The A19 is faster than the A15 on every kind of GPU work measured: 1.3× to 2.2× as much per frame on a cool box, 1.04× to 1.6× after 20 minutes
The A19’s budget per 60 Hz frame at 1080p ÷ the A15’s, one row per test, the largest lead first. Filled dot: on a cool box (the one-minute budget). Ring: after 20 minutes (each budget × the share of speed that test keeps on that chip, at most 100 %). At 1× both fit the same.
- cool box
- after 20 minutes
- No single test says how much faster a chip is: the lead depends on the kind of work, and heat takes away more of it on some kinds than on others.
- The opaque-layer control is left out: it measures how little hidden work costs, not a load.
- The table below also gives the A19 ÷ A12 and A15 ÷ A12 ratios.
The numbers as a table
| Test | A19 ÷ A15, cool | after 20 min | A19 ÷ A12, cool | after 20 min | A15 ÷ A12, cool | after 20 min |
|---|---|---|---|---|---|---|
| Full-screen shader passes | 2.2× | 1.6× | 3.7× | 2.8× | 1.7× | 1.7× |
| Blended full-screen layers | 1.4× | 1.1× | 2.0× | 1.6× | 1.4× | 1.4× |
| Alpha-tested layers | 1.3× | 1.04× | 1.8× | 1.4× | 1.4× | 1.3× |
| Sprites, instanced | 1.3× | 1.04× | 2.1× | 1.3× | 1.6× | 1.3× |
| Sprites, one draw each | 1.5× | 1.2× | 2.2× | 1.5× | 1.4× | 1.2× |
| Meshes, instanced | 1.8× | 1.5× | 2.1× | 1.9× | 1.2× | 1.2× |
| Meshes, one draw each | 1.6× | 1.5× | 2.1× | 1.7× | 1.3× | 1.2× |
What each test is
- Full-screen shader passes
- heavy shader effects (lighting, post-processing)
- Blended full-screen layers
- see-through layers (smoke, glass)
- Alpha-tested layers
- layers with holes in them (leaves, fences)
- Sprites, instanced
- 2D sprites, drawn as one batch
- Sprites, one draw each
- 2D sprites, drawn one by one
- Meshes, instanced
- 3D models, drawn as one batch
- Meshes, one draw each
- 3D models, drawn one by one
- Opaque layers (control)
- the control: hidden layers the GPU never has to shade
At 1080p, one 60 Hz frame holds 11 full-screen shader passes on the A19: 2.2× the A15 and 3.7× the A12
Passes of the reference shader per frame at 60 Hz, higher is better: at 1080p and at native 4K (3840×2160).
- A12
- A15
- A19
- A budget is the largest load whose 99th-percentile frame time stays within one 60 Hz frame, measured off-screen: it is GPU work per frame, not frames on screen.
The numbers as a table
| Chip | Resolution | Passes per 60 Hz frame | Spread | Status |
|---|---|---|---|---|
| A12 | 1080p | 3 | all equal over 3 runs (CV 0 %) | final |
| A15 | 1080p | 5 | all equal over 3 runs (CV 0 %) | final |
| A19 | 1080p | 11 | all equal over 3 runs (CV 0 %) | final |
| A12 | 2160p (4K) | nothing fits | — | final |
| A15 | 2160p (4K) | 1 | all equal over 3 runs (CV 0 %) | final |
| A19 | 2160p (4K) | 2 | all equal over 3 runs (CV 0 %) | final |
Why it slows down
Under constant load a chip heats up, and the system slows it down to keep it cool (thermal throttling). The A12 box has a fan; the A15 and A19 boxes do not. No temperature was read: what was measured is how much each box slows under constant load, and how it recovers after rest. Rest brings it back: after a 2-minute run, 4 minutes of rest returned the A19 to within 1 % of its rested speed.
Same load for 20 minutes: each chip slows while still reporting nominal (frames per second)
The same fixed load on every box (full-screen shader passes at 4 passes, 1080p), off-screen, from a rested box. Before its first change of thermal state the A15 had lost 17 % (by 8:40) and the A19 had lost 7 % (by 1:50); the A12 lost 23 % and never left nominal.
- A12
- A15
- A19 unit 1
- A19 unit 2
- first “fair”
- first “serious”
- first “critical”
- 60 fps: above it is off-screen throughput
- Each point is the mean of a 10-second window. tvOS reports four thermal states (nominal, fair, serious, critical), not a temperature; the marks show the first second each box reported a hotter state.
- Frames per second above 60 are off-screen throughput: the work is drawn as fast as the GPU allows, never shown.
- The dashed line is the second A19 unit, on the same load: the two units track each other.
- Every 20-minute run here is a single run.
The numbers as a table
| Box | First 5 s, fps | At 1 min, fps | Last minute, fps | Last ÷ first 5 s | Thermal state | Rest at nominal before |
|---|---|---|---|---|---|---|
| A12 | 69.8 | 66.2 | 53.9 | 77 % | nominal throughout | 300 s |
| A15 | 108.4 | 89.2 | 74.0 | 68 % | serious at 8:49 | 300 s |
| A19 unit 1 | 215.0 | 201.5 | 124.3 | 58 % | serious at 1:59, critical at 5:39 | 300 s |
| A19 unit 2 | 212.4 | 199.0 | 128.4 | 60 % | fair at 1:56, serious at 2:01, critical at 5:31 | 300 s |
On the A19, rendering at 1080p and upscaling with MetalFX spatial fits 4× the shader passes of native 4K (2× on the A15)
Passes of the reference shader per 60 Hz frame when the frame is drawn at 1080p or 1440p and scaled to 3840×2160 by MetalFX, against drawing it at native 4K. Higher is better.
- A12
- A15
- A19
- The scaler’s own cost is inside each budget: a frame is the passes at the input size, one MetalFX scale to 4K and a one-pixel read-back. “Nothing fits” means not even one pass does.
- Timed alone and settled, from 1080p to 4K, the spatial scaler costs 9.2 ms on the A12 (55 % of a 60 Hz frame), 7.0 ms on the A15 (42 %) and 3.1 ms on the A19 (18 %); the temporal scaler costs 20.1 ms on the A15 (121 % of a 60 Hz frame: more than a whole frame) and 9.9 ms on the A19 (60 %), and is not supported on the A12.
- Inside a real frame, the spatial scaler from 1080p plus the frame’s fixed cost (the intercept of a straight-line fit of frame time against passes) is 10.0 ms of the 16.7 on the A12, 8.2 ms on the A15 and 3.6 ms on the A19.
