Calibrating the 12AX7: which datasheet do you believe?

The 12AX7 is the most-used small-signal triode in guitar amplification — it is the gain stage, the cathode follower, and the phase inverter in nearly every circuit in this archive. So when our SPICE model of it mispredicts a stage by a third, that is worth chasing to the bottom. This is the write-up of that chase. The short version: the tube's published data is remarkably consistent across four manufacturers and two continents — except for one chart, which turns out to be the odd one out; and the amplifier voltage chart that started the whole investigation is itself partly to blame.

Background: what the model has to do

Circuit Codex ships a SPICE model for each tube, fitted to the tube's published "average characteristics" and dedicated to the public domain. The model form is the Koren equation (a standard, non-proprietary triode expression); the numbers in it are fitted here from datasheet facts, never copied from another model collection. The method is described in models/METHODOLOGY.md.

The original models are single-anchor fits: each triode's parameters are solved so the model reproduces exactly one published operating point — for the 12AX7, the classic Va = 250 V, Vg = −2 V → Ia = 1.2 mA, gm = 1600 µmho point. That is enough to verify a stage whose plate sits near 250 V, but a 12AX7 gain stage in a tweed amp often runs its plate down at 140–190 V, far below the anchor, and there the model was under-reading plate current badly. On the 5F4 the affected nodes deviate from the printed factory chart by up to ~39 %.

The obvious fix — trace the datasheet's full plate-characteristic curve family and least-squares fit the whole thing — ran into a metrology problem, which is the subject of this study.

The finding that stopped the curve fit

RCA's own 12AX7-A datasheet disagrees with itself. On page 1 it prints a table of typical operation; on page 3 it prints an "average plate characteristics" graph (drawing no. 92CM-6879). Read the graph at the same point the table describes — Va = 250 V, Vg = −2 V — and it gives a plate current well below 1 mA, roughly half the 1.2 mA the table on page 1 states.

Half is not a rounding error. Fitting a model to the graph would move every 12AX7 stage in the corpus in the wrong direction relative to the factory amplifier charts. Before fitting anything, we needed to know which of RCA's two numbers is correct — and that meant going to other manufacturers.

Sources (all datasheets via Frank Philipse's tube-data archive):

• RCA 12AX7-A — /sheets/049/1/12AX7A.pdf (Radio Corporation of America, Harrison NJ, "Data 1, 10-60")
• General Electric 12AX7 — /sheets/093/1/12AX7.pdf (ET-T509B, 6-53)
• Sylvania 12AX7 — /sheets/137/1/12AX7.pdf (Engineering Data Service, September 1955)
• Philips / Mullard ECC83 — /sheets/030/e/ECC83.pdf (Philips, 5-5-1960)

(The RCA sheet is filed on the archive under folder 049 and the GE sheet under 093; each PDF's own letterhead is the authority for its provenance, and that is what is cited here.)

What the tables say

Every manufacturer prints the same two-point "Class A₁ amplifier, each section" typical operation. These are printed numeric values, exact as read:

Four independent data departments, over seven years (1953–1960), agree to the last digit — including a second operating point at 100 V that the RCA table also carries. That second point is the prize: it is a published, low-plate-voltage fact about the tube, exactly in the regime where our model was failing, and it does not come from any amplifier chart.

The tables are also internally consistent: amplification factor µ = gm × rp holds at both points (1250 µmho × 80 kΩ = 100; 1600 µmho × 62.5 kΩ = 100). Nothing in the tabulated data is in tension.

Which chart is the outlier

To decide whether the RCA 92CM-6879 plate graph or the tabulated 1.2 mA is correct, three independent lines of evidence all point the same way.

1. RCA's other graph agrees with the table. The same RCA sheet prints, on page 4, an "average characteristics" graph (drawing 92CM-6880) of rp, gm and µ against grid voltage at Eb = 100, 200 and 300 V. Read at Eb = 100 V, Vg = −1 V it gives rp ≈ 0.080 MΩ, gm ≈ 1250 µmho, µ ≈ 100 (read precision ≈ ±0.005 MΩ / ±100 µmho) — matching the table. Only the 92CM-6879 plate graph on page 3 is out of step, and only in one direction.

2. GE drew its own curves, and they match the table. GE's sheet (093) contains two independently drawn graphs in conventional orientation (Ef = 12.6 V): an average plate-characteristics family (Ia vs Va) and an average transfer family (Ia vs Vg at Eb = 50…300 V). Both are legible and both confirm the table (read precision ≈ ±0.1 mA):

GE's plate-characteristics graph is the same kind of chart as RCA's 92CM-6879, drawn by a different manufacturer, and it lands on 1.2 mA where RCA's lands near 0.6 mA. That is as clean a cross-check as this kind of question allows.

3. Real resistance-coupled operating points agree with the table. Philips' ECC83 sheet prints measured resistance-coupled amplifier tables — actual DC operating points for a stage with a plate-load resistor and cathode bias. For the 100 kΩ-load table, each row's implied plate voltage (Vb − Ia·Ra) and grid bias (−Ia·Rk) give a low-plate operating point, exactly the 5F4's regime:

These currents are consistent with the tabulated behaviour and far above what the 92CM-6879 plate family would predict at the same voltages.

