Decoder CPU-efficiency comparison¶
Curated source-specific evidence, not a new run or qualification of current source. Dates, measurements, failure outcomes and original source identifiers retain their recorded meaning. Old tags/assets have been retired and original commits may no longer resolve after the privacy rewrite. Where identifiers were sanitized earlier, the original hardware/run association cannot be independently recovered from the public derivative. Current acceptance lives in the roadmap.
Accepted baseline and scope¶
Application d61cf03de9e0a528e0b606cfa40d81a8306281e3, qualification documentation
ab4f5fa2fad3ec1ecc680721d5867381264a47d2; initial checkout clean. The operator
accepted Phase 3, corrections and the first Viterbi patch. All performance
comparisons here use that already optimized receiver, not the older scalar
build used for the preceding 20% improvement.
Baseline source/build was frozen before editing and reproduced the installed
decoder exactly. Native flags -O2 -std=c++20 -Wall -Wextra -pthread, g++
14.2.0-19, Pi 5 Cortex-A76/aarch64, Debian Trixie, kernel
6.18.50+rpt-rpi-2712, four cores/7.9 GiB RAM. New-path SDK 4.0.23 and
legacy 3.1.69 unchanged. No fast-math, precision/reduction/contraction changes,
new SIMD, threads, queues, dependencies or global CPU-policy changes.
Validation reuses the already-required native nlohmann JSON header; the focused
CI job now installs it explicitly for the new test manifest reader.
| Artifact | SHA-256 |
|---|---|
| Accepted installed/rebuilt native decoder | 9541bc612f847031ec42d1ea55671335ca10aa43f7f1647a317663e77ea01ff1 |
| Frozen receiver source | e31d4ed29027ba075a76d3424efac1a3add524cff361b81edd3099aaceba0c57 |
| Frozen channelizer source | 28d78432217e760ab1a81ac8e9c518d7c8e356f8c25b24862e066fbfe118c81f |
| Accepted memory64 receive module | e417e5acfaf22225e3da54151a4a650f6e1654a45a18009e227deab7e9c2c2d9 |
| Combined native decoder | ecf784265ac39b79d8b9ad31c3bae57996796b4f9b8960ce5af34a701321667c |
| Combined memory64 receive module | aea445a443eb97341eb9389c110f75e7450246b59b993381354e1f438692718d |
| Combined memory64 DSP audit | 8c09c5f6e7d0dad043af0a39bb8ff7d1f8bf92b3070b1ef127f91f7016d0d8e4 |
| Combined receiver source | 205b1876a694f5abf38bd373e6d65fb2a82d82bafb197b9a22de5bc5151111f4 |
| Combined channelizer source | 1893579d2d7b4516636d4fd7d83ebac71750509fa8d63380712ef326698f0479 |
Normal memory64 module 32,353 → 32,362 bytes, inspected i64 memory/address
ABI, 16-MiB initial/64-MiB maximum, exact artifact/export/import allowlists.
These qualification artifacts are not deployed legacy-browser modules.
Native nm finds no audit, frozen-reference or DSP-tracing symbols.
Three isolated candidates, then the combination¶
- Viterbi destination butterflies: evaluate both predecessors for each destination, retain the lower predecessor on strict-comparison ties, write each next metric once, then swap the two metric-buffer pointers. Remove redundant initialization and 256-byte per-step copying. Keep all required transitions, integer costs, survivor bytes, final selection, traceback and hard-disagreement re-encoding unchanged.
- SRRC mirrored delay lines: duplicate 32-entry I/Q rings into 64 entries; remove two modulo/index constructions per symmetric pair. Same 15 symmetric pairs plus center, operands, accumulation and effective FMA behavior. 256 additional bytes per instance, no hot-path allocation.
- FIR active-phase packing: change only coefficient-bank addressing from
phase × 1,023tophase × taps. Same design/normalization/tap order and existing mirrored contiguous sample ring. Current active coefficients occupy 31,744 bytes, now contiguous rather than spread over about 1,043,584 bytes. The original 1,047,552-byte bounded allocation remains provisioned; do not confuse allocated size with active working set.
Each was implemented and measured without the others before stacking. All three passed the frozen DSP/scalar comparisons and had repeatable native and browser gains. No neutral/regressive DSP experiment was retained. A fixture-only frequency-shift sign error was corrected during setup: channelization applies the conjugate NCO, so the production-like capture signal belongs at -100 kHz, not +100 kHz. No receiver code/assertion/tolerance was weakened to fix that setup.
