GPU

The PS5's GPU is an RDNA2 part inside the Oberon APU, PCI ID 1002:13fb on the original PS5 and the Slim. The PS5 Pro has a different GPU, see PS5 Pro. This page covers what the kernel does with it and the settings you can change.

Status

• amdgpu drives the GPU and the HDMI output on every kernel.
• Hardware video decode and encode (VCN) work; HEVC encoding only with I-frames. See Hardware video (VCN).

Required modprobe options for a working display

Every working image ships /etc/modprobe.d/ps5-amdgpu.conf with:

/etc/modprobe.d/ps5-amdgpu.conf

options amdgpu dpm=0 gpu_recovery=0

Without this file the GPU comes up without a picture. When you bring up a distro of your own, put it in place before you look at any other display problem.

Clocks and boost

The console's firmware starts the CPU at 3.2 GHz and the GPU below the 2.23 GHz the PS5 OS runs it at. How the full clocks are reached depends on the kernel, see Kernel setup:

Kernel	CPU and GPU clocks	What to do
rmuxnet's kernel	ps5-cpufreq and ps5-gpufreq scale them with the load.	Nothing for the GPU. CPU boost (3.3 to 3.5 GHz) is off until you turn it on, below.
PS5 Bazzite Deck	The upstream GPU governor; CPU boost through ps5-perf.	On by default.
Upstream (current packages)	The upstream GPU governor (ps5_gpufreq.temp_limit, see PS5 Pro); the CPU stays at 3.2 GHz until ps5_control --boost on.	Install ps5-linux-tools, below.
Older upstream kernels	Fixed at the firmware's clocks until ps5_control --boost on.	Install ps5-linux-tools, below.

The upstream GPU governor is a devfreq driver: 500 to 2,230 MHz on the original PS5 and the Slim, 500 to 2,350 MHz on the Pro, with a thermal limit of 85 °C by default (ps5_gpufreq.temp_limit=<50-95> in cmdline.txt). Read the current clock from cur_freq under its directory in /sys/class/devfreq/.

rmuxnet's kernel: cpufreq and devfreq

ps5-cpufreq drives the CPU P-states (800 MHz to 3.2 GHz, schedutil governor) through the SMU. The three boost states are off at boot. Turn them on:

echo 1 | sudo tee /sys/devices/system/cpu/cpufreq/boost

ps5-gpufreq is a devfreq driver at /sys/class/devfreq/ps5-gpufreq/ with the simple_ondemand governor: 400 to 2,230 MHz on the original PS5 and the Slim, 500 to 2,350 MHz on the PS5 Pro, in 400 MHz steps up to 2,000 MHz and then the maximum. cur_freq shows the clock; min_freq and max_freq pin it, and governor set to performance holds the maximum.

Don't use ps5_control --boost on this kernel

The drivers already own the clocks. ps5_control --boost on pins them at maximum until a full restart and doesn't improve frame rates. Use the cpufreq boost switch above instead.

Older kernels: ps5-linux-tools

On kernels without a clock driver, ps5_control from ps5-linux-tools raises the clocks through the power controller:

cd ps5-linux-tools
sudo ./ps5_control --fan on
sudo ./ps5_control --boost on

Turn the fan on whenever you turn the boost on, as the PS5 OS does. --fan max runs the fan harder than on if you want to hold the boost clocks under a heavy load:

sudo ./ps5_control --fan max

To apply both at every boot:

sudo ./install.sh

This installs ps5fan.service and ps5boost.service. On a kernel with the fan curve (see Fan control) --fan isn't needed.

40 CU unlock

The die has 40 compute units (CUs), but every console leaves the factory with 36 or 38 of them enabled; the others are switched off in the silicon, usually because they failed a test, sometimes only to even out production. Which case you have is per console, which is also why identical consoles benchmark differently. The PS5 kernel patch can switch the disabled CUs back on.

Kernel parameter: amdgpu.ps5_cu_unlock=N

Value	Behaviour
0	Off (default)
1	Probe
2	Unlock SE0/SH0 only
3	Unlock all harvested CUs
4	Probe all

Without the parameter, dmesg shows the hardware default: 2 shader engines × 2 shader arrays × 10 CUs, of which 36 are active here:

amdgpu: SE 2, SH per SE 2, CU per SH 10, active_cu_number 36

Original PS5 and Slim only

Don't use this on a PS5 Pro: all of its 60 compute units are already active.

