Disabling ECC (Pro VRAM)
WIP: Procedure is documented for Radeon PRO W6800 in lunnova’s guide.
#vllm-rdnaand#generalreport the same ECC-on-by-default behavior on V620. Treat the two-reboot kernel-param path as community until you confirmrocm-smion your own cards. See Verification status.
Workstation/data-center Navi 21 cards (PRO V620, PRO W6800) ship with on-board ECC enabled. That costs about 7% of VRAM (~2 GB on a 32 GB card). Consumer RX 6800 / 6900 cards have no ECC and already show the full 16 GB.
| ECC | Typical rocm-smi VRAM (32 GB Pro card) |
|---|---|
| On (factory default) | ~30 700 MiB (~30 GB) |
| Off | ~32 768 MiB (full 32 GB) |
#vllm-rdna (pinned): extra ~2 GB per GPU if you disable it. Some members keep ECC on for bit-flip
protection; others disable it for long-context KV cache.
Check whether ECC is on
sudo dmesg | grep -i ecc
# "MEM ECC is active" / "GECC is enabled" → ECC is on
# "MEM ECC is not presented" → consumer card / no ECC
rocm-smi --showmeminfo vram
# ~30.7 GB total on a 32 GB Pro card → ECC is eating the rest
Linux: amdgpu.ras_enable=0 (two reboots)
There is no Radeon control-panel toggle on Linux. The community path (lunnova, 6.x kernels) is:
- Add the kernel parameter
amdgpu.ras_enable=0. - Reboot twice, leaving the parameter in place both times.
GRUB (Ubuntu / Debian / Fedora):
# Ubuntu/Debian: edit GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub
# Fedora: sudo grubby --update-kernel=ALL --args="amdgpu.ras_enable=0"
sudo nano /etc/default/grub
# …add amdgpu.ras_enable=0 to GRUB_CMDLINE_LINUX_DEFAULT, then:
sudo update-grub # Ubuntu/Debian
# Fedora uses grubby above instead of update-grub
sudo reboot
After the first reboot, dmesg should mention that GECC will be disabled on the next boot:
GECC will be disabled in next boot cycle if set amdgpu_ras_enable and/or amdgpu_ras_mask to 0x0
Reboot again. On the second boot you want:
amdgpu: GECC is disabled
and rocm-smi --showmeminfo vram should report the full ~32 768 MiB.
To turn ECC back on: remove the kernel parameter and reboot twice.
An older
amdgpu-no-ecc.patchis not needed on 6.x kernels. Keep it only if you are still on 5.x and the two-reboot path does nothing — details in lunnova’s article.
Notes
- Consumer gfx1030 (RX 6800 / 6800 XT / 6900 XT / 6950 XT): skip this page.
- Secure Boot / signed kernels: same caveats as power tuning — unsigned module overrides will not load.
- This does not replace the 120 W power-cap patch. ECC and the VBIOS power floor are independent.