- The temporal scaler was timed with still, unjittered content; its cost with real motion was not checked.
The numbers as a table
| Chip | Drawn as | Passes per 60 Hz frame | Spread | Status |
|---|---|---|---|---|
| A12 | native 4K | nothing fits | — | final |
| A12 | 1440p + spatial | nothing fits | — | final |
| A12 | 1080p + spatial | 1 | all equal over 3 runs (CV 0 %) | final |
| A12 | 1440p + temporal | not supported on this chip | — | final |
| A12 | 1080p + temporal | not supported on this chip | — | final |
| A15 | native 4K | 1 | all equal over 3 runs (CV 0 %) | final |
| A15 | 1440p + spatial | 1 | all equal over 3 runs (CV 0 %) | final |
| A15 | 1080p + spatial | 2 | all equal over 3 runs (CV 0 %) | final |
| A15 | 1440p + temporal | nothing fits | — | final |
| A15 | 1080p + temporal | nothing fits | — | final |
| A19 | native 4K | 2 | all equal over 3 runs (CV 0 %) | final |
| A19 | 1440p + spatial | 5 | 4–5 over 3 runs (CV 9.4 %) | final |
| A19 | 1080p + spatial | 8 | all equal over 3 runs (CV 0 %) | final |
| A19 | 1440p + temporal | 1 | all equal over 3 runs (CV 0 %) | final |
| A19 | 1080p + temporal | 3 | all equal over 3 runs (CV 0 %) | final |
| chip | box | scaler | input size | output size | status | unit | cost ms from rest | cost ms settled | gpu cost ms settled | with scaler ms from rest | with scaler ms settled | without ms from rest | without ms settled | share of 60hz frame settled | output megapixels per s | thermal | interruptions | n | date |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| A12 | A12 | spatial | 1920x1080 | 3840x2160 | measured | ms per frame | 9.23 | 9.22 | 9.22 | 11.18 | 11.18 | 1.96 | 1.96 | 0.55 | 742 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A12 | A12 | spatial | 2560x1440 | 3840x2160 | measured | ms per frame | 9.32 | 9.28 | 9.28 | 12.75 | 12.75 | 3.42 | 3.46 | 0.56 | 651 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A12 | A12 | temporal | 1920x1080 | 3840x2160 | unsupported | ms per frame | — | — | — | — | — | — | — | — | — | — | 0 | 1 | 2026-11-14 |
| A12 | A12 | temporal | 2560x1440 | 3840x2160 | unsupported | ms per frame | — | — | — | — | — | — | — | — | — | — | 0 | 1 | 2026-11-14 |
| A15 | A15 | spatial | 1920x1080 | 3840x2160 | measured | ms per frame | 5.73 | 6.95 | 6.95 | 6.89 | 8.46 | 1.16 | 1.52 | 0.42 | 1,032 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A15 | A15 | spatial | 2560x1440 | 3840x2160 | measured | ms per frame | 5.46 | 6.96 | 6.96 | 7.47 | 9.61 | 2.01 | 2.65 | 0.42 | 910 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A15 | A15 | temporal | 1920x1080 | 3840x2160 | measured | ms per frame | 14.56 | 20.09 | 20.1 | 15.72 | 21.61 | 1.16 | 1.52 | 1.21 | 409 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A15 | A15 | temporal | 2560x1440 | 3840x2160 | measured | ms per frame | 16.94 | 21.29 | 21.29 | 18.95 | 23.94 | 2.01 | 2.65 | 1.28 | 363 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A19 | A19 unit 1 | spatial | 1920x1080 | 3840x2160 | measured | ms per frame | 2.84 | 3.06 | 3.06 | 3.55 | 3.79 | 0.7 | 0.73 | 0.18 | 2,253 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A19 | A19 unit 1 | spatial | 2560x1440 | 3840x2160 | measured | ms per frame | 2.83 | 3.1 | 3.1 | 4.08 | 4.38 | 1.25 | 1.28 | 0.19 | 1,950 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A19 | A19 unit 1 | temporal | 1920x1080 | 3840x2160 | measured | ms per frame | 9.3 | 9.92 | 9.92 | 10 | 10.65 | 0.7 | 0.73 | 0.6 | 793 | nominal→nominal | 0 | 1 | 2026-11-14 |
| A19 | A19 unit 1 | temporal | 2560x1440 | 3840x2160 | measured | ms per frame | 10.62 | 11.41 | 11.41 | 11.87 | 12.69 | 1.25 | 1.28 | 0.69 | 672 | nominal→nominal | 0 | 1 | 2026-11-14 |
On one thread each chip runs about 1.6× as fast as the one before it (CoreMark, iterations per second)
Settled figures (after the first 30 seconds), higher is better; each panel has its own unit and its own zero-based axis. On all threads, Jolt physics steps the same scene 1.5× as fast on the A15 as on the A12, and 1.4× as fast on the A19 as on the A15.
- A12
- A15
- A19
- Core counts as each box reports them: A12: 6 cores (2 performance + 4 efficiency); A15: 5 cores (2 performance + 3 efficiency); A19: 5 cores (2 performance + 3 efficiency). All-thread figures are not per core.
- Memory bandwidth and decompression use MB = 10^6 bytes; decompression is MB/s of output, over a generated 24 MiB corpus of text, JSON, vertex data, texture and audio.
- Every CPU figure here is a single run.
The numbers as a table
| Test | A12, settled | A15, settled | A19, settled |
|---|---|---|---|
| CoreMark, 1 thread (iterations/s) | 20,743 | 33,313 | 52,273 |
| Jolt physics, 1 thread (steps/s) | 52 | 80.7 | 114 |
| Jolt physics, all threads (steps/s) | 102 | 154 | 223 |
| Memory bandwidth, 1 thread (MB/s) | 20,591 | 31,659 | 56,297 |
| Memory bandwidth, all threads (MB/s) | 19,449 | 27,198 | 52,854 |
| zstd-3 decompression, 1 thread (MB/s) | 776 | 1,041 | 1,269 |
| zstd-3 decompression, all threads (MB/s) | 2,391 | 2,866 | 4,444 |
| lz4 decompression, 1 thread (MB/s) | 2,789 | 4,120 | 5,393 |
| lz4 decompression, all threads (MB/s) | 10,367 | 11,461 | 17,781 |
All-thread physics for 20 minutes: the A12 keeps 97 % of its minute-1 speed, the A15 78 %, the A19 67 % (steps per second)
Jolt physics on every core for 20 minutes from a rested box; each point is a 10-second window. Stability (worst loop ÷ best loop): A12 91.9 %, A15 64.3 % and A19 58.4 %.
