Skip to content

Commit f83b6bb

Browse files
committed
Re-verify the two flagged research figures; one was wrong
The substrate research carried two figures it could not check, because skia.org, blog.chromium.org, source.android.com and vulkan.org were all unreachable when it was written. They still are, so this used search plus the primary sources that do resolve. Graphite MotionMark: confirmed as written. ~15% on MotionMark 1.3, measured on Apple Silicon, not Android. The caveat the draft already carried -- Android's HWUI uses Ganesh, not Graphite -- is the load-bearing part and is unchanged. Snapdragon 680 Impeller-vs-Skia: partly wrong. One sentence had merged two different benchmarks. The ~22% average frame time and ~60% variance figures do belong to that device; the "4.05 ms vs 2.81 ms GPU raster" numbers come from a separate republished benchmark with no Snapdragon attribution. Each figure is defensible; pairing them was not. Re-verification also found a primary source the draft did not have, and a better one than either. flutter/flutter#192147 measures Impeller's GLES backend on an Adreno 610 -- a floor-segment GPU -- at 12.0 ms/frame against Skia GL's 6.7 ms, 2-3x slower, with 10% of frames late against 0%. The diagnosed cause is per-draw CPU overhead: the backend re-issues all GL state per draw with minimal caching, 25-35 GL calls per draw command, and a scene with no blur, no shadows and no platform views spends 6.5 ms per frame inside EncodeCommandsInReactor on the raster thread. That is the strongest evidence available for section 5's substrate choice, and it states directly what our own draw-call measurement implies: on a floor-segment device the number of draws costs you, not their complexity. MolecularStructure issues 5,120 drawPath calls per frame, 154,000 per second at 30 fps. Co-Authored-By: Claude Opus 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01LYN6fepNP94KixTL6tB4a2
1 parent 2f0fa9c commit f83b6bb

2 files changed

Lines changed: 64 additions & 11 deletions

File tree

‎docs/SPEC.md‎

Lines changed: 14 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1130,5 +1130,17 @@ the fallback.
11301130
Manim and LaTeX install and tens of minutes per scene; that belongs in a nightly job and is
11311131
run by hand until one exists.
11321132
- **Bundle integrity and player UX beyond scrubbing.**
1133-
- **Two secondary figures** in the substrate research want re-verifying; several primary sources
1134-
were unreachable behind an egress proxy when it was written.
1133+
- ~~**Two secondary figures** in the substrate research want re-verifying.~~ **Done, and one was
1134+
wrong.** The Graphite MotionMark number is confirmed as written (~15% on MotionMark 1.3,
1135+
Apple Silicon, not Android). The Snapdragon 680 Impeller-vs-Skia entry had **merged two
1136+
different benchmarks**: the ~22% frame-time and ~60% variance figures belong to that device,
1137+
the "4.05 ms vs 2.81 ms GPU raster" numbers belong to another with no Snapdragon attribution.
1138+
Corrected in `docs/research/android-rendering-substrate.md`.
1139+
1140+
Re-verification also turned up a primary source the draft did not have, and it is a better
1141+
one: [flutter/flutter#192147](https://github.com/flutter/flutter/issues/192147) measures
1142+
Impeller's GLES backend at **2–3× slower than Skia GL on an Adreno 610** — 12.0 ms/frame
1143+
against 6.7 ms — with the cause being per-draw CPU overhead, 25–35 redundant GL calls per
1144+
draw, not anything on the GPU. That sharpens §5's substrate choice and says plainly what
1145+
§11's draw-call measurement implies: **on a floor-segment device the number of draws costs
1146+
you, not their complexity.**

‎docs/research/android-rendering-substrate.md‎

Lines changed: 50 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -146,12 +146,39 @@ a large one on Canvas.
146146
No source gives "N Manim paths per frame on a Snapdragon 680". The honest answer is that
147147
the number is workload-shaped, and the credible anchors are:
148148

