SwiftPython

Measurements / September 2026

Mac and iPad duplex measurements

The same echo workload was run on a MacBook Pro and a physical iPad mini. Each device completed 85 trials. Every returned payload was checked against its input.

29–30 September 2026 UTC · Download PDF · Per-trial data (JSON)

Scroll tables and charts sideways to see both devices.

Test systems

Test system MacBook Pro iPad mini
Processor M4 Pro, 14 cores A17 Pro, 6 cores
Memory 48 GiB 7.72 GiB reported
OS macOS 26.2 (25C56) iPadOS 26.7 (23H24)
Python 3.13.3 3.13.15*
Host / worker CLI host / native sidecar SwiftUI host / app extension
Build Swift 6.3.3, optimized Swift 6.3.3, Release

*The iPad Python version was read from the pinned framework header.

Receipt latency

Figure 1. Receipt latency. Open symbols are trial statistics; filled symbols are medians. Lines span the observed minimum and maximum. Five trials per device and payload.
Figure 1. Receipt latency. Open symbols are trial statistics; filled symbols are medians. Lines span the observed minimum and maximum. Five trials per device and payload.
Payload Mac p50 Mac p95 Mac p99 iPad p50 iPad p95 iPad p99
64 B 71.3 90.5 104.6 91.8 125.6 145.3
1 KiB 71.8 91.8 108.2 92.8 126.2 146.1
16 KiB 80.9 102.0 116.0 115.3 148.9 169.4
256 KiB 204.2 250.1 285.9 296.0 326.5 362.3

Table 1. Microseconds, medians of trial percentiles. Timing ends when Swift receives the complete Python echo. Payload verification and acknowledgement follow that timestamp.

Verified transfer throughput

Swift sends while Python returns the received bytes. The reported rate counts bytes in one direction; combined traffic is twice that rate. Timing includes full output verification and acknowledgement.

Figure 2. Trial rates and medians. Both panels start at zero; the upper panel uses MB/s and the lower panel GB/s. The managed route changes input storage; Python still constructs the output bytes.
Figure 2. Trial rates and medians. Both panels start at zero; the upper panel uses MB/s and the lower panel GB/s. The managed route changes input storage; Python still constructs the output bytes.
Payload / frame window / ack Mac MB/s iPad MB/s n
16 KiB / default / 1 272.3 [253.5, 277.1] 157.1 [138.2, 162.3] 5
16 KiB / 64 / 1 321.1 [290.1, 324.5] 170.3 [156.6, 174.4] 5
16 KiB / 64 / 16 386.2 [372.2, 411.6] 222.1 [213.4, 225.8] 5
256 KiB / 16 / 4 2,438.5 [2,181.8, 2,464.5] 1,046.5 [1,043.1, 1,058.6] 5
Managed 256 KiB / 16 / 4 2,705.6 [2,638.6, 3,103.5] 1,518.8 [1,516.3, 1,523.5] 3
1 MiB / 16 / 4 2,896.0 [2,829.9, 3,342.1] 1,171.1 [969.0, 1,204.8] 5
Managed 1 MiB / 16 / 4 4,654.0 [4,328.9, 5,107.3] 2,062.7 [2,057.9, 2,071.5] 3

Table 2. Median [minimum, maximum], decimal MB/s per direction. n is the number of trials on each device. Default output credit is 64 KiB; the 64-frame setting at 16 KiB permits 1 MiB. Ack is the number of output frames between acknowledgements.

At 1 MiB, managed ingress measured 4.654 GB/s on the Mac and 2.063 GB/s on the iPad, including echo processing and full byte comparison.

Concurrent sessions

Each session sends a 64-byte message, receives and verifies the echo, then acknowledges it before sending the next message. Sessions begin their measured loops at a shared barrier.

Figure 3. Three trials per configuration and device. Symbols show individual trials and their median; lines show the observed range. Message rate counts completed echo cycles across all sessions.
Figure 3. Three trials per configuration and device. Symbols show individual trials and their median; lines show the observed range. Message rate counts completed echo cycles across all sessions.
Sessions / workers Mac msg/s Mac p50 Mac p99 iPad msg/s iPad p50 iPad p99
4 / 1 18,696 130.2 727.8 12,951 200.2 805.5
16 / 1 19,778 704.2 2,110.1 13,580 1,039.8 2,919.1
4 / 4 19,150 128.1 706.9 15,513 165.1 684.0

Table 3. Median message rates and receipt percentiles. Latency is in microseconds. Four-session trials use 2,000 measured messages per session; sixteen-session trials use 1,000.