It is a lottery. On many consoles all 40 CUs work and the GPU gets faster; on others the extra units produce green dots, a purple tint, flickering or crashes. Try it, and turn it off again at the first sign of trouble.

To unlock all CUs, add this to cmdline.txt on the FAT32 USB partition:

cmdline.txt

amdgpu.ps5_cu_unlock=3

Or let this one-liner do it from Linux (it finds cmdline.txt in /boot/efi or /efi):

E=$(for d in /boot/efi /efi; do sudo test -f $d/cmdline.txt && echo $d && break; done); sudo sed -i -e 's/amdgpu\.ps5_cu_unlock=[0-9]*/amdgpu.ps5_cu_unlock=3/' -e '/ps5_cu_unlock/!s/[[:space:]]*$/ amdgpu.ps5_cu_unlock=3/' $E/cmdline.txt && sudo cat $E/cmdline.txt

Reboot and run the loader, then check the result. It should show 40 (or 36 without the unlock):

sudo dmesg | grep active_cu_number

To turn it off again:

E=$(for d in /boot/efi /efi; do sudo test -f $d/cmdline.txt && echo $d && break; done); sudo sed -i 's/ *amdgpu\.ps5_cu_unlock=[0-9]*//' $E/cmdline.txt && sudo cat $E/cmdline.txt

The patch enables the CUs in two places, the CC register and the SPI; changing only one of them makes the units show up without any work being sent to them. Source: ps5-linux-patches commit 5b0dc56.

Both shader engines need the same CU count

The GPU only uses extra CUs when its two shader engines (SEs) end up with the same number active. 38 CUs, with one SE at 20 and the other at 18, performs like 36. If a game glitches after the unlock, try RADV_DEBUG=nohiz %command% in its Steam launch options first, then 38 CUs instead of 40.

Disabled CUs come in pairs, and a pair can be bad

CUs are disabled two at a time, so a console has 36, 38 or 40 usable CUs and never an odd number. Heavy artefacts with all 40 on usually mean one of your disabled pairs is defective rather than merely switched off. Finding which one means testing 38 CUs with different pairs masked, and nothing about the console predicts the answer.

Runtime unlock with UMR (no kernel rebuild)

UMR, AMD's GPU register debugger, can switch the extra CUs on while the system runs, which is a quick way to test your console before you change cmdline.txt. It is packaged in the AUR for Arch and as an RPM for Fedora and Bazzite; on Debian-based distros, build it from source.

UMR's topology view can keep showing the CUs as disabled while the runtime unlock is in effect. Run a benchmark before and after and compare the scores rather than relying on the topology output.

Performance compared to the BC-250

The AMD BC-250, a mining board built on the same APU, is the usual reference. At the same CU count and clock the BC-250 scores a few percent higher in Superposition than the PS5, with or without the 40 CU unlock, and the PS5's scores fall short of what scaling the BC-250's results by CU count would predict. The cause isn't known; CU detection isn't the problem, and the leading theory is that the PS5's clock doesn't hold steady under the benchmark. Without the boost the two platforms score closer together.

A dead end: lds_size

Setting adev->gfx.cu_info.lds_size = 128; in amdgpu_device.c has been suggested as a fix. A rebuild with it produced a black screen and no amdgpu output. It is listed here so nobody tries it again, not as a fix.

VRAM

Tools show a default allocation of 512 MB, but VRAM is dynamic: the driver takes more as needed. With the default split and about 14 GB usable, the ceiling works out at roughly 7 GB; the formula and how to raise it are below.

To set a fixed minimum, edit vram.txt on the FAT32 USB partition. The loader reads the value as hexadecimal bytes, not decimal megabytes (strtoull(buf, NULL, 16) in ps5-linux-loader/source/loader.c). A decimal megabyte value gives the GPU almost nothing, and the boot ends in a black screen.

The default (512 MB) is:

vram.txt

20000000

Size	vram.txt value
512 MB (default)	20000000
1 GB	40000000
2 GB	80000000
4 GB	100000000
6 GB (recommended for gaming)	180000000
8 GB	200000000

What vram.txt actually controls

Every VRAM split is dynamic; 512 MB is not a special "fully dynamic" mode. The vram.txt value only sets the minimum amount of reserved VRAM, not the maximum. The real ceiling for dynamic allocation is:

(vram.txt split) + (remaining RAM not in that split) / 2

That is why a game that wants a little more than 8 GB of GPU memory crashes with amdgpu_vm_validate() failed / Not enough memory for command submission while free still shows plenty of RAM.