Correctness and weak-signal evidence¶
Verbatim accepted DSP reference is hash-checked before trace-only adapters rename classes/add callbacks. No reference arithmetic, coefficients or decisions are edited. The separate scalar Viterbi oracle still uses independent encoder/branch arithmetic.
- Exhaustive 128 mappings / 65,536 soft pairs × four costs, saturated/neutral/ noisy/tied inputs, every step's metrics, reachability, winner/survivor, final state, traceback/pre-ASM/RS bytes and hard disagreements pass. Audit projects the new destination step onto equivalent logical predecessor relaxations; physical visitation differs, value comparisons do not.
- Every channelized Float32 bit pattern, rational source position, demod/SRRC state and output, soft symbol from all recovery hypotheses, diagnostic/lock/ rejection/lane counters and frame completion cursor agrees with the frozen receiver on the same target. No new epsilon/tolerance. GCC's existing native FMA contraction is retained; WASM has its existing separate operations. Cross-target weak-fixture soft/hypothesis totals can differ as before; that is not silently relabeled cross-target floating-point equality.
- Independent OTA remains 92 ordered 892-byte frames / 17 corrected symbols:
70e147554060f5be9aa117737a6061fcc3f54f46accd58b3f9608862e8450677. Ordered payloads, lane corrections, disagreements and source cursors match. Nonzero lanes: zero-based frame 10[1,0,0,1], 36[0,0,0,1], 55[4,4,2,4]. - Native, ASan/UBSan/leak, clean/corrected/uncorrectable/offset, instance, reset, discontinuity and varied-chunk tests passed. Wider active FIR banks at both rates also pass exact native/sanitized/browser comparisons.
- 32 deterministic weak-signal cases, native/sanitized/actual local memory64: 2.4 and 5 MS/s; three seeds × three noise levels; three carrier/clock-offset fades; noise-only; explicit reset/source gap; and long dropout/reacquisition. Every accepted frame belongs to independent known truth. Noise-only accepts zero. Both strong and recovery cases must decode, not merely match zero.
- Matrix native performs about 1.65 billion exact DSP comparisons plus about 1.08 billion scalar checks; actual browser audits 1,005 candidates, 8,265,120 trellis steps and 1,040,755 ties. Native has 1,006 candidates due to existing target FP behavior, each compared to its own unchanged reference.
- Longer 15.7554125-source-second / 75,625,980-byte varied fixture passed native/browser exact DSP comparisons: 414 validated / 22 rejected / 464 corrected symbols / two loss-and-recovery events on both targets. Includes mixed noise, clock/carrier offsets, fades, repeated signal loss and reacquisition; no fabricated source-gap discontinuity between its parts.
- Real local ARM64 Chromium 154.0.8037.92, owned temporary profile, loopback test server, actual workers/memory64/i64 ABI, Seccomp 2/NoNewPrivs 1. No remote CDP, Node-only substitute, experimental flags or sandbox bypass.
- Shared wire/ownership/interchange checks and gateway protected-queue, continuous-job/auth/read-only/mock TLS/IPv6/fault-isolation tests passed.
Fixture provenance and transforms¶
Independent source: Lucas Teske's
GOES sample index,
GOES-16 2017-01-17 16:25:08 UTC / 5 MS/s. Retain only the established
33,554,432-byte prefix, SHA-256
b34eadddd172d8713be686d29b28541fe86fa3158bdf17f9b57a4ba48530a326.
Prepared CU8: 16,777,194 bytes, SHA-256
144a1d22d34394807320ebc2ee9a4a49c8010bdb7a1e1ca4f42133ce98018cc6.
No new live capture or full-recording download.
Generator documents:
offline 12/25 resampling, 63-tap Blackman-windowed sinc at 1.08 MHz, Float32
storage, linear clock warp, complex carrier rotation, seeded Box–Muller Gaussian
noise, rounded/clipped CU8. Seeds 0x13579bdf / 0x2468ace1 / 0x5f3759df;
noise sigma 0.002 / 0.0045 / 0.008 in normalized IQ units. Offsets
+1,800 / -3,300 / +7,000 Hz, clock +900 / -850 / +1,300 ppm;
3%-amplitude fade at 0.20–0.42 s; 2.3-s RF dropout.