- A12
- A15
- A19
- first “serious”
- first “critical”
- The same work on every chip: Jolt’s result hash was identical in every run.
- A single 20-minute run per chip.
The numbers as a table
| Chip | First loop | Minute 1 median | Best loop | Worst loop | Stability | Last minute ÷ minute 1 | Thermal state |
|---|---|---|---|---|---|---|---|
| A12 | 100.9 | 101.7 | 102.5 | 94.2 | 91.9 % | 97 % | nominal throughout |
| A15 | 187.5 | 158.0 | 187.5 | 120.5 | 64.3 % | 78 % | serious at 4:09 |
| A19 | 274.2 | 243.9 | 274.2 | 160.1 | 58.4 % | 67 % | serious at 1:24, critical at 2:26 |
Every box lets one app use about 2,098 MB, whether it has 2.9, 3.9 or 7.9 GB of memory
MB = 2^20 bytes. The limit is the app’s footprint plus the memory the system still reports available to it, read at every step of an allocate-until-killed test.
- A12
- A15
- A19
- The memory warning arrives first: plan on less than the limit. In these runs the first warning came with 1,088–1,664 MB held.
- Metal textures, buffers and render targets count exactly as heap memory does, and memory compression does not raise the limit.
The numbers as a table
| Box | What was allocated | Held at the last step, MB | First warning, MB | 4096² RGBA8 textures | 4K RGBA16F targets | Runs |
|---|---|---|---|---|---|---|
| A12 | Metal buffers | 2,048 | 1,344 | 32 | 32 | 1 |
| A12 | heap, random-filled | 2,048 | 1,088–1,408 | 32 | 32 | 3 |
| A12 | heap, zero-filled | 2,048 | 1,664 | 32 | 32 | 1 |
| A12 | 4K rgba16Float render targets | 2,048 | 1,344 | 32 | 32 | 1 |
| A12 | Metal textures, 4096² RGBA8 | 1,988 | 1,283–1,603 | 31 | 31 | 3 |
| A15 | Metal buffers | 2,048 | 1,664 | 32 | 32 | 1 |
| A15 | heap, random-filled | 2,048 | 1,664 | 32 | 32 | 3 |
| A15 | heap, zero-filled | 2,048 | 1,664 | 32 | 32 | 1 |
| A15 | 4K rgba16Float render targets | 2,056 | 1,606 | 32 | 32 | 1 |
| A15 | Metal textures, 4096² RGBA8 | 2,000 | 1,548 | 31 | 31 | 3 |
| A19 unit 1 | heap, random-filled | 2,048 | 1,664 | 32 | 32 | 1 |
| A19 unit 1 | 4K rgba16Float render targets | 2,056 | 1,606 | 32 | 32 | 1 |
| A19 unit 1 | Metal textures, 4096² RGBA8 | 2,000 | 1,548 | 31 | 31 | 1 |
A classic deferred frame at 4K takes 3.2 ms on the A12, 2.6 on the A15 and 1.9 on the A19: at most 19 % of a 60 Hz frame
Apple’s “Rendering a scene with deferred lighting” sample, unmodified shaders, settled milliseconds per frame (lower is better); 67,202 triangles submitted per frame.
- A12
- A15
- A19
- A light frame, useful as what a classic deferred frame with many lights costs, not as a stress test. The rates behind it are far above 60 fps: off-screen throughput, not frames on screen.
- Every scene figure here is a single run.
The numbers as a table
| Chip | Resolution | From rest, ms | Settled, ms | Settled, fps (off-screen) | CPU encode, ms |
|---|---|---|---|---|---|
| A12 | 1080p | 1.37 | 1.36 | 735 | 0.42 |
| A12 | 2160p (4K) | 3.22 | 3.24 | 309 | 0.57 |
| A15 | 1080p | 0.67 | 0.88 | 1,135 | 0.26 |
| A15 | 2160p (4K) | 1.95 | 2.59 | 386 | 0.27 |
| A19 | 1080p | 0.57 | 0.62 | 1,615 | 0.15 |
| A19 | 2160p (4K) | 1.72 | 1.86 | 537 | 0.17 |
Budgets to plan with
For each chip, test and resolution: the budget at 60 Hz, the largest load that held 60 frames a second for a 40 s hold on a rested box; its spread over repeated launches; and the 20-minute budget, the budget times the share of speed that test keeps after 20 minutes on that chip. Plan against the 20-minute figure. It is measured where the 20-minute run was at that resolution and scaled from 1080p where the 1080p share was applied to another resolution. Rows that say “not measured yet” are still to come from the campaign.
How the chips compare, test by test
One chip’s budget ÷ another’s, on a cool box and after 20 minutes (each budget × the share of speed that test keeps on that chip, at most 100 %). A dash means one of the two chips has no measured budget there yet.
