All CPU related settings contains in one `cpu` object in config file, CPU backend allow specify multiple profiles and allow switch between them without restrictions by pool request or config change. Default auto-configuration create reasonable minimum of profiles which cover all supported algorithms.
Example below demonstrate all primary ideas of flexible profiles configuration:
*`"rx/wow"` Exact match to algorithm `rx/wow`, defined 4 threads without CPU affinity.
*`"cn"` Default failback profile for all `cn/*` algorithms, defined 2 threads with CPU affinity, another failback profiles is `cn-lite`, `cn-heavy` and `rx`.
*`"cn-lite"` Default failback profile for all `cn-lite/*` algorithms, defined 2 double threads with CPU affinity.
Each line represent one thread, first element is intensity, this option was known as `low_power_mode`, possible values is range from 1 to 5, second element is CPU affinity, special value `-1` means no affinity.
#### Short array format
```json
[-1, -1, -1, -1]
```
Each number represent one thread and means CPU affinity, this is default format for algorithm with maximum intensity 1, currently it all RandomX variants and cryptonight-gpu.
[MSR mod](https://xmrig.com/docs/miner/randomx-optimization-guide/msr). Enabled (`true`) or disabled (`false`). It gives up to 15% speedup depending on your system. _(**Note**: Userspace MSR writes are no longer enabled by default; the flag `msr.allow_writes=on` must be set for Linux Kernels 5.9 and after.)_
Restore MSR register values to their original values on exit. Used together with `wrmsr`. Enabled (`true`) or disabled (`false`).
#### `cache_qos`
[Cache QoS](https://xmrig.com/docs/miner/randomx-optimization-guide/qos). Enabled (`true`) or disabled (`false`). It's useful when you can't or don't want to mine on all CPU cores to make mining hashrate more stable.
#### `numa`
NUMA support (better hashrate on multi-CPU servers and Ryzen Threadripper 1xxx/2xxx). Enabled (`true`) or disabled (`false`).
#### `scratchpad_prefetch_mode`
Which instruction to use in RandomX loop to prefetch data from scratchpad. `1` is default and fastest in most cases. Can be off (`0`), `prefetcht0` instruction (`1`), `prefetchnta` instruction (`2`, a bit faster on Coffee Lake and a few other CPUs), `mov` instruction (`3`).
Enable (`true`) or disable (`false`) huge pages support for RandomX JIT code, by default `false`. It gives a very small boost on Ryzen CPUs, but hashrate is unstable between launches. Use with caution.
Force enable (`true`) or disable (`false`) hardware AES support. Default value `null` means miner autodetect this feature. Usually don't need change this option, this option useful for some rare cases when miner can't detect hardware AES, but it available. If you force enable this option, but your hardware not support it, miner will crash.
Mining threads priority, value from `1` (lowest priority) to `5` (highest possible priority). Default value `null` means miner don't change threads priority at all. Setting priority higher than 2 can make your PC unresponsive.
Use continuous, persistent memory block for mining threads, useful for preserve huge pages allocation while algorithm switching. Possible values `false` (feature disabled, by default) or `true` or specific count of 2 MB huge pages. It helps to avoid loosing huge pages for scratchpads when RandomX dataset is updated and mining threads restart after a 2-3 days of mining.
Allow override automatically detected Argon2 implementation, this option added mostly for debug purposes, default value `null` means autodetect. This is used in RandomX dataset initialization and also in some other mining algorithms. Other possible values: `"x86_64"`, `"SSE2"`, `"SSSE3"`, `"XOP"`, `"AVX2"`, `"AVX-512F"`. Manual selection has no safe guards - if your CPU doesn't support required instuctions, miner will crash.