Chunks 1,022/4,094/8,192/32,768/65,536 bytes; reset case adds an explicit
123,456-source-sample discontinuity.
Longer workload concatenates strong, seed-0 high-noise, offset/fade-1, noise-only, long-dropout and seed-2 mid-noise, twice. These are transformed fixtures, not new independent recordings or evidence of improved sensitivity. Raw/generated payloads stay ignored; durable evidence contains only hashes/metadata/counters.
Performance method and results¶
Harness uses the actual
unchanged Decoder::process/publish/snapshot: input copying/validation,
channelization, demodulation/FEC, evidence/hash, canonical record encode/verify/
zero/disposal and one serialized snapshot per source second. Entry-point/service
loop, real IPC/network/SDR are not invoked. File load/hash, construction/filter
setup, compilation and warmup are separately recorded/excluded. Live checks
qualify IPC separately.
Pin only the test process to CPU 3; warm twice, alternate order, five measured pairs with eight replays for strong workloads and two for the longer varied workload. Browser is measured separately with actual worker wall time including ABI copies/drain, not claimed native CPU/PMU or native record/IPC cost. No global CPU policy changes. All fixed-workload diagnostics and outputs agree.
The Viterbi/SRRC experiments used the initial harness (baseline SHA
9d4c861ceb9ce282a7909161d6e8b2623fd99fc4d780a8dae227d37cd0607b1f);
FIR/combined use the extended, matched harness (baseline SHA
ff905c753809f7a8b07bbbbf5efd5078b620e79d24da527f078b35a4aecaaacd,
combined 6ae038ed105e269d277ff1a4c1e8f10222b6c14fed2e2121e1023824df967f5c).
It adds pre-timing integrity/workload selection, not production changes.
Compare each candidate with its matched baseline; never add isolated savings
or infer precise kernel costs from cross-harness absolute differences.
The added whole-input hash raises benchmark peak RSS, not decoder resident RAM.
| Isolated candidate, 5 MS/s | Native whole CPU reduction, mean ± sample SD | Browser wall reduction |
|---|---|---|
| Viterbi | 17.135% ± 0.159 points | 14.909% ± 0.185 points |
| SRRC | 3.780% ± 0.147 points | 2.022% ± 0.426 points |
| FIR packing | 0.888% ± 0.109 points | 0.990% ± 0.149 points |
Combined native means ± sample SD, each five alternating pairs:
| Workload | Baseline CPU seconds/source second | Candidate | Paired reduction |
|---|---|---|---|
| Independent 5 MS/s | 0.589709 ± 0.000741 | 0.478970 ± 0.000226 | 18.779% ± 0.109 points |
| Derived strong 2.4 MS/s | 0.566040 ± 0.000623 | 0.452251 ± 0.000475 | 20.103% ± 0.094 points |
| Varied 2.4 MS/s, 31.51 source seconds/run | 0.459224 | 0.392406 | 14.550% ± 0.097 points |
Reduction ranges: 18.676–18.950% / 19.978–20.242% / 14.408–14.643%. The varied input spends more time without decodable frames; savings are not uniform across signal conditions. Each run has matching deterministic counters.
| Native workload | Speed relative to real time, baseline → candidate | Cycles/source sample | Instructions/source sample |
|---|---|---|---|
| 5 MS/s | 1.696× → 2.088× | 282.49 → 229.48 | 691.53 → 637.57 |
| Strong 2.4 MS/s | 1.767× → 2.211× | 564.80 → 451.35 | 1,373.19 → 1,260.77 |
| Varied 2.4 MS/s | 2.178× → 2.548× | 458.27 → 391.65 | 1,140.07 → 1,056.93 |
Scoped internal PMU enabled/running times agree; no unreported multiplex scaling. CPU time includes user/system thread time; these counters exclude kernel.