The real fix: raise the GTT ceiling with the ttm.pages_limit kernel parameter instead of (or as well as) raising vram.txt:

ttm.pages_limit = (desired GTT size in GB) × 1024 × 1024 / 4

For a console used for gaming, 6 GB in vram.txt (180000000) is the usual choice: the GPU then has 6 GB it never has to compete for, and the ceiling becomes 6 + (14 - 6) / 2 = about 10 GB. On a server or a desktop that doesn't game, leave the default: the 6 GB are then locked away from everything else. Raising ttm.pages_limit gives the same ceiling without locking anything, and is the tool for games that need more than that.

Example for a 10 GB GTT ceiling:

cmdline.txt

ttm.pages_limit=2621440

Performance tips

• Micro-stuttering? Remove idle=halt from cmdline.txt.
• Glitches in one game? Add RADV_DEBUG=nohiz %command% to that game's Steam launch options. Keep it per game: it disables a depth optimisation and costs performance everywhere, and only a few games need it.
• The AMD BC-250 documentation covers the same GPU and much of it carries over. Its section on VRAM and GTT doesn't match how the split works on the PS5; use the VRAM section above for that.

Video engine (VCN)

The GPU's video engine (VCN 2.0, IP 2.0.3, driver vcn_v2_0) decodes and encodes video in hardware in every current PS5 kernel. This page lists what it handles, how it was made to work, and measured results.

What	Status
Decode H.264, HEVC Main/Main 10, VP9 Profile 0/2, VC-1	✅ Working, bit-exact with software decoding
Decode MPEG-2	✅ Working, very close but not bit-exact
Encode H.264	✅ Working
Encode HEVC, I-frames only	✅ Working
Encode HEVC, normal GOP with P-frames	❌ Broken: corrupt colours at every tested bitrate
Decode H.264 High 10, AV1, VP8, Theora, MPEG-4 Part 2	❌ Not supported by the hardware

Frame sizes that decode: H.264 up to 4096×2304, HEVC and VP9 up to 8192×4352. The engine advertises H.264 at 4096×4096 but fails on it.

How it works

The video engine runs AMD's Navi 10 VCN firmware with one change: as shipped, that firmware waits at boot for a handshake from the SMU, which never comes on the PS5, so the patched copy doesn't wait. It lives in the patch set repository as firmware/amdgpu/ps5_vcn.bin and is compiled into the kernel.

On 1002:13fb, the driver also:

• turns off VCN doorbells (the rings use direct register writes instead),
• resets the VCPU before starting and after a failed stop,
• clears stale ring state before starting.

Getting it

Which kernels have it

Every current kernel: the upstream packages, rmuxnet's release and PS5 Bazzite Deck. On an old kernel, update or build from the current patch set, see Kernel setup.

The repository's .config sets:

CONFIG_EXTRA_FIRMWARE="amdgpu/ps5_vcn.bin"
CONFIG_EXTRA_FIRMWARE_DIR="../ps5-linux-patches/firmware"

That path is relative to the kernel source tree, so linux and ps5-linux-patches must be cloned side by side, exactly as in the build steps.

Check that it works

Install libva-utils (Arch) or vainfo (Debian/Ubuntu), then run:

vainfo

It lists H.264, HEVC, VP9, MPEG-2 and VC-1 decoding, plus H.264 and HEVC encoding, when the engine is up.

Using it with FFmpeg

Decode:

ffmpeg -hwaccel vaapi -i input.mkv -f null -

Transcode to H.264:

ffmpeg -hwaccel vaapi -hwaccel_output_format vaapi -i input.mkv -c:v h264_vaapi -b:v 12M output.mkv

HEVC comes out clean only when every frame is an I-frame; with P-frames the colours are wrong. Force I-frames with -g 1 -bf 0:

ffmpeg -hwaccel vaapi -hwaccel_output_format vaapi -i input.mkv -c:v hevc_vaapi -g 1 -bf 0 -b:v 25M output.mkv

An all-I-frame stream needs a much higher bitrate for the same quality, so when file size matters, h264_vaapi is the better encoder.