149-
- **Impeller vs Skia on a Snapdragon 680 / 4 GB device**: GPU raster averaged 4.05 ms/frame
150-
(Skia) vs 2.81 ms (Impeller), with Impeller improving average frame render time ~22% and
151-
frame-time variance ~60% *for Flutter UI workloads* (widely republished benchmark; treat
152-
as order-of-magnitude, secondary source). The useful reading is that **GPU raster on a
153-
budget phone for a full-screen UI is single-digit milliseconds** — the GPU is not our
154-
problem, the CPU-side path work is.
149+
- **Impeller vs Skia on a Snapdragon 680 / 4 GB device**: Impeller improved average frame
150+
render time ~22% and frame-time variance ~60% *for Flutter UI workloads*.
151+
**Correction (re-verified 2026-09):** the "4.05 ms vs 2.81 ms GPU raster" figures were
152+
previously attributed to this same device. They are not from it — they come from a
153+
*separate* republished benchmark (~30% GPU raster reduction, ~7.71 ms vs ~6.57 ms total
154+
frame time) with no Snapdragon 680 attribution. Two benchmarks had been merged into one
155+
sentence. The percentages and the millisecond figures are each defensible; pairing them
156+
was not.
157+
158+
The reading that survives either way is that **GPU raster on a budget phone for a
159+
full-screen UI is single-digit milliseconds** — the GPU is not our problem, the CPU-side
160+
path work is. What follows makes that concrete with a primary source.
161+
162+
- **Adreno 610 (Snapdragon 662/665), a floor-segment GPU — the CPU-side cost, measured**
163+
([flutter/flutter#192147](https://github.com/flutter/flutter/issues/192147), open):
164+
165+
| Backend | Total | Raster | Late frames |
166+
|---|---|---|---|
167+
| Skia GL | 6.7 ms | 3.4 ms | 0% |
168+
| Impeller GLES | 12.0 ms | 8.5 ms | 10% |
169+
170+
Impeller is **2–3× slower than Skia** here, and the diagnosed cause is per-draw CPU
171+
overhead rather than anything on the GPU: the GLES backend re-issues all GL state per
172+
draw with minimal caching — **25–35 GL calls per draw command** — and a scene with no
173+
blur, no shadows and no platform views spends 6.5 ms/frame inside `EncodeCommandsInReactor`
174+
on the raster thread.
175+
176+
This is the strongest available evidence for §5's conclusion, and it is evidence the
177+
earlier draft did not have. It also says the thing directly: **on a floor-segment device,
178+
what costs you is the number of draws, not their complexity.** Our own measurement agrees —
179+
MolecularStructure issues 5,120 `drawPath` calls per frame, 154,000 per second at 30 fps
180+
(SPEC §11). A second data point on per-draw cost, from the other direction: Skia's reported
181+
text ceiling of ~10,000 `DrawText` calls/sec on desktop-class hardware.
155182
- Skia's own text throughput ceiling in one reported case was ~10,000 `DrawText` calls/sec
156183
([SkiaSharp#2771](https://github.com/mono/SkiaSharp/issues/2771)) — i.e. roughly
157184
300 calls per frame at 30 fps before text alone eats the budget, on desktop-class
@@ -622,13 +649,27 @@ Secondary (used only for order-of-magnitude, flagged inline): republished Impell
622649
Snapdragon 680 benchmarks; React Native Skia batching write-ups; [apilevels.com](https://apilevels.com/)
623650
cumulative reach figures.
624651

652+
### Re-verification, 2026-09
653+
654+
The two figures flagged below were re-checked. `skia.org`, `blog.chromium.org`,
655+
`source.android.com` and `vulkan.org` remain unreachable from this environment, so the check
656+
used search and the retrievable primary sources.
657+
658+
- **Graphite MotionMark: confirmed as written.** ~15% MotionMark 1.3 improvement, measured on
659+
Apple Silicon (a MacBook Pro M3), not Android. The caveat the draft already carried —
660+
Android's HWUI uses Ganesh, not Graphite — is the load-bearing part and is unchanged.
661+
- **Snapdragon 680 Impeller-vs-Skia: partly wrong, corrected above.** The percentages belong
662+
to that device; the millisecond figures belong to a different benchmark and had been merged
663+
into the same sentence. The conclusion drawn from them did not depend on the pairing, so
664+
§1.4 and §5 stand — and flutter/flutter#192147 now supports them with a primary source and
665+
a harsher number.
666+
625667
### Sources that could not be retrieved
626668

627669
The research network blocked `skia.org`, `skia.googlesource.com`, `source.android.com`,
628670
`blog.chromium.org`, `blog.google`, `groups.google.com` (skia-discuss), `rive.app`,
629671
`news.ycombinator.com`, `vulkan.org` (Ian Elliott's *Vulkan on Android* Vulkanised 2025/2026
630672
decks) and `arxiv.org`. Claims that would normally be sourced there were either obtained
631673
from the GitHub mirror of the Skia repo, from `developer.android.com`, or from search-engine
632-
summaries and are flagged as secondary above. **Two figures worth re-verifying against the
633-
original when those domains are reachable:** the Graphite Motionmark number, and the
634-
Snapdragon 680 Impeller-vs-Skia frame times.
674+
summaries and are flagged as secondary above. The two figures once flagged here have been
675+
re-checked; see *Re-verification, 2026-09* above.

0 commit comments

Comments
 (0)