| Native workload | Block p50/p95/p99 ms, baseline → candidate | Worst measured block ms |
|---|---|---|
| 5 MS/s, fixed 65,536-byte blocks | 2.429/6.469/6.621 → 2.298/4.670/4.752 | 7.090 → 5.039 |
| Strong 2.4 MS/s, varied blocks | 0.581/8.742/8.821 → 0.548/6.763/6.830 | 9.072 → 7.566 |
| Varied 2.4 MS/s, varied blocks | 0.590/8.727/8.843 → 0.561/6.741/6.816 | 11.220 → 8.496 |
Percentiles are run means; maximum is the worst across runs. Do not compare raw p50 across different chunk distributions. Benchmark RSS is unchanged within allocator/page noise: 44,080–44,128 KiB (5 MS/s), 23,616–23,648 KiB (strong 2.4), 208,560–208,608 KiB (75-MB varied input plus pre-timing hash). This fixture-loaded test RSS is not live decoder resident memory. Native temperature 65.3–71.4 °C, clock about 2.400 GHz, all sampled throttling flags 0x0. Live accepted decoding continued at roughly 0.58 cores; gateway/demand/background load and thermal samples are retained per run.
Actual local browser measurements, seven pairs/two replays for strong input, five pairs/one replay for varied input, with identical frozen/candidate workloads:
| Browser workload | Wall seconds/source second, baseline → candidate | Paired wall reduction | Mean block p99 ms |
|---|---|---|---|
| 5 MS/s | 0.918151 → 0.754943 | 17.775% ± 0.283 points | 9.847 → 7.089 |
| Strong 2.4 MS/s | 0.874047 → 0.705763 | 19.253% ± 0.112 points | 13.539 → 10.557 |
| Varied 2.4 MS/s | 0.715390 → 0.618329 | 13.568% ± 0.024 points | 13.418 → 10.471 |
Both modules retain 16 MiB linear memory. Native CPU/PMU gains are not inferred browser results. Browser per-worker CPU, hardware PMU and whole-renderer RSS attribution are unavailable in this harness; report wall time/linear memory, not fabricated CPU or GPU utilization. Matching ordered bytes, lane evidence, all 32 metric fields and zero output drops accompany every timed run.
Durable payload-free evidence retains every paired run, artifact hashes, thermal/clock/load conditions, native/browser matrix totals and sampled profile/PMU summaries. No waveform, decoded frame bytes, stack snapshots or credentials are included. Larger raw profiles, fixtures and local logs remain ignored build evidence.
Profiles, counters and code inspection¶
Profiles use IP-only cycles:u sampling, no live IQ/stack/payload collection.
Baseline: roughly 35% Viterbi / 29% demod+SRRC / 22% channelization /
7% timing recovery. Viterbi-only profile: about 22% Viterbi, while other
percentages rise because total work is lower; that does not mean FIR slowed down.
Matched baseline/combined sampled profiles, normalized by their measured whole
CPU/source means (estimates, not instrumented function timers):
| Hotspot | Baseline share / CPU seconds per source second | Combined share / normalized cost |
|---|---|---|
| Viterbi | 33.83% / 0.19950 | 22.83% / 0.10935 |
| Demodulation/Costas/AGC/SRRC | 28.04% / 0.16535 | 32.40% / 0.15519 |
| Channelization/DC/NCO/FIR | 23.49% / 0.13852 | 26.48% / 0.12683 |
| Timing recovery | 7.09% / 0.04181 | 9.16% / 0.04387 |
| Reed–Solomon | 2.12% / 0.01250 | 2.77% / 0.01327 |
IP-only 199-Hz sampling, zero lost samples; no stack/payload capture. Profiles include setup/hash while normalized CPU means are steady-loop scoped, so small differences (e.g. timing/RS) are sampling/scope noise, not proof those unchanged algorithms slowed down. FIR is inlined into channelizer processing; percentages do not isolate tap arithmetic from DC/NCO/resampler work.
Disassembly confirms existing native FMA in symmetric SRRC and contiguous mirrored sample access in FIR. No manual SIMD was added. Linux sysfs reports 64-KiB L1 data cache / 64-byte lines; allocated bank size alone was not used as proof of cache cost. Branch, L1-data and dTLB events are available on Cortex-A76. An exploratory eight-event run multiplexed (~70% running); it is not used as exact qualification evidence. Final groups disable harness PMU and keep only four external events at a time, all 100.00% running. Scoped steady-loop cycles/instructions above are separate from these whole user-process groups, which include pre-timing load/hash/construction/output and exclude kernel:
| Combined PMU event / source sample | Baseline | Candidate |
|---|---|---|
| Branches | 66.709 | 68.337 |
| Branch misses | 2.20681 | 0.37400 |
| L1 data refill/miss alias | 1.28439 | 0.19164 |
| Data-TLB refill/miss alias | 0.0813949 | 0.00186921 |
Branch count is slightly higher while mispredictions are much lower; fewer instructions alone do not explain the gain. FIR-only packing changes refill aliases 215,613,773 → 46,712,240 (L1) and 13,658,947 → 336,352 (dTLB) over 167,771,940 source samples, with essentially unchanged instructions/ branches. That supports locality, while its measured whole gain is only 0.9%. Linux generic L1/dTLB access denominators alias the same hardware event here (identical totals), not independent load-only traffic measurements. Refill events include speculation/other process memory; no claim of exact stall cycles, cache latency or independently measured TLB access rate. Per-stage PMU attribution, energy, GPU counters and large-address memory operation remain unqualified.