Browsers and streaming

Firefox decodes through VAAPI by itself: H.264 at 720p, 1080p and 4K, HEVC and HEVC Main 10 at 1080p, and VP9 at 1080p all play without dropped frames. about:support, under Media, shows which codecs are decoded in hardware. Chrome needs flags, see Browser video.

Moonlight streaming from the PS5 works through Sunshine with H.264 VCN encoding, see Game streaming.

Results

The short version: decoding in hardware costs around 1% of one CPU core, where software decoding of the heaviest streams (HEVC Main 10 at 4K) costs most of a core, and H.264 hardware transcoding runs at full frame rate at about 1 to 2.5% of a core.

CPU figures are FFmpeg's CPU use between seconds 2 and 9.5 of each run, where 100% is one logical CPU.

Decoding: CPU use

Codec	Size	Software	VCN	VCN frame check
H.264 Baseline	720p	4.35%	0.54%	32 exact frames
H.264 Main	1080p	17.19%	0.60%	32 exact frames
H.264 High	1080p	17.66%	0.67%	32 exact frames
HEVC Main	1080p	29.77%	0.54%	32 exact frames
HEVC Main 10	1080p	35.12%	0.47%	32 exact frames
VP9 Profile 0	1080p	19.40%	0.60%	32 exact frames
VP9 Profile 2	1080p	23.68%	0.54%	32 exact frames
MPEG-2 Main	1080p	7.76%	0.74%	57.09 dB, not exact
VC-1 Advanced	720p	19.67%	0.87%	32 exact frames
H.264 High	4K	66.89%	1.74%	32 exact frames
HEVC Main 10	4K	83.48%	1.07%	12 exact frames
VP9 Profile 0	4K	75.12%	1.14%	12 exact frames
H.264 High 10	1080p	21.47%	n/a	Unsupported
AV1 Main	1080p	16.99%	n/a	Unsupported
Theora	1080p	8.96%	n/a	Unsupported
VP8	1080p	14.18%	n/a	Unsupported
MPEG-4 Part 2 Simple	1080p	8.03%	n/a	Unsupported

Decoding: frames per second

Codec	Size	Software	VCN
H.264 Baseline	720p	29.7	29.4
H.264 Main	1080p	29.7	29.4
H.264 High	1080p	29.7	29.4
HEVC Main	1080p	29.7	29.4
HEVC Main 10	1080p	29.7	29.4
VP9 Profile 0	1080p	29.6	29.4
VP9 Profile 2	1080p	29.4	29.4
MPEG-2 Main	1080p	28.8	28.5
VC-1 Advanced	720p	59.4	58.8
H.264 High	4K	29.4	29.4
HEVC Main 10	4K	30.0	30.0
VP9 Profile 0	4K	29.7	29.7
H.264 High 10	1080p	29.4	n/a
AV1 Main	1080p	30.9	n/a
Theora	1080p	21.9	n/a
VP8	1080p	29.7	n/a
MPEG-4 Part 2 Simple	1080p	29.7	n/a

Transcoding: CPU use

Codec	Size	Target (Mbps)	Software	VCN	VCN output
H.264 Baseline	1080p	12	90.12%	1.00%	Checked
H.264 Main	1080p	12	102.15%	1.00%	Checked
H.264 High	1080p	12	103.02%	1.00%	Checked
HEVC Main	1080p	12	292.23%	0.94%	Corrupt colours
HEVC Main 10	1080p	12	298.39%	1.00%	Corrupt colours
H.264	1080p	4	87.78%	0.94%	Checked
HEVC	1080p	4	322.99%	0.80%	Corrupt colours
H.264	1080p	8	95.00%	0.94%	Checked
HEVC	1080p	8	309.06%	0.94%	Corrupt colours
H.264	1080p	25	118.16%	1.20%	Checked
HEVC	1080p	25	284.08%	1.20%	Corrupt colours
H.264	4K	25	315.14%	2.27%	Checked
HEVC	4K	25	419.63%	1.87%	Corrupt colours
H.264	4K	50	356.79%	2.41%	Checked
HEVC	4K	50	367.30%	2.14%	Corrupt colours