At 1080p
| Test | A19 ÷ A15, cool | after 20 min | A19 ÷ A12, cool | after 20 min | A15 ÷ A12, cool | after 20 min |
|---|---|---|---|---|---|---|
| Full-screen shader passes | 2.2× | 1.6× | 3.7× | 2.8× | 1.7× | 1.7× |
| Blended full-screen layers | 1.4× | 1.1× | 2.0× | 1.6× | 1.4× | 1.4× |
| Alpha-tested layers | 1.3× | 1.04× | 1.8× | 1.4× | 1.4× | 1.3× |
| Sprites, instanced | 1.3× | 1.04× | 2.1× | 1.3× | 1.6× | 1.3× |
| Sprites, one draw each | 1.5× | 1.2× | 2.2× | 1.5× | 1.4× | 1.2× |
| Meshes, instanced | 1.8× | 1.5× | 2.1× | 1.9× | 1.2× | 1.2× |
| Meshes, one draw each | 1.6× | 1.5× | 2.1× | 1.7× | 1.3× | 1.2× |
At 2160p (4K)
| Test | A19 ÷ A15, cool | after 20 min | A19 ÷ A12, cool | after 20 min | A15 ÷ A12, cool | after 20 min |
|---|---|---|---|---|---|---|
| Full-screen shader passes | 2.0× | 1.5× | — | — | — | — |
| Blended full-screen layers | 1.4× | 1.2× | 2.3× | 1.9× | 1.6× | 1.6× |
| Alpha-tested layers | 1.2× | 1.01× | 1.8× | 1.4× | 1.5× | 1.4× |
| Sprites, instanced | 1.4× | 1.1× | 2.3× | 1.5× | 1.6× | 1.3× |
| Sprites, one draw each | 1.4× | 1.1× | 2.3× | 1.6× | 1.7× | 1.4× |
| Meshes, instanced | 1.7× | 1.4× | 2.3× | 2.0× | 1.4× | 1.4× |
| Meshes, one draw each | 1.6× | 1.5× | 2.2× | 1.8× | 1.3× | 1.2× |
Full-screen shader passes
full-screen passes of the reference shader
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 3 passes | all equal over 3 runs (CV 0 %) | 2 passesmeasured | final, 2026-11-15 |
| A12 | 1440p | 2 passes | all equal over 3 runs (CV 0 %) | 1 passscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | nothing fits | — | — | final, 2026-11-15 |
| A15 | 1080p | 5 passes | all equal over 3 runs (CV 0 %) | 4 passesmeasured | final, 2026-11-15 |
| A15 | 1440p | 3 passes | all equal over 3 runs (CV 0 %) | 2 passesscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 1 pass | all equal over 3 runs (CV 0 %) | 0 passesscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 11 passes | all equal over 3 runs (CV 0 %) | 6 passesmeasured | final, 2026-11-15 |
| A19 | 1440p | 6 passes | all equal over 3 runs (CV 0 %) | 3 passesscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 2 passes | all equal over 3 runs (CV 0 %) | 1 passscaled from 1080p | final, 2026-11-15 |
Blended full-screen layers
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 75 layers | 72–76 over 3 runs (CV 2.3 %) | 58 layersmeasured | final, 2026-11-15 |
| A12 | 1440p | 41 layers | all equal over 3 runs (CV 0 %) | 32 layersscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 18 layers | 18–19 over 3 runs (CV 2.6 %) | 14 layersscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 108 layers | all equal over 3 runs (CV 0 %) | 84 layersmeasured | final, 2026-11-15 |
| A15 | 1440p | 64 layers | 63–64 over 3 runs (CV 0.7 %) | 50 layersscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 29 layers | all equal over 3 runs (CV 0 %) | 22 layersscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 150 layers | 150–152 over 3 runs (CV 0.6 %) | 95 layersmeasured | final, 2026-11-15 |
| A19 | 1440p | 86 layers | 86–87 over 3 runs (CV 0.6 %) | 54 layersscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 42 layers | 40–42 over 3 runs (CV 2.2 %) | 26 layersscaled from 1080p | final, 2026-11-15 |
Opaque layers (control)
opaque full-screen layers (control)
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 1,024 layers | 1,024–1,040 over 3 runs (CV 0.7 %) | 1,024 layersmeasured | final, 2026-11-15 |
| A12 | 1440p | 584 layers | all equal over 3 runs (CV 0 %) | 584 layersscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 260 layers | all equal over 3 runs (CV 0 %) | 260 layersscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 2,240 layers | all equal over 3 runs (CV 0 %) | 2,240 layersmeasured | final, 2026-11-15 |
| A15 | 1440p | 1,312 layers | all equal over 3 runs (CV 0 %) | 1,312 layersscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 544 layers | 536–552 over 3 runs (CV 1.2 %) | 544 layersscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 3,264 layers | 3,200–3,264 over 3 runs (CV 0.9 %) | 3,264 layersmeasured | final, 2026-11-15 |
| A19 | 1440p | 1,952 layers | all equal over 3 runs (CV 0 %) | 1,952 layersscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 848 layers | all equal over 3 runs (CV 0 %) | 848 layersscaled from 1080p | final, 2026-11-15 |
Alpha-tested layers
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 96 layers | 93–96 over 3 runs (CV 1.5 %) | 81 layersmeasured | final, 2026-11-15 |
| A12 | 1440p | 56 layers | 54–56 over 3 runs (CV 1.7 %) | 47 layersscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 25 layers | all equal over 3 runs (CV 0 %) | 21 layersscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 132 layers | all equal over 3 runs (CV 0 %) | 107 layersmeasured | final, 2026-11-15 |
| A15 | 1440p | 79 layers | 79–80 over 3 runs (CV 0.6 %) | 64 layersscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 37 layers | 35–37 over 3 runs (CV 2.6 %) | 30 layersscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 168 layers | 167–172 over 3 runs (CV 1.3 %) | 111 layersmeasured | final, 2026-11-15 |
| A19 | 1440p | 100 layers | 100–102 over 3 runs (CV 0.9 %) | 66 layersscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 46 layers | 44–46 over 3 runs (CV 2.1 %) | 30 layersscaled from 1080p | final, 2026-11-15 |
Sprites, instanced
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 27,000 sprites | 24,761–27,000 over 5 runs (CV 3.3 %) | 25,411 spritesmeasured | final, 2026-11-15 |
| A12 | 1440p | 12,000 sprites | all equal over 3 runs (CV 0 %) | 11,293 spritesscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 6,625 sprites | 6,500–6,625 over 3 runs (CV 0.9 %) | 6,235 spritesscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 42,000 sprites | all equal over 3 runs (CV 0 %) | 32,400 spritesmeasured | final, 2026-11-15 |
| A15 | 1440p | 22,000 sprites | 20,672–22,000 over 3 runs (CV 2.9 %) | 16,971 spritesscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 10,750 sprites | all equal over 3 runs (CV 0 %) | 8,292 spritesscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 55,885 sprites | 55,849–56,429 over 3 runs (CV 0.5 %) | 33,831 spritesmeasured | final, 2026-11-15 |