Remaining opportunities — not implemented¶
- Per-component DSP SIMD/locality, not tap-reduction SIMD: demod/SRRC and channelizer together are about 0.282 CPU seconds/source second in the sampled combined replay. Parallel I/Q work could help while retaining each component's operation/contraction order. Whole benefit is unmeasured; changed reductions or relaxed FMA are outside this authorization. Require same-target bitwise intermediate/weak-signal, native/memory64 and portability qualification.
- Viterbi survivor/addressing detail: roughly 0.109 normalized CPU/source remains. Profile traceback versus survivor writes separately before proposing packing/vectorization. Any representation change needs the unchanged scalar step/tie/final/pre-FEC comparisons, memory/reset checks and separate approval.
- Timing/record/diagnostic overhead: timing is about 0.044 normalized CPU/ source; record/SHA/JSON work is small in this profile (SHA setup is included). Measure exact invocation/copy cost first. No fewer control-loop updates, skipped validation, delayed diagnostics or altered record identity are implemented/proposed as shortcuts. New transport/storage remains Phase 4+.
These are ranked investigation ideas, not measured savings or additional authorization. No speculative percentage gains are added to the retained patch.
Reproduce¶
Repository-root commands; generated sources/artifacts/payloads remain ignored:
core_dir=components/rfsentry-core
baseline_dir="$core_dir/build/cpu-efficiency-baseline-source"
mkdir -p "$baseline_dir"
git archive d61cf03de9e0a528e0b606cfa40d81a8306281e3 \
components/rfsentry-core components/rfsentry-api components/rfsentry-decoder \
| tar -xf - -C "$baseline_dir"
components/rfsentry-core/tools/test.sh
export VITERBI_BASELINE_ROOT=build/cpu-efficiency-baseline-source
export VITERBI_EVIDENCE_DIR=build/cpu-efficiency-evidence/combined
./scripts/node-toolchain.sh run node "$core_dir/tools/viterbi-benchmark.mjs" build
VITERBI_PAIRS=5 VITERBI_REPEATS=8 ./scripts/node-toolchain.sh run node \
"$core_dir/tools/viterbi-benchmark.mjs" measure
./scripts/node-toolchain.sh run node "$core_dir/tools/efficiency-counters.mjs"
./scripts/node-toolchain.sh run node "$core_dir/tools/viterbi-benchmark.mjs" profile
Weak/varied preparation and focused checks:
./scripts/node-toolchain.sh run node "$core_dir/tools/weak-fixtures.mjs"
./scripts/node-toolchain.sh run node "$core_dir/tools/efficiency-workloads.mjs"
"$core_dir/build/native/dsp_cases" "$core_dir/build/cpu-efficiency-evidence/varied"
CORE_DSP_FIXTURE_DIR=build/cpu-efficiency-evidence/varied \
./scripts/node-toolchain.sh run node "$core_dir/tools/local-browser.mjs" dsp-audit
nice -n 10 make -C components/rfsentry-api -j1 test test-iq-queue-sanitize all integration
Supply VITERBI_INPUT, VITERBI_WORKLOAD, VITERBI_RESULT_NAME for the
generated production/varied workloads; CORE_BASELINE_DIR selects the frozen
memory64 artifact for browser benchmarking. Both directions of record/BigInt
interchange remain in the full suite. Production-rate native replay:
VITERBI_INPUT=build/weak-2400000-strong.cu8 \
VITERBI_WORKLOAD=build/cpu-efficiency-evidence/varied/2400000-strong.json \
VITERBI_RESULT_NAME=native-production VITERBI_PAIRS=5 VITERBI_REPEATS=8 \
./scripts/node-toolchain.sh run node "$core_dir/tools/viterbi-benchmark.mjs" measure
Generate workload expectations only from a passed exact differential report:
efficiency-workloads.mjs build/dsp-native.jsonl; longer expectations use the
varied differential report. Native and browser expectations are target-specific,
never relaxed comparisons. Freeze the baseline receive module by building the
archived d61cf03 core with the same pinned SDK/flags, then copying its
receive.wasm/descriptor to build/cpu-efficiency-evidence/baseline-receive.*.