With four sessions, moving from one worker to four increased the median message rate by 19.8% on iPad and 2.4% on Mac; these measurements do not identify the limiting component.

Completion and a paused consumer

In the final cohort, all three standard sixteen-session trials and ten additional trials completed on each device. Separate delayed-consumer trials pause consumption for 500 ms, then verify the remaining output and complete shutdown.

Output credit Mac queue high-water iPad queue high-water Trials / device
64 KiB 64 KiB 64 KiB 3
1 MiB 1 MiB 1 MiB 3

Table 4. Observed queue high-water during delayed-consumer trials. Each trial sends 1,024 frames of 16 KiB after 30 warmups.

Idle-session snapshots

One worker is measured before sessions open, with sessions open, and after they close. The table pairs the first two snapshots. The figure shows the open-session counts.

Figure 4. Host and active-worker threads with 1, 4 and 16 idle sessions. Points are the three trial snapshots; lines connect their medians. Overlapping points may appear as one.
Figure 4. Host and active-worker threads with 1, 4 and 16 idle sessions. Points are the three trial snapshots; lines connect their medians. Overlapping points may appear as one.
Device Sessions Host threads Worker threads Host RSS, MiB Worker RSS, MiB
Mac 1 8 / 12 24 / 27 31.0 / 34.8 41.5 / 42.2
Mac 4 8 / 21 24 / 36 30.9 / 41.2 41.5 / 42.5
Mac 16 8 / 57 24 / 72 31.0 / 66.8 41.3 / 44.2
iPad 1 15 / 18 24 / 27 123.3 / 125.3 42.8 / 43.3
iPad 4 16 / 28 24 / 36 123.2 / 131.5 42.8 / 43.7
iPad 16 15 / 63 24 / 72 123.2 / 156.7 42.8 / 45.3

Table 5. Before / open, medians of three trials. RSS is resident set size, not peak allocation or exclusive ownership. MiB is 1,048,576 bytes.

Sampling

Mac RSS comes from process snapshots; iPad RSS and thread counts come from Mach queries within each process. The iPad host includes SwiftUI and extension hosting. Its standby extension is outside the active-worker figures. Absolute host memory therefore describes different application setups.

Thread counts returned close to their starting values after sessions closed. Memory snapshots were taken before process exit. Every selected trial also checked that the benchmark worker processes had exited.

Measurement method

Data and timing

The payload is a repeated byte sequence, without compression. Python echoes the input, and Swift compares every returned byte. Elapsed time uses the monotonic uptime clock. Bulk intervals include send, echo, verification and acknowledgement. Latency samples end at receipt; full-cycle samples include verification and acknowledgement.

Payload Mac full-cycle p50, µs iPad full-cycle p50, µs
64 B 91.4 110.2
1 KiB 91.6 111.0
16 KiB 100.8 132.6
256 KiB 221.1 315.5

Runs and aggregation

Each trial uses fresh processes. Startup, warmup and the iPad's 11-second relaunch wait are outside measurement. Cases are shuffled within suites using seed 20260929. Builds do not overlap measurement.

Tables report medians of trial statistics, including trial p99s. Ranges span observed minima and maxima, not confidence intervals. Counts below are configurations × repeats. *Concurrency lists sessions/workers.

Suite Measured frames / session Warmup frames Trials / device
Latency: 64 B / 1 KiB 25,000 100 2 × 5 = 10
Latency: 16 KiB / 256 KiB 15,000 / 5,000 100 / 50 2 × 5 = 10
Bulk: five settings in Table 2 131,072 / 65,536 / 32,768 100 / 100 / 30 5 × 5 = 25
Managed: 256 KiB / 1 MiB 65,536 / 32,768 100 / 30 2 × 3 = 6
Concurrency: 4/1, 16/1, 4/4* 2,000 / 1,000 100 3 × 3 = 9
Additional 16-session runs 1,000 100 1 × 10 = 10
Idle: 1 / 4 / 16 sessions No measured traffic 0 3 × 3 = 9
Paused consumer: two credits 1,024 30 2 × 3 = 6
Total per device 85

Cohort and scope

The comparison uses the final 85-trial cohort from each device. Earlier source and setup trials, smoke tests, and instrumented follow-ups are outside these cohorts. No observation within either selected cohort was removed. Background system services remained active. All recorded iPad thermal samples were nominal.

Hardware, OS, Python version and worker hosting differ. The tables compare complete platform setups; they do not isolate any one of those factors. Measurements were collected on 29-30 September 2026 UTC. Per-trial statistics and parameters accompany this report in data.json.

Figures

Download the vector figures: latency, throughput, concurrent sessions, idle sessions.