GPU power limit on macOS

• tools/nvgfx-smi (RUSD) reports a 50 W power limit, requested and enforced, and the only throttle reason under load is sw_power_cap. Windows runs the same GPU at 115 W.
• On macOS the RM gets no platform power input: every _DSM except NBCI answers "not supported", there is no NVPCF device, no AC/battery state, and nv_is_chassis_notebook() is FALSE. The Linux open driver without NVIDIA's closed nvidia-powerd behaves the same, so 50 W looks like the RM default with no platform input.
• Firmware (SSDT-21): the GPU node \_SB.PC00.RP01.PXSX has the GPS _DSM (rev 0x200, functions 0, 0x12, 0x13, 0x20 PSHARESTATUS, 0x2A PSHAREPARAMS; only when NVS GPSS != 0). \_SB.NPCF (HID NVDA0820) has the NVPCF _DSM (rev 0x200, functions 0 to 10; only when PXSX.NPCS != 0). Both read Arg3 as a Buffer.

nvrmgps and nvrmnvpcf do not need nvrmacpibl. Either of them also turns on:

• nv_is_chassis_notebook(): TRUE if ACPI has the NPCF device or a PNP0C0A battery.
• nv_acpi_is_battery_present(): Linux rule, _BIX[4] (or _BIF[3]) battery technology == 1. Not _STA (BAT0 reports _STA 0 under macOS here). Evaluated once.
• nv_acpi_get_powersource(): _PSR on the ACPI0003 device (ADP0), *ac_plugged = (_PSR == 1). RM reads it once at init (osinit.c, mobile config). No AC change notifications yet.

The new ACPI calls are refused when interrupts are off or in the ISR. Logs: NVRM-xnu: GPS _DSM 0x2a rev 0x200 in N B -> st (n B, ioret x), NVPCF _DSM ..., _PSR on ADP0 -> ..., first 20 calls then every 256th.

ABI in kexts/NVRM/nvrm_physctl_abi.h, sent through the normal escape selector (1) on an NVRM/NVAccel RM client. Root only (kauth_cred_issuser), and only with nvrmphysctl=1.

struct NVRMPhysCtrlEsc {      // 4128 bytes
    u32 version;              // 1
    u32 cmd;                  // RM control
    u32 paramsSize;           // <= 4096
    u32 status;               // out: NV_STATUS, or 0x80000001 off / 0x80000002 bad version / 0x80000003 bad size
    u32 mode;                 // 0 GSP subdevice, 1 GSP device, 2 GSP client, 3 kernel internal call
    u32 hClient, hObject;     // mode 3 only
    u32 reserved;
    u8  params[4096];         // in/out
};

Modes 0 to 2 call GPU_GET_PHYSICAL_RMAPI(pGpu)->Control(hInternalClient, hInternal{Subdevice,Device,Client}): straight to GSP, like the RM's own internal controls. Mode 3 calls rmapiGetInterface(RMAPI_GPU_LOCK_INTERNAL) on a CPU-RM object the caller owns, so controls flagged RMCTRL_FLAGS_INTERNAL (the 0x20800Axx family) pass the internal-call check. Both run in rm_nvgfx_phys_control() (kexts/NVRM/nvrm_physctl.c) with the API lock and all GPU locks held, after FULL_GPU_SANITY_CHECK.

sudo tools/nvgfx-smi/nvgfx-smi --phys CMD [WORD...] [--size N] [--target sub|dev|client] [--kernel]

CMD and WORDs are hex, WORDs are little-endian 32-bit params, --size pads the params to N bytes. Without --kernel it goes to GSP on RM's internal client; with --kernel it is an internal call on the tool's own subdevice (or device/client). The reply prints the status and the params as words.

Caution: these controls write GPU power policy. Start with GET controls; a bad SET can hang GSP or the GPU.

• N-73: GSP INTERNAL_PMGR_PFM_REQ_HNDLR_CONFIGURE_TGP_MODE (0x20800AD2, 1-byte bEnable) = 1 lifts the enforced limit from the VBIOS 50 W to 115 W. Burn 12.1 to 17.8 TFLOP/s, about 2.3 GHz, up to 116 W, 66 C, no throttle.
• nvrmtgp=1 (kexts/NVRM/nvrm_power.cpp): after the GPU is up, sends TGP mode on while _PSR says AC and off on battery, polled every 5 s, only on change, through rm_nvgfx_phys_control (internal client, physical API).
• Sysctls: debug.nvgfx_tgp_enable (RW, 1 auto, 0 force off), debug.nvgfx_tgp_state, debug.nvgfx_tgp_ac, debug.nvgfx_tgp_last_status, debug.nvgfx_tgp_applies.
• thread_call_cancel_wait is not exported to kexts (install CHECK caught it); stop uses cancel, wait, free.