The baseline hash must match the table. Browser command:
CORE_BASELINE_DIR=build/cpu-efficiency-evidence \
./scripts/node-toolchain.sh run node "$core_dir/tools/local-browser.mjs" receive-benchmark
Supply the same VITERBI_INPUT/VITERBI_WORKLOAD for browser production replay;
CORE_BENCH_PAIRS=5 CORE_BENCH_REPEATS=1 for the longer input. Current tool
defaults preserve the older scalar report's reproducibility; this follow-up
always specifies the optimized baseline paths explicitly. Isolated source/
binary snapshots remain in ignored build/cpu-efficiency-evidence/{viterbi,srrc,fir};
their hashes are in the durable evidence. Recreate each from d61cf03 with only
its corresponding production region changed; never stack before measuring it.
Recorded live comparison (2026-10-11)¶
Baseline: 20:03:29–20:13:30 UTC, 601.151 seconds. Candidate: 20:19:29–20:29:30 UTC, 601.402 seconds.
The metadata-only soak loads no live payloads and performs no IQ subscription. No build, benchmark or profiling load ran during these comparison intervals.
| Live interval measurement | Accepted baseline | Combined candidate |
|---|---|---|
| Decoder CPU seconds / wall second (one core = 1.0) | 0.583149 | 0.462963 |
| Validated frames | 34,012 | 34,027 |
| Corrected symbols, lanes 0/1/2/3 | 26 / 29 / 48 / 368 | 26 / 39 / 43 / 348 |
| Rejected frames, ASM / RS | 0 / 1 | 0 / 0 |
| Lock losses / reacquisitions | 0 / 0 | 0 / 0 |
| Source gaps / known missing samples | 0 / 0 | 0 / 0 |
| IPC overflow blocks / samples | 0 / 0 | 0 / 0 |
| Output failures | 0 | 0 |
| Worker RSS / locked / swap, KiB | 7,824 / 7,984 / 0 | 7,824 / 7,984 / 0 |
| Sampled queued bytes maximum, 2,097,152-byte capacity | 246,528 | 246,528 |
| Sampled last processing lag maximum, ms | 22.500 | 20.523 |
Live decoder CPU is 20.610% lower, consistent with the matched production-rate replay. Gateway CPU is about 0.0102 → 0.0098 cores, RSS 26,144 KiB throughout. Temperature 63.7–66.4 °C, clocks about 2.400 GHz, throttling 0x0 in both intervals. Records are all counted and deliberately disposed by the existing bounded internal sink; zero output failures does not claim reliable delivery. API diagnostics/counters advance and remain current with stable configuration/run identities during each interval.
Queue/last-lag maxima are 30-second samples, not continuous high-water data. Diagnostic lifetime peak block/lag are 29.678/63.532 ms → 11.142/32.383 ms, but the accepted decoder has a much older run: those are not comparable interval percentiles. Use the matched recorded-workload p50/p95/p99/max tables for spike qualification. Different live RF intervals and slightly different durations explain different frame/correction counts; they do not prove better sensitivity or bit equivalence. Those guarantees come from the frozen-input differential tests.
```sh sudo ./scripts/node-toolchain.sh run node \
The announced decoder restart separately lost 376 overflow blocks / 12,320,768 source samples (5.134 seconds at 2.4 MS/s). Zero gaps in the later run do not erase that loss. Configured memory-cgroup limits were not effective on that measurement host; observed kernel locks and swap values are separate evidence.
The source-specific hosted runs 38170599492 and 38170599487 passed on
9f77ecaae17aab39b6bb313b41cade2beefaf7b4. Operator acceptance of this follow-up
was pending at recording; later Server progression is owned by the roadmap.
4-GiB execution, wider browsers and aggregate multi-stream capacity remain separate gates. See supporting data.