| A19 | 1440p | 31,239 sprites | 31,212–31,622 over 3 runs (CV 0.6 %) | 18,911 spritesscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 15,392 sprites | 15,360–15,531 over 3 runs (CV 0.5 %) | 9,317 spritesscaled from 1080p | final, 2026-11-15 |
Sprites, one draw each
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 25,500 sprites | 24,750–25,500 over 3 runs (CV 1.4 %) | 23,547 spritesmeasured | final, 2026-11-15 |
| A12 | 1440p | 11,875 sprites | 11,750–11,875 over 3 runs (CV 0.5 %) | 10,965 spritesscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 6,500 sprites | 6,187–6,500 over 3 runs (CV 2.3 %) | 6,002 spritesscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 36,500 sprites | 36,500–37,000 over 3 runs (CV 0.7 %) | 28,003 spritesmeasured | final, 2026-11-15 |
| A15 | 1440p | 22,000 sprites | all equal over 3 runs (CV 0 %) | 16,879 spritesscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 10,750 sprites | all equal over 3 runs (CV 0 %) | 8,247 spritesscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 55,612 sprites | 54,000–56,247 over 3 runs (CV 1.7 %) | 34,750 spritesmeasured | final, 2026-11-15 |
| A19 | 1440p | 30,890 sprites | 30,818–30,916 over 3 runs (CV 0.1 %) | 19,302 spritesscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 15,023 sprites | 15,017–15,028 over 3 runs (CV 0 %) | 9,387 spritesscaled from 1080p | final, 2026-11-15 |
Meshes, instanced
meshes of ~1,000 triangles, instanced
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 10,000 meshes | 9,875–10,375 over 3 runs (CV 2.1 %) | 9,563 meshesmeasured | final, 2026-11-15 |
| A12 | 1440p | 9,500 meshes | 9,375–9,500 over 3 runs (CV 0.6 %) | 9,085 meshesscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 7,875 meshes | 7,875–8,000 over 3 runs (CV 0.8 %) | 7,531 meshesscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 11,875 meshes | all equal over 3 runs (CV 0 %) | 11,875 meshesmeasured | final, 2026-11-15 |
| A15 | 1440p | 11,375 meshes | 11,375–11,500 over 3 runs (CV 0.5 %) | 11,375 meshesscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 10,750 meshes | 10,750–11,000 over 3 runs (CV 1.1 %) | 10,750 meshesscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 21,000 meshes | all equal over 3 runs (CV 0 %) | 18,120 meshesmeasured | final, 2026-11-15 |
| A19 | 1440p | 19,750 meshes | all equal over 3 runs (CV 0 %) | 17,041 meshesscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 17,750 meshes | all equal over 3 runs (CV 0 %) | 15,316 meshesscaled from 1080p | final, 2026-11-15 |
Meshes, one draw each
meshes of ~1,000 triangles, one draw each
| Chip | Resolution | Budget at 60 Hz | Spread | 20-minute budget | Status |
|---|---|---|---|---|---|
| A12 | 1080p | 9,800 meshes | 9,800–9,900 over 3 runs (CV 0.5 %) | 9,701 meshesmeasured | final, 2026-11-15 |
| A12 | 1440p | 9,300 meshes | 9,200–9,300 over 3 runs (CV 0.5 %) | 9,206 meshesscaled from 1080p | final, 2026-11-15 |
| A12 | 2160p (4K) | 7,900 meshes | 7,900–8,100 over 3 runs (CV 1.2 %) | 7,820 meshesscaled from 1080p | final, 2026-11-15 |
| A15 | 1080p | 12,431 meshes | 12,000–12,907 over 3 runs (CV 3 %) | 11,350 meshesmeasured | final, 2026-11-15 |
| A15 | 1440p | 11,400 meshes | all equal over 3 runs (CV 0 %) | 10,409 meshesscaled from 1080p | final, 2026-11-15 |
| A15 | 2160p (4K) | 10,600 meshes | 10,400–10,922 over 3 runs (CV 2 %) | 9,678 meshesscaled from 1080p | final, 2026-11-15 |
| A19 | 1080p | 20,400 meshes | all equal over 3 runs (CV 0 %) | 16,966 meshesmeasured | final, 2026-11-15 |
| A19 | 1440p | 19,200 meshes | all equal over 3 runs (CV 0 %) | 15,968 meshesscaled from 1080p | final, 2026-11-15 |
| A19 | 2160p (4K) | 17,200 meshes | all equal over 3 runs (CV 0 %) | 14,304 meshesscaled from 1080p | final, 2026-11-15 |
Rendering lower and upscaling to 4K
Shader passes per 60 Hz frame when the frame is drawn at 1080p or 1440p and scaled to 3840×2160 by MetalFX in the same frame.
Shader passes + MetalFX spatial to 4K
passes of the reference shader at the input size + MetalFX spatial to 3840x2160
| Chip | Drawn at, then to 4K | Budget at 60 Hz | Spread | Status |
|---|---|---|---|---|
| A12 | 1080p | 1 pass | all equal over 3 runs (CV 0 %) | final, 2026-11-15 |
| A12 | 1440p | nothing fits | — | final, 2026-11-15 |
| A15 | 1080p | 2 passes | all equal over 3 runs (CV 0 %) | final, 2026-11-15 |
| A15 | 1440p | 1 pass | all equal over 3 runs (CV 0 %) | final, 2026-11-15 |
| A19 | 1080p | 8 passes | all equal over 3 runs (CV 0 %) | final, 2026-11-15 |
| A19 | 1440p | 5 passes | 4–5 over 3 runs (CV 9.4 %) | final, 2026-11-15 |
Shader passes + MetalFX temporal to 4K
passes of the reference shader at the input size + MetalFX temporal to 3840x2160
| Chip | Drawn at, then to 4K | Budget at 60 Hz | Spread | Status |
|---|---|---|---|---|
| A12 | 1080p | not supported on this chip | — | final, 2026-11-15 |
| A12 | 1440p | not supported on this chip | — | final, 2026-11-15 |
| A15 | 1080p | nothing fits | — | final, 2026-11-15 |
| A15 | 1440p | nothing fits | — | final, 2026-11-15 |
| A19 | 1080p | 3 passes | all equal over 3 runs (CV 0 %) | final, 2026-11-15 |
| A19 | 1440p | 1 pass | all equal over 3 runs (CV 0 %) | final, 2026-11-15 |
Method
Who measured, on what
One A12, one A15, two A19 units, measured by one developer with his own test app: native Metal tests written for this, with no code from a game or from another benchmark’s renderer. Every A19 figure is from A19 unit 1 unless it names A19 unit 2. Unit 2 ran the early tests only, and is there to show whether two units of the same box agree.
What each test does
- Full-screen shader passes. Each pass is its own render pass over the whole screen that reads the previous pass and stores its result, so the GPU cannot skip any. The reference shader does, per pixel, 8 bilinear samples of the previous pass, then 32 loop iterations of two 4-wide multiply-adds and one 4-wide fract.
- Blended layers. Alpha-blended full-screen layers in one render pass. Opaque layers are the control: the GPU removes hidden surfaces before shading, so they cost little. Alpha-tested layers discard some pixels.
- Sprites. Textured quads, drawn instanced or with one draw call each; their size scales with the screen.
- Meshes. Meshes of ~1,000 triangles, instanced, and the same with one draw call each. They are limited by triangle throughput, not by pixels.
