEARCH_MODE_SA(px, py, incr, hash160, out);
      break;
   }
}

to this because doing switch-case in kernel is very bad idea:

__device__ __noinline__ void CheckHashSEARCH_MODE_SA(uint64_t* px, uint64_t* py, int32_t incr, uint32_t* hash160, uint32_t* out)
{
   
   CheckHashCompSEARCH_MODE_SA(px, (uint8_t)(py[0] & 1), incr, hash160, out);
      
}

also maxFound can be completely removed to search puzzle, because we need only one return result anyway

Thanks.