RTX 5090 Laptop (Colorful G16 Pro): Dynamic Boost capped at 120W after fixing nvidia-powerd — is 175W reachable on Linux?

Summary

On a Colorful G16 Pro (i9-14900HX + RTX 5090 Laptop GB203M, Deepin 23 / kernel 6.6.155, driver 595.99.02 proprietary), the GPU was stuck at the 95W default TGP. I fixed the nvidia-powerd setup and Dynamic Boost now works, but SBIOS only ever grants 120W under load — never more, while Windows + Colorful Control Center (turbo profile) reportedly reaches 175W on this chassis.

Questions:

  1. Has anyone with this or a similar OEM (Colorful / Tongfang / Clevo-based) gotten the full 175W boost budget on Linux?
  2. Is the boost budget purely SBIOS/NVPCF-controlled, or is there anything driver-side that can raise the cap?
  3. Is there a known mapping between OEM performance profiles (EC/WMI) and the NVPCF budget that nvidia-powerd negotiates?

What was wrong and what I fixed

Deepin’s NVIDIA driver package (595.99.02) ships /usr/share/doc/NVIDIA_GLX-1.0/nvidia-dbus.conf but does not install it to /etc/dbus-1/system.d/. Consequences:

  • nvidia-powerd ran but could not register its bus name → Dynamic Boost dead
  • GPU pinned at Default Power Limit 95W, compute loads ~800MHz, “SW Power Cap” throttle active

Fix (verified working, powerd now owns nvidia.powerd.server on the system bus):

sudo mkdir -p /etc/dbus-1/system.d
sudo cp /usr/share/doc/NVIDIA_GLX-1.0/nvidia-dbus.conf /etc/dbus-1/system.d/
sudo systemctl restart nvidia-powerd

Current state after fix

  • Idle: Current Power Limit 120W (Default still 95W) — boost active
  • Under sustained compute load (nvidia-smi dmon, 1s samples): power flat at 117–118W, SM clock 900–930MHz, GPU temp 47→63°C over 90s, mem clock 14001MHz
  • nvidia-smi -pl → “Changing power management limit is not supported in current scope” (expected on mobile)
  • powerd journal shows no errors (no “SBIOS requested to disable Dynamic Boost”, no AutoFL errors)

Things I already ruled out

  • CPU power is not the bottleneck: lowering Intel RAPL PL1 from 110W → 55W → 35W does not change the GPU cap at all (still exactly 118W). Total system draw during compute is ~135W against a 330W adapter.
  • AC power is connected (ADP1 online=1); battery reports “Not charging 98%” — if the OEM firmware gates boost on battery state this may be it, but I have no way to verify on Linux.
  • All CPUs on performance governor; powerd’s documented schedutil/powersave interaction looks orthogonal since the cap never moves regardless of CPU limits.
  • No WMI driver claims this machine: no colorful/tongfang driver in kernel 6.6, /sys/firmware/acpi/platform_profile does not exist, ~20 unclaimed WMI GUIDs in /sys/bus/wmi/devices/.

Data points

state power limit SM clock compute throughput
before fix (powerd dead) 95W 800MHz 100% (baseline)
after fix 120W (dynamic) 930MHz ~115%
Windows + Colorful Control Center turbo (reviews/other users) up to 175W ~2GHz+ n/a

System info

  • Laptop: Colorful G16 Pro (DMI: COLORFUL / G16 Pro), i9-14900HX, RTX 5090 Laptop 24GB (10de:2c18)
  • OS: Deepin 23 (beige), kernel 6.6.155-amd64-desktop-hwe
  • Driver: 595.99.02 proprietary
  • nvidia-powerd: 595.99.02, active, DBus connected, no errors
  • /proc/driver/nvidia/gpus/…/power: “Notebook Dynamic Boost: Supported”

Any pointers appreciated — especially from anyone with a Colorful/Tongfang 50-series laptop. I can attach an nvidia-bug-report.log.gz if needed.

[SOLVED] Root cause found and fixed — full 175W unlocked

Update: I reverse-engineered the firmware side and got the full 175W working on Linux. Sharing the full analysis and fix here since this thread covers everything needed.

Root cause (three layers):

  1. Deepin’s driver package doesn’t install nvidia-dbus.conf to /etc/dbus-1/system.d/, so nvidia-powerd couldn’t register its bus name → Dynamic Boost dead (fixed earlier in this thread, got me from 95W to 120W).
  2. Even with powerd running, the cap stayed at 120W. Decompiling the DSDT revealed why: this OEM (Colorful, Tongfang-based chassis) keeps a firmware-side authorization flag — _SB.NPCF.CTGP (inside the NVPCF device, HID NVDA0820) — which is only set to 1 when the Windows Control Center switches performance profiles. Under Linux it stays 0 forever, and inside NPCF._DSM func 0x02: If CTGP == One: TGPA = ACBT, Else: TGPA = Zero. So powerd’s negotiation only ever received the base budget. This also explains the “SBIOS has requested to disable Dynamic Boost DC controller” error: the SBIOS refuses boost while the profile authorization is unset.
  3. The activation path is _SB.PC00.LPCB.EC0.RCAP(0,0) — the same function the Windows tool triggers via WMI (GUID 644C5791-B7B0-4123-A90B-E93876E0DAAD, method WSAA, sub-command 0x0300, profile written to EC register APMD at offset 0x95).

The fix:

bash
set the firmware authorization flag (firmware’s own path, not a blind write)
echo ‘_SB.PC00.LPCB.EC0.RCAP 0 0’ > /proc/acpi/call

confirm
echo ‘_SB.NPCF.CTGP’ > /proc/acpi/call; cat /proc/acpi/call # → 0x1

renegotiate NVPCF (mandatory — the cap won’t update without this)
sudo systemctl restart nvidia-powerd

nvidia-smi -q -d POWER now reports Current Power Limit = 175W (Max).

Measured on this machine (RTX 5090 Laptop, LLVM pipeline fixed 500-token LLM decode benchmark):

State Power Limit SM clock Throughput
powerd dead 95W ~800MHz ~29 tok/s
powerd fine, CTGP=0 120W ~930MHz ~33 tok/s
CTGP=1 + renegotiation 175W ~1545MHz 49–53 tok/s

Sustained load: 168–169W flat at 69°C (firmware turbo thermal target 75°C still enforced — all protections remain active).

Answers to my original questions:

  1. Budget is controlled by SBIOS/NVPCF plus an OEM authorization flag — powerd alone can’t raise it.
  2. Yes, driver-side action is possible: call the firmware’s own activation path (RCAP).
  3. The mapping exists in the DSDT: WMI profile switch → EC APMD (0x95) → RCAP() → NPCF.CTGP → TGPA = ACBT.

I’ve published the complete tool with safety checks, a systemd service for persistence across reboots, and full reverse-engineering notes (including an adaptation checklist for other Tongfang/Colorful machines) on GitHub:

Thanks to @IzzieBoopers 's 5080-Unlock_Linux for demonstrating that OEM WMI/EC flag toggling is the right approach on 50-series laptops, and to generix for the powerd documentation pointers.