- Upscaling. Shader passes at 1080p or 1440p, one MetalFX scale (spatial or temporal) to 3840×2160, and a one-pixel read-back, all in the frame.
- A real frame. Apple’s “Rendering a Scene with Deferred Lighting” sample inside the same off-screen loop, with its shaders unchanged.
- CPU. CoreMark 1.0 under EEMBC’s run rules (one thread); Jolt Physics’ performance test (its ragdoll scene) on one thread and on all; a STREAM-style triad for memory bandwidth; and decompression with zstd, lz4 and Apple’s Compression framework over a generated 24 MiB corpus. One thread runs at user-interactive priority; all threads means one per logical core, with work handed out as threads free up.
- Memory. Allocate in 64 MB steps, writing to every page and rendering into every texture so the memory is really in use, until tvOS ends the app. Every step is written to disk first, so the last good step survives the kill.
How a budget is found
Every GPU test runs off-screen with a fixed time step, so every run draws the same frames, with at most 3 frames in flight, timed by when each frame finishes. The load doubles until a step fails, then a bisection finds the edge. A step passes when the 99th-percentile frame interval of a 3 s probe is at most 16.0 ms (4 % headroom under one 60 Hz frame). The largest passing load must then hold 16.67 ms for one 40 s hold. There is one hold per launch: if it misses, the load is cut by the size of the miss and held again in a new launch, after another rest, because on a warming box a second hold in the same launch only chases the temperature. The figure is the median of the repeats; the campaign takes three launches per figure, and the tables say how many each figure has so far.
On a real 60 Hz display
Budgets are measured off-screen. To check that they hold on a screen, the same work was drawn at loads around the budget, then copied or scaled into the 3840×2160 drawable once per display refresh for a minute, counting the refreshes that showed no new frame. At each chip’s 1080p budget, 6 of 7 one-minute runs on a 60 Hz display missed no refresh at all; the worst missed 0.17 % of refreshes (A12, blended layers). About 10 % above the budget, 5 of 6 runs missed refreshes, up to 5.9 %. With CAMetalDisplayLink instead, runs at or below the budget missed up to 6 % of refreshes, so the runs that count use CADisplayLink.
The on-screen runs (49)
| Box | Test | Drawn at | Load | × budget | Display link | Refreshes | Missed | Missed share | Longest run of misses |
|---|---|---|---|---|---|---|---|---|---|
| A15 | Full-screen shader passes | 1920×1080 | 4 | 0.8 | CAMetalDisplayLink | 899 | 12 | 1.34 % | 3 |
| A12 | Full-screen shader passes | 1920×1080 | 3 | 1 | CAMetalDisplayLink | 719 | 24 | 3.34 % | 2 |
| A12 | Full-screen shader passes | 1920×1080 | 3 | 1 | CAMetalDisplayLink | 3,596 | 71 | 1.97 % | 3 |
| A15 | Full-screen shader passes | 1920×1080 | 5 | 1 | CAMetalDisplayLink | 3,596 | 58 | 1.61 % | 3 |
| A19 unit 1 | Full-screen shader passes | 1920×1080 | 11 | 1 | CAMetalDisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Blended full-screen layers | 1920×1080 | 73 | 0.97 | CAMetalDisplayLink | 3,596 | 214 | 5.95 % | 3 |
| A15 | Blended full-screen layers | 1920×1080 | 108 | 1 | CAMetalDisplayLink | 3,596 | 39 | 1.09 % | 2 |
| A12 | Full-screen shader passes | 2112×1080 | 3 | 1.1 | CAMetalDisplayLink | 3,596 | 182 | 5.06 % | 3 |
| A15 | Full-screen shader passes | 2112×1080 | 5 | 1.1 | CAMetalDisplayLink | 3,596 | 162 | 4.51 % | 4 |
| A19 unit 1 | Blended full-screen layers | 1920×1080 | 152 | 1.01 | CAMetalDisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Blended full-screen layers | 1920×1080 | 80 | 1.07 | CAMetalDisplayLink | 3,596 | 489 | 13.6 % | 3 |
| A15 | Blended full-screen layers | 1920×1080 | 119 | 1.1 | CAMetalDisplayLink | 3,596 | 234 | 6.51 % | 4 |
| A12 | Full-screen shader passes | 1728×1080 | 3 | 0.9 | CAMetalDisplayLink | 3,596 | 66 | 1.84 % | 2 |
| A15 | Full-screen shader passes | 1728×1080 | 5 | 0.9 | CAMetalDisplayLink | 3,596 | 38 | 1.06 % | 3 |
| A19 unit 1 | Full-screen shader passes | 2112×1080 | 11 | 1.1 | CAMetalDisplayLink | 3,596 | 94 | 2.61 % | 1 |
| A12 | Blended full-screen layers | 1920×1080 | 66 | 0.88 | CAMetalDisplayLink | 3,596 | 169 | 4.7 % | 3 |
| A15 | Blended full-screen layers | 1920×1080 | 97 | 0.9 | CAMetalDisplayLink | 3,596 | 33 | 0.92 % | 2 |
| A19 unit 1 | Blended full-screen layers | 1920×1080 | 167 | 1.11 | CAMetalDisplayLink | 3,596 | 5 | 0.14 % | 1 |
| A15 | Full-screen shader passes | 1920×1080 | 1 | 0.2 | CAMetalDisplayLink | 3,596 | 47 | 1.31 % | 2 |
| A12 | Full-screen shader passes | 1920×1080 | 1 | 0.33 | CAMetalDisplayLink | 3,596 | 151 | 4.2 % | 3 |
| A15 | Blended full-screen layers | 1920×1080 | 86 | 0.8 | CAMetalDisplayLink | 3,596 | 64 | 1.78 % | 2 |
| A12 | Blended full-screen layers | 1920×1080 | 58 | 0.77 | CAMetalDisplayLink | 3,596 | 55 | 1.53 % | 2 |
| A19 unit 1 | Full-screen shader passes | 1728×1080 | 11 | 0.9 | CAMetalDisplayLink | 3,596 | 0 | 0 % | 0 |
| A15 | Full-screen shader passes | 1920×1080 | 4 | 0.8 | CAMetalDisplayLink | 3,596 | 37 | 1.03 % | 3 |
| A12 | Full-screen shader passes | 1920×1080 | 2 | 0.67 | CAMetalDisplayLink | 3,596 | 118 | 3.28 % | 3 |
| A15 | Blended full-screen layers | 1920×1080 | 65 | 0.6 | CAMetalDisplayLink | 3,596 | 57 | 1.59 % | 3 |
| A12 | Blended full-screen layers | 1920×1080 | 44 | 0.59 | CAMetalDisplayLink | 3,596 | 103 | 2.86 % | 3 |
| A19 unit 1 | Blended full-screen layers | 1920×1080 | 137 | 0.91 | CAMetalDisplayLink | 3,596 | 1 | 0.03 % | 1 |
| A12 | Full-screen shader passes | 1920×1080 | 1 | 0.33 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Full-screen shader passes | 1920×1080 | 1 | 0.33 | CAMetalDisplayLink | 3,596 | 104 | 2.89 % | 2 |
