panvk_faux123_v1.1.3 vs panvk_faux123_pixel7_v0.11.0

On Mali-G715 MC7 (Google Pixel 9). The first driver named is compared against the second, which is the baseline.

Stockpanvk_faux123_pixel7_v0.11.0PanVK-kbase RC3 (Mesa 26.3.0-devel (git-5a07217f03))

Custompanvk_faux123_v1.1.3PanVK-kbase RC3-lowmem (Mesa 26.3.0-devel (git-44ce4ed1d7))

Graphics system version (Vulkan)1.4.363 to 1.4.363

Optional features (extensions)189 to 189+0 -0

Features that pass their check (validated)50 to 6718 pass only on panvk_faux123_v1.1.3 1 pass only on panvk_faux123_pixel7_v0.11.0

Speed test scenes (benchmark)4 faster 1 slower27% faster overall (geometric mean)

Games each driver could run (an estimate from its feature list): green likely, yellow maybe, orange limited, red no

StockDX92002+DX102006+DX112009+DX122015+Switch

CustomDX92002+DX102006+DX112009+DX122015+Switch

compute_nbody8% faster
hdr_2160p197% faster
oit_linked_lists3% faster
pipeline_barriers_2160p2% slower
subpasses4% faster

Custom driver against stock, per scene. Bars to the right are faster, to the left slower.

panvk_faux123_pixel7_v0.11.0 details panvk_faux123_v1.1.3 details

Feature checks that differ (19)

Feature	panvk_faux123_pixel7_v0.11.0	panvk_faux123_v1.1.3	Change
textureCompressionBC VkPhysicalDeviceFeatures	pass	fail	lost
depthClamp VkPhysicalDeviceFeatures	untested	pass	gained
largePoints VkPhysicalDeviceFeatures	untested	pass	gained
compat.helper_loop_terminator core	untested	pass	gained
compat.nested_map core	untested	pass	gained
core.blend_advertised_formats core	untested	pass	gained
core.compute_big_stack core	untested	pass	gained
core.host_coherent_readback core	untested	pass	gained
core.mrt_partial_kill core	untested	pass	gained
core.ubo_update_buffer core	untested	pass	gained
core.ubo_written_after_record core	untested	pass	gained
core.varying_subset core	untested	pass	gained
mem.image_padding core	untested	pass	gained
perf.fb_local_barrier_large_draws core	untested	pass	gained
perf.fb_local_barrier_overdraw core	untested	pass	gained
stage.geometry_max_output core	untested	pass	gained
stage.geometry_point_size core	untested	pass	gained
stress.scratch_per_batch core	untested	pass	gained
sync.first_submit_clear_then_compute core	untested	pass	gained

Speed test scenes (benchmark)

Scene	panvk_faux123_pixel7_v0.11.0 ms	panvk_faux123_v1.1.3 ms	Difference	Starts	GPU clock, MHz	Picture vs stock
compute_nbody	11.24 typical 11.12	10.40 typical 10.22	8% faster	3 of 3 vs 3 of 3	? vs ? clock unknown	no large difference 0.0 no small difference worst 0.0 64x36
hdr_2160p	59.06 typical 58.20	19.86 typical 20.11	197% faster	3 of 3 vs 3 of 3	? vs ? clock unknown	no large difference 0.0 no small difference worst 0.7 64x36
oit_linked_lists	18.18 typical 18.02	17.71 typical 17.66	3% faster	3 of 3 vs 3 of 3	? vs ? clock unknown	no large difference 0.0 no small difference worst 0.0 64x36
pipeline_barriers	7.73	n/a	too light, see pipeline_barriers_2160p	3 of 3 vs 3 of 3	? vs ? clock unknown	no large difference 0.0 no small difference worst 0.0 64x36
pipeline_barriers_2160p	13.30 typical 13.33	13.63 typical 13.56	2% slower	3 of 3 vs 3 of 3	? vs ? clock unknown	no large difference 0.0 no small difference worst 0.0 64x36
subpasses	8.66 typical 8.57	8.34 typical 8.34	4% faster	3 of 3 vs 3 of 3	? vs ? clock unknown	not checked

Time to draw one frame, per scene: the average (mean) frame time of the scene's best round, which is how many frames the driver delivered; lower is faster. "Typical" beside it is the median frame, which can sit well under the average when a driver's frames come in two speeds. A pair of reports made before per-round averages were recorded is compared on the median. A scene whose rounds disagree by more than 15% is left out. A scene marked "paced" was held at that mean frame time by the driver or the display queue, together with at least one other scene of the same run; its time says how the frames were spaced, not how fast the driver rendered, so it is not compared. A scene marked "too light" took under 8 ms at the standard size (1920 by 1080) on both drivers, where the screen's frame queue spaces the frames more than the chip does; its heavier run is the one compared: the row ending in _2160p is the scene drawn at 3840 by 2160, and compute_nbody_2x is that scene with twice the particles. "Faster" and "slower" describe panvk_faux123_v1.1.3 against panvk_faux123_pixel7_v0.11.0. "Picture vs stock" compares one fixed frame per scene as a coarse colour grid (0 is identical, above 40 is a different picture); it finds a scene that was not drawn. Where both reports carry the finer 64x36 grid, the second figure is how many of its 2,304 cells differ by more than 4 of 255, and "worst" is the largest single cell; this is what finds a small rendering error in a picture that otherwise matches. A row marked 16x9 comes from a report made before the finer grid existed. "Starts" is how many of a scene's process starts reached a measured round. A GPU clock marked "not comparable" differed by more than 5% of the device's maximum between the two timings.

Technical differences (Vulkan)

Extensions only in panvk_faux123_v1.1.3 (0)

None

Extensions only in panvk_faux123_pixel7_v0.11.0 (0)

None

Core features that differ (0)

None

Limits that differ (0)

Limit	panvk_faux123_pixel7_v0.11.0	panvk_faux123_v1.1.3

Formats supported by only one driver (0 gained, 0 lost)

Only in panvk_faux123_v1.1.3:

None

Only in panvk_faux123_pixel7_v0.11.0:

None