nvgfx-smi

Read-only GPU telemetry for NvGfx on macOS: power limit, power draw, clocks, pstate, temperature and throttle reasons. It talks to the NVRM kext the same way the NVK userland does and needs no root.

./build.sh                      # xcrun clang, x86_64, headers from ~/src/deps/open-gpu-610.57.04 (override with OPEN_GPU=...)
./nvgfx-smi                     # same as --once: human-readable report plus every RM status
./nvgfx-smi --json              # one report as JSON
./nvgfx-smi --watch 200         # CSV, one line every 200 ms until Ctrl-C (default 1000 ms)
./nvgfx-smi --verbose           # also log transport details and thermal sensors to stderr

CSV columns: time,pstate,gpc_mhz,mem_mhz,power_mw_inst,power_mw_avg,limit_req_mw,limit_enf_mw,tgp_mw,temp_c,throttle_mask,throttle. A load to watch it against: ../nvgfx-burn/nvgfx-burn 60.

1. Open an NVAccel user client (open type 0x4E560000|0, falls back to NVRM type 0), bind a /dev/null fd to it (selector 3).
2. NV_ESC_CARD_INFO for the gpu id, alloc NV01_ROOT_CLIENT, NV0000_CTRL_CMD_GPU_GET_ID_INFO_V2, alloc NV01_DEVICE_0 and NV20_SUBDEVICE_0 (escape selector 1, as in nvrm_xnu.c).
3. Alloc RM_USER_SHARED_DATA (class 0x00DE, "RUSD") under the subdevice with polledDataMask = CLOCK|PERF|POWER|THERMAL.
4. Map it the way NVK maps system memory: open a second ctl client, send NV_ESC_RM_MAP_MEMORY (read-only flag) on the first one naming the second's fd, then IOConnectMapMemory64(type 0, kIOMapReadOnly) on the second. The kext's clientMemoryForType hands out the RUSD pages.
5. Each sample reads the RUSD sections with the timestamp seqlock and also issues the plain controls below.

Only GET controls are sent. Everything is freed on exit (freeing the root client frees the tree).

RUSD fields, populated or not:

Field notes:

• pstate: P0 is full performance; P4 is what the GPU sits in at the desktop.
• gpc_mhz, mem_mhz: RUSD actual clocks. The PMUMON control gives the same graphics clock; its "active GPC" value stayed at 2790 MHz in every sample, so it is not a live reading.
• limit_req_mw / limit_enf_mw: the GPU power limit the RM asked for and the one being enforced.
• power_mw_inst / power_mw_avg: instant and averaged GPU power.
• throttle: decoded RUSD_CLK_THROTTLE_REASON_* bits (gpu_idle, app_clock_setting, sw_power_cap, hw_slowdown, sync_boost, sw_thermal_slowdown, hw_thermal_slowdown, hw_power_brake, display_clock_setting).
• The perf level clock domains (0x1 at 180 MHz, 0x10 at 405 MHz) do not match the graphics or memory clock; the open headers do not name the domain ids, so treat them as unidentified.

Idle (desktop, nothing running), --once:

pstate            RUSD P4   control P4
graphics clock    1417 MHz (target 1417 MHz)   PMUMON gpc 1417 MHz active 2790 MHz
memory clock      9001 MHz (target 9001 MHz)   PMUMON dram 9001 MHz
power limit       requested 50000 mW, enforced 50000 mW
power draw        instant 14749 mW, average 12539 mW, TGP n/a mW
temperature       GPU 59.8 C (RUSD)   control 59 C   memory 62.0 C
throttle reasons  none (0x0)

--watch 200 with nvgfx-burn 10 started at about 5 s (every fourth line):

time,pstate,gpc_mhz,mem_mhz,power_mw_inst,power_mw_avg,limit_req_mw,limit_enf_mw,tgp_mw,temp_c,throttle_mask,throttle
0.000,P4,1455,9001,10913,12339,50000,50000,n/a,57.8,0x0,none
4.891,P4,1440,9001,10996,11188,50000,50000,n/a,57.4,0x0,none
5.704,P0,1477,12001,38537,51400,50000,50000,n/a,63.6,0x4,sw_power_cap
6.522,P0,1905,12001,59811,51284,50000,50000,n/a,67.4,0x4,sw_power_cap
8.960,P0,1597,12001,43191,49282,50000,50000,n/a,65.8,0x4,sw_power_cap
12.203,P0,1807,12001,52794,50444,50000,50000,n/a,68.2,0x4,sw_power_cap
14.631,P0,1725,12001,48614,49995,50000,50000,n/a,68.4,0x4,sw_power_cap
16.255,P4,1417,9001,13818,14596,50000,50000,n/a,61.1,0x0,none

nvgfx-burn reached about 12.0 to 12.3 TFLOP/s FP32 during that run. Under load the average draw sits right on the 50 W limit, the graphics clock moves between about 1600 and 1900 MHz, and the only throttle reason is sw_power_cap. The temperature stays under 70 C, so it's the power limit holding the GPU back.