| A15 | Full-screen shader passes | 1920×1080 | 1 | 0.2 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Full-screen shader passes | 1920×1080 | 3 | 1 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A15 | Full-screen shader passes | 1920×1080 | 5 | 1 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Blended full-screen layers | 1920×1080 | 73 | 0.97 | CADisplayLink | 3,596 | 6 | 0.17 % | 1 |
| A15 | Blended full-screen layers | 1920×1080 | 108 | 1 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Full-screen shader passes | 2112×1080 | 3 | 1.1 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A15 | Full-screen shader passes | 2112×1080 | 5 | 1.1 | CADisplayLink | 3,596 | 114 | 3.17 % | 1 |
| A12 | Blended full-screen layers | 1920×1080 | 80 | 1.07 | CADisplayLink | 3,596 | 106 | 2.95 % | 1 |
| A15 | Blended full-screen layers | 1920×1080 | 119 | 1.1 | CADisplayLink | 3,596 | 211 | 5.87 % | 1 |
| A12 | Full-screen shader passes | 1728×1080 | 3 | 0.9 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A15 | Full-screen shader passes | 1728×1080 | 5 | 0.9 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Blended full-screen layers | 1920×1080 | 66 | 0.88 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A15 | Blended full-screen layers | 1920×1080 | 97 | 0.9 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A12 | Full-screen shader passes | 1920×1080 | 3 | 1 | CADisplayLink | 899 | 0 | 0 % | 0 |
| A15 | Full-screen shader passes | 1920×1080 | 4 | 0.8 | CADisplayLink | 899 | 0 | 0 % | 0 |
| A19 unit 1 | Full-screen shader passes | 1920×1080 | 11 | 1 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A19 unit 1 | Blended full-screen layers | 1920×1080 | 152 | 1.01 | CADisplayLink | 3,596 | 0 | 0 % | 0 |
| A19 unit 1 | Full-screen shader passes | 2112×1080 | 11 | 1.1 | CADisplayLink | 3,596 | 77 | 2.14 % | 1 |
| A19 unit 1 | Blended full-screen layers | 1920×1080 | 167 | 1.11 | CADisplayLink | 3,596 | 4 | 0.11 % | 1 |
Two figures matter, because no chip holds its first seconds: from rest (the first seconds on a rested box, what short benchmarks elsewhere report) and settled (after the first half-minute). The budgets are settled figures; the 20-minute figures go further.
Temperature
Heat is the easiest way to get a wrong number out of these boxes, so this is what was done about it.
- What was controlled. Before any warm-up or timing, the app sat idle, drawing nothing but a label, until tvOS had reported its normal thermal state for a set time. A box that would not settle was not measured: the run was refused, not taken hot. Every budget launch followed 240 s at nominal. Before the 20-minute runs, CPU tests and scene runs the waits actually were 120–900 s on the A12, 120–989 s on the A15, 120–1,070 s on the A19.
- What the signal is, and is not. tvOS gives an app four thermal states, nominal, fair, serious and critical, and no temperature. “Nominal” is a range. That the range was narrow enough here rests on the two A19 units agreeing. On the 16 budgets both units ran, the median difference was 0.4 % and the largest 9.8 % (sprites, one draw each at 4K). In the same 20-minute run they were 1.2 % apart at one minute and 3.3 % apart in the last minute, and first reported “serious” at 1:59 and 2:01. The three repeats of every budget are whole passes over the test list, taken hours apart, so drift between them shows up as spread between repeats: Over the 78 budgets with repeats so far, the spread between repeats (coefficient of variation) is 0.1 % at the median and 9.4 % at most. And nominal is not rested: after 2 minutes of full load (it reported serious at 1:54), A19 unit 2 took 69 s to read nominal again; just back at nominal it ran 2.7 % slow against its own rested run, and after 240 s at nominal it was still 0.9 % slow (one unit, one series of 6 runs, a Debug build). That is why every launch rests first, and longer after a long run.
The rest measurements
chip box role test load size seconds state at launch seconds to reach nominal rest at nominal s fps fps first second fps last second vs rested reference first serious s thermal end build interruptions date time A19 A19 unit 2 rested reference pass 4 1920x1080 10 nominal 0 20 212 213 211 1 — nominal Debug 0 2026-11-14 0832 A19 A19 unit 2 heating run pass 4 1920x1080 120 nominal 0 20 200 214 190 0.95 114 serious Debug 0 2026-11-14 0838 A19 A19 unit 2 recovery pass 4 1920x1080 10 serious 68.6 20 206 209 199 0.97 — nominal Debug 0 2026-11-14 0842 A19 A19 unit 2 recovery pass 4 1920x1080 10 nominal 0 60 207 209 200 0.98 — nominal Debug 0 2026-11-14 0844 A19 A19 unit 2 recovery pass 4 1920x1080 10 nominal 0 120 209 210 207 0.99 — nominal Debug 0 2026-11-14 0845 A19 A19 unit 2 recovery pass 4 1920x1080 10 nominal 0 240 210 211 208 0.99 — nominal Debug 0 2026-11-14 0848 - Each chip slows while still reporting nominal. On the same fixed load, frames per second fell 5 % below the first ten seconds for the A12 within 60 s, the A15 within 20 s and the A19 within 30 s of starting, still reporting nominal. Over the run, the loss was the A12 23 % over the whole run without leaving nominal, the A15 17 % before it reported anything hotter (at 8:40) and the A19 7 % before it reported anything hotter (at 1:50). An app cannot wait for the thermal state to change before shedding load. This is why the results are given from rest and settled, and why the chart of frames per second against time is shown.
- What heat does. The short-test numbers are a cool box’s numbers, taken about a minute into the load. The 20-minute runs show what is left once the box is warm, as their own charts and their own column in the tables, never mixed into the short-test charts.
