Current work.

Vendor tools routed each core in isolation to produce these preliminary counts. Independent AI review inspected the source reports, finite integer checks, register-transfer-level (RTL) simulations and width bounds. RTL describes how values move between circuit registers.

The tested isolation change added 16 logic tables and 48 ordinary registers to P9, which still used eight DSP blocks. Review accepted no useful isolation benefit for that integration.

P9_logic3 still lacks qualified activity-based power, throughput and wider product-fit results.

In a separate three-channel sensing model, 12 readings meet the old balanced 13-reading quality limit, but variance is 8.580858% above the optimized 13-reading plan. A capable ordinary exact planner can match the allocation. In a separate count-and-rail energy example, the combined change uses 8/1943 (0.411734%) more energy than the ordinary lower-rail option under a chosen 25% affected share and separate 1% plus 1% costs.

In the separate software accounting review, FI was slower than capable ordinary gmpy2 in all 26 preserved first-call and warm groups. H22 covers 2,500 cases against the prior FI hybrid.

The matched warm-service comparison includes creating the record, numerical guards, carrying state, decoding, validation, integrity checks and reading the complete record. Setup and priming are reported separately.

An earlier counts-only record reduced traced memory in eight matched groups and passed the timing rule in six; two timing comparisons remained uncertain. That record loses the required ordered journal.

The signal job stores 90,112 raw bytes per frame, then requires encoding, decoding and a full comparison. Even ideal raw capacity needs three parallel work lanes; charging for the actual record raises that to four. Scheduling the work within each frame remains unresolved.

A separate amplifier test used two models with supplied coefficients. Only lag and gain were calibrated on the training capture, then held fixed for validation. Both fixed models failed on the retained validation capture, which had already been seen. The provider’s use of the training, test and validation captures to prepare the coefficients remains unknown. One model had lower aggregate error, stated calculation work and coefficient storage; the other had lower peak error.

Costs against the existing FI baseline and on the target device remain unresolved.

Each row below records one fresh process. Time runs from launch to exit, including startup, fitting, prediction, required records and readback. Peak resident RAM includes Python and the numerical libraries. Retained storage counts the required inputs, complete outputs and work record once. The observing parent process and the equipment that captured the recordings are outside this software job.

The full record contains all 36 settings and 180 process-cost observations. Mean times and maximum memory values summarize separate observations. In the five-block paired analysis, timing remained uncertain in 138 of 396 directed method comparisons. Although the selector chose from early data, a different penalty method achieved lower aggregate error on the later Train and Val points.

Trimming weak directions in the linear fit, or penalizing large fitted coefficients, lowered both errors against the untreated history-model control on each recording. Train and Val still had much higher peaks than cubic. The tested large penalty, with relative strength 0.01, raised aggregate error to about 0.0170–0.0176 and peak error to 0.398–0.476. The supplied history model with calibrated gain also did worse than cubic on Val.

The trimmed fit uses the technical method TSVD with a relative cutoff of 0.001. The history models still store and evaluate 84 coefficient slots even when fewer fit directions are retained.

The pilot fits and chooses a model separately for each recording, with that work included in its costs. In the earlier test, a model fixed on Train failed when transferred to Val over a different scored range. The supplied model’s fitting and capture choices remain unknown. Its sampling descriptions conflict—100 MHz and 200 MHz—and the original evaluation and alignment rules have not been recovered.

These are ordinary fitting methods. A distinct FI implementation and the buyer’s acceptance limits have yet to be defined. This ADL9006 study and the SKY85747-11 radio comparison below use different error measures.

The datasheet lists +12 dBm output with −47 dB error and +16 dBm output with −43 dB error for 802.11ax MCS11 HE80 at 5 V and 25 °C. A job requiring +16 dBm rejects the +12 dBm operating point because it delivers less RF output. The lower RF output is a service loss; supply-power saving remains unquantified.

With the same signal reference, timing, corrections and normalization for all errors, the −47 dB source limit allows at most 0.0026126 added normalized RMS error while retaining a −43 dB total limit. RMS means root mean square, a measure of signal size. The allowance is about −51.6585 dB. It uses the triangle bound to cover the worst alignment between the source and added errors.

At a source error of −43 dB, no guaranteed added-error allowance remains. Adding squared errors would permit a larger allowance only if their cross-term were zero in the same norm. Statistical independence alone does not prove that condition for a finite recording. The datasheet leaves the exact error-measurement setup unresolved, so the common-reference assumption remains necessary.

At +16 dBm average output on ANT, the datasheet’s coupling limits imply −3 to +0.5 dBm average at CPLR, the feedback port, if the output and coupling limits share a power reference, impedance and bandwidth. Receiver headroom remains unresolved: waveform peaks, bandwidth, path loss, gain and linear input limits have not been matched.

Conventional designs can use the same integrated feedback coupler. Current draw has not been matched at these exact output and quality conditions. Feedback reception, conversion, processing and thermal costs remain unquantified.

The software model gives the receiver an ideal 400 MB/s service rate and a 4 MiB queue for records waiting to be delivered. MB means one million bytes; MiB means 1,048,576 bytes.

Once a frame is refused for lack of capacity, the controller stops accepting new frames for the rest of that run. Every later planned frame is counted as refused, while queued records continue to drain whenever service is available. The model tests neither recovery nor renewed admission. The final queue contains accepted data still waiting for delivery when the chosen time window ends.

The ordinary R controller and candidate C follow the same generic rule and tie in every listed case. The successful 2 ms stall result covers one fixed position in the burst schedule.

The source-rate bound uses the admitted range of average clock rates. Physical arrival timing, jitter, rounding allowances and packet/host capacity costs remain open. The software work used generated synthetic records. The +16 dBm and +12 dBm radio-output jobs have separate service definitions.