Transcoding: frames per second

Codec	Size	Target (Mbps)	Software	VCN
H.264 Baseline	1080p	12	29.4	29.4
H.264 Main	1080p	12	29.4	29.4
H.264 High	1080p	12	29.4	29.4
HEVC Main	1080p	12	18.4	29.4
HEVC Main 10	1080p	12	15.4	29.4
H.264	1080p	4	29.4	29.4
HEVC	1080p	4	26.8	29.4
H.264	1080p	8	29.4	29.4
HEVC	1080p	8	21.4	29.4
H.264	1080p	25	29.4	29.4
HEVC	1080p	25	14.4	29.4
H.264	4K	25	29.4	29.4
HEVC	4K	25	8.0	29.4
H.264	4K	50	29.4	29.4
HEVC	4K	50	6.4	29.4

Encoding quality: PSNR (dB)

Higher is better. HEVC VCN results use a normal GOP with P-frames, which is why they are so low (corrupt colours).

Codec	Size	Target (Mbps)	Software	VCN
H.264 Baseline	1080p	12	51.16	48.19
H.264 Main	1080p	12	52.54	49.90
H.264 High	1080p	12	52.20	49.90
HEVC Main	1080p	12	51.76	15.46
HEVC Main 10	1080p	12	52.14	15.31
H.264	1080p	4	41.66	39.68
HEVC	1080p	4	43.25	15.85
H.264	1080p	8	47.54	45.12
HEVC	1080p	8	47.98	15.61
H.264	1080p	25	60.88	59.90
HEVC	1080p	25	60.20	15.42
H.264	4K	25	45.72	42.68
HEVC	4K	25	47.32	15.03
H.264	4K	50	53.24	51.30
HEVC	4K	50	53.84	15.01

Output file bitrate (Mbps)

Codec	Size	Target	Software	VCN
H.264	1080p	4	4.08	4.01
HEVC	1080p	4	4.10	4.10
H.264	1080p	8	8.17	8.00
HEVC	1080p	8	8.20	7.96
H.264	1080p	25	24.81	24.86
HEVC	1080p	25	25.62	24.90
H.264	4K	25	25.39	24.90
HEVC	4K	25	25.59	25.05
H.264	4K	50	50.67	49.78
HEVC	4K	50	51.24	49.81

How to read the results

Term	Meaning
CPU scale	100% = one logical CPU. FFmpeg CPU use between seconds 2 and 9.5.
FPS	Paced runs, including startup. Not maximum throughput.
HEVC VCN encoding	Normal GOP with P-frames. Corrupt colours at every tested bitrate; I-frame-only output is clean.
PSNR	Measured over the first 60 encoded frames. Higher means less pixel error.
File bitrate	Includes container overhead.
Repeats	Decoding and 12 Mbps tests: two runs. Other bitrates: one run.
Exact frames	Decoded pixels match software decoding exactly.

Limits and open problems

What every PS5 Linux user runs into sooner or later, whatever the distro.

Limit	What happens	Status
GPU memory ceiling	A game that needs more than the GTT ceiling crashes with Not enough memory for command submission. Raise ttm.pages_limit rather than a fixed vram.txt carve-out, see VRAM.	Documented fix
40 CU unlock	More GPU power on units whose disabled pairs are good; artefacts, a purple tint or crashes on units where one isn't, see 40 CU unlock. Don't use it on a Pro.	Silicon lottery
HEVC encoding	The video engine corrupts colours with P-frames (red screen when streaming or recording). Use H.264, or HEVC with I-frames only, see VCN.	Known not to work
GPU fault at session switches	Switching between a Gamescope session and a desktop can trigger a GPU page fault and ring gfx_0.0.0 timeout, most often at 4K120 on some TVs. It is on the compositor side; keep the desktop at 4K60 and switch sessions at the boot instead, see Boot to the desktop. On a PS5 Pro the fault isn't recovered (GPU Recovery Failed) and the screen stays black until a restart. If you hit it, save the coredump before restarting: sudo cp /sys/class/drm/card0/device/devcoredump/data ~/gpu-coredump.txt.	Open
Failed to pin framebuffer: -ENOMEM	Reported under memory pressure with a desktop and Steam running. No confirmed fix; more zram helped in one report, see Memory and swap.	Open
Graphical glitches in one game	Glitches in a single game. Add RADV_DEBUG=nohiz %command% to that game's launch options, see Performance tips.	Workaround
No AV1 decode, no GPU load sensor on the base PS5	AV1 video falls back to the CPU; KDE's system monitor shows no GPU usage, see Browser video.	Hardware limits