- What differs between the boxes. The A12 model has a fan; the A15 and A19 models do not. One of the two A19 units, A19 unit 1, the one every A19 figure comes from, reported its highest thermal state while idle where it sits, before any test. Placement and ventilation matter as much as the chip.
- The raw data. Every run records the thermal state for every second. The downloadable series gives it for each 10-second window (the hottest state reported in it), and each 20-minute run lists when it first reported each state.
What was left out, and why
- Runs in which the app lost the screen (a system prompt or the screen saver on top of it) are not used.
- A hold that missed is not a budget; a load that never confirmed shows as “not confirmed”.
- Launches that tvOS ended before anything was measured were run again; nothing from them is used.
- Only Release builds count: a Debug build changes the CPU-bound tests a lot.
- A19 unit 2 was unavailable after the early tests, so the later A19 figures are unit 1 alone.
What was not measured
- A heavy scene (a port of Apple’s “Modern Rendering with Metal” sample) is planned and has not been run.
- Ray tracing and mesh shaders: the boxes report whether they have them, but their speed was not measured.
- The MetalFX temporal scaler with real motion; it was timed with still, unjittered content.
- Box2D physics: the results are in the CPU download but are not charted here.
- Other streaming boxes and phones: they are not compared.
The three boxes
What each box reports to an app at run time, read by the test app itself (except cooling, which comes from a teardown and press reviews). MB = 2^20 bytes. The A19 column is A19 unit 1; unit 2 reports the same figures.
| A12 | A15 | A19 | |
|---|---|---|---|
| tvOS (build) | 27.2.1 (24K157) | 27.2 (24K122) | A19 unit 1: 27.2.1 (24K157); A19 unit 2: 27.2.1 (24K157) |
| Model identifier | AppleTV11,1 | AppleTV14,1 | AppleTV18,1 |
| Memory | 2.94 GB | 3.88 GB | 7.87 GB |
| What one app can use | 2,098 MB | 2,098 MB | 2,098 MB |
| Available to the app at launch | 2,089 MB | 2,089 MB | 2,091 MB |
| CPU cores | 6 (2 performance + 4 efficiency) | 5 (2 performance + 3 efficiency) | 5 (2 performance + 3 efficiency) |
| L2 cache | 8 MiB performance, 2 MiB efficiency | 12 MiB performance, 4 MiB efficiency | 8 MiB performance, 4 MiB efficiency |
| Cooling | fan (iFixit teardown) | no fan (press reviews; no teardown found) | no fan (press reviews; no teardown found) |
| GPU | Apple A12 GPU | Apple A15 GPU | Apple A19 GPU |
| Highest Apple GPU family | apple5 | apple8 | apple10 |
| Apple GPU families supported | apple1, apple2, apple3, apple4, apple5 | apple1, apple2, apple3, apple4, apple5, apple6, apple7, apple8 | apple1, apple2, apple3, apple4, apple5, apple6, apple7, apple8, apple9, apple10 |
| Metal 3 | no | yes | yes |
| Metal 4 | no | yes | yes |
| Ray tracing | no | yes | yes |
| Unified memory | yes | yes | yes |
| MetalFX spatial scaler | yes | yes | yes |
| MetalFX temporal scaler | no | yes | yes |
| MetalFX framework loads | yes | yes | yes |
| Max buffer length, MB | 1,536 | 2,048 | 4,096 |
| Recommended max working set, MB | 2,048 | 2,731 | 5,461 |
| GPU counter sets | timestamp | timestamp | timestamp |
| Counter sampling at blit boundary | no | no | no |
| Counter sampling at dispatch boundary | no | no | no |
| Counter sampling at draw boundary | no | no | no |
| Counter sampling at stage boundary | yes | yes | yes |
| Counter sampling at tile dispatch boundary | no | no | no |
| Native screen | 3840×2160 | 3840×2160 | 3840×2160 |
| Max frames per second | 60 | 60 | 60 |
Raw data
The files behind every chart and table on this page, as published on 16 November 2026.
- tvbench-data.jsonEverything on this page, one JSON file
- tvbench-budgets.csvBudgets per chip, test and resolution, 84 rows
- tvbench-sustained.csvThe 20-minute runs, one row per run, 28 rows
- tvbench-sustained_series.csvThe 20-minute runs every 10 seconds, with the thermal state, 3,720 rows
- tvbench-cpu.csvCPU results, every figure and worker count, 410 rows
Schema tvbench-data-1. Every row carries how it was measured, its date and its number of runs (n). MB: 2^20 bytes, except STREAM and compression (10^6 bytes). Frames per second above 60: off-screen throughput, not frames on screen. Provisional: a budget row not yet measured in the requested campaign: the latest line outside it is shown instead.
For coding agents
The same findings, written as rules an AI coding agent can load while it builds or tunes a tvOS app or game: what fits in a frame on each chip, what heat takes away, the memory limit, what each box supports and how to tell the boxes apart at run time, with each rule’s caveat. It is for developers who work with an agent such as Claude Code and want it to plan against these budgets instead of guessing. Pack version 1.0, from the data of 16 November 2026.
- SKILL.mdThe Claude Code skill: the rules, and when to use them, 15 KB
- apple-tv-4k-budgets.mdThe same rules as plain Markdown, for other tools (AGENTS.md style), 14 KB
- budgets.jsonThe full tables the skill points to, 38 KB
To install the skill in Claude Code, save SKILL.md and budgets.json into the folder .claude/skills/apple-tv-4k-budgets/ in a project (or ~/.claude/skills/apple-tv-4k-budgets/ to have it in every project):
mkdir -p .claude/skills/apple-tv-4k-budgets
curl -o .claude/skills/apple-tv-4k-budgets/SKILL.md https://icallwindowseat.com/tvbench/pack/SKILL.md
curl -o .claude/skills/apple-tv-4k-budgets/budgets.json https://icallwindowseat.com/tvbench/pack/budgets.jsonClaude Code can then load it on its own whenever a task matches the skill’s description, such as building or tuning a tvOS app or choosing quality tiers per Apple TV model.
Changelog
- Data file regenerated (campaign final-1115b, budget runs through 16 November 2026).
- First version of the page, built while the final measurement campaign is still running: figures marked provisional are from earlier single runs and are being replaced.
- The final measurement campaign finished on all three chips. Every budget is now the median of three or more confirmed repeats, and no figure is provisional.
Measured by Josh, who makes I Call Window Seat and RevealHaven. Neither app was used as a test.