Verdict: the tabulated 1.2 mA / 0.5 mA two-point data is correct, corroborated by four tables, three independently drawn graphs, and one set of measured resistance-coupled points. The RCA 92CM-6879 plate-characteristics graph is the lone outlier — a drafting artefact on one sheet, not a fact about the tube — and it is deliberately not used for calibration. (Reading that graph precisely is also hampered by its rotated axes and coarse scan; there is no need to rely on the exact figure, only on the fact that it sits well below every other source.)

The recalibration recipe

With a trustworthy second point in hand, the fix is a two-point fit rather than a single-anchor solve. Keeping the tube's amplification factor (µ = 100) and the space-charge exponent (EX = 1.5, the 3/2-power law) fixed as before, the fit now frees three parameters — KP, KG1 and the knee term KVB — and solves them so the model reproduces:

• plate current at both tabulated points (0.5 mA at 100 V/−1 V and 1.2 mA at 250 V/−2 V), and
• transconductance at the 250 V anchor (1600 µmho), the point the model test suite checks.

The single freed parameter compared with the original models is KVB: in the single-anchor fit it was a fixed default with negligible effect near 250 V, but it is precisely the term that shapes how current falls away at low plate voltage, so it must be free to honour the 100 V point. The fit is a small deterministic Nelder-Mead solve in the existing pure-Python pipeline; it never touches the disputed plate graph or any amplifier chart (calibrating to the amp charts we later verify against would be circular).

Resulting parameters (12AX7, one section):

How the two models compare against the datasheet and against the independent Philips resistance-coupled points (which were not used in the fit):

The recalibration does exactly what it was meant to: it removes the large systematic under-read at low plate voltage. It does not make everything perfect — the two-point fit slightly over-reads current in the 160–190 V band, and its transconductance at 100 V is still low (an inherent limit of a single-exponent triode expression, which cannot match plate current and transconductance at two widely separated voltages at once). But for DC operating-point work — what this archive verifies — it is a clearly better description of the tube.

What it does to the amplifiers

Rebuilding the model and re-simulating every circuit shows the recalibration helping where the model was genuinely wrong — the low-plate gain stages — and drifting slightly on stages that were already accurate. No circuit that passed before fails after; the automated voltage checks stay green.

Averaged over every chart-valued node in the corpus, mean deviation moves from 8.2 % to 7.9 % — a modest net improvement that understates the real story, which is a large improvement on the worst (lowest-voltage) stages traded against small drift on stages that were already within a few percent.

The amplifier chart that cannot be satisfied

The investigation began with the 5F4, whose second gain stage deviates ~37 %. The recalibration barely moves it (37.1 % → 34.0 %), and that is the most instructive result in the study, because it shows the 5F4 deviation was never mainly a model problem.

The 5F4's printed factory chart shows its V2A stage at a plate of 140 V with a cathode of 2.2 V. A 2.2 V cathode across the stage's 1.5 kΩ cathode resistor means 1.47 mA of plate current. But a 12AX7 biased at Vg = −2.2 V is essentially at cutoff at Vp = 140 V — at that grid voltage the tube needs roughly 280 V on the plate to pass 1.47 mA. The chart's plate voltage and its cathode voltage cannot both be true for a 12AX7: they describe two different operating points.

The self-consistent operating point for that stage — 100 kΩ plate load from ~280 V, 1.5 kΩ cathode bias — sits near 190 V, which is what both the old and new models produce, and which the Philips resistance-coupled data independently supports (at Vp ≈ 189 V the tube passes ~1.1 mA; the single-anchor model is only −4 % off there). The 5F4's phase-inverter nodes are inconsistent in the same way, as its circuit notes already record. No physically valid 12AX7 model can reproduce those printed numbers, because the printed numbers do not describe a physically valid operating point. Those chart values should be treated as disputed, the way the JTM45's phase-inverter cathode reading already is — not chased with model parameters.

Status

Adopted corpus-wide on 2026-07-18. The recalibrated two-point 12AX7 fit is now the shipping model — models/12ax7.inc carries the two-point parameters (KP 679.8, KG1 658.0, KVB 25 690.8) — and every amp in the archive has been re-simulated and re-verified against it, with 0 failures: no circuit that passed under the single-anchor model falls outside its chart tolerance under the two-point one. The low-plate gain stages improve, the already-accurate stages drift only slightly, exactly as the corpus table above lays out.

The record of how that decision was reached is worth keeping, because it was not adopted on the strength of the 5F4 alone. The bar first proposed for the change was that the 5F4's worst 12AX7 stage improve materially with no regressions; the 5F4's worst nodes are its phase-inverter, whose printed chart is internally inconsistent, so they cannot improve and in fact drift slightly. That bar was never met — and the honest conclusion there was that the 5F4 page needs its disputed nodes marked, not a model tuned to hit impossible targets. Those disputed phase-inverter nodes stay marked as disputed; the adoption did not touch them, because they were never a model problem.

What carried the adoption was the separate, corpus-wide question the study handed over: judged on overall accuracy across every circuit rather than on the 5F4, the two-point fit is the better description of the tube — the recipe and data are sound, and it is the correct calibration. That is the basis on which it now ships, replacing the single-anchor models and their documented caveat that they under-read plate current below ~200 V.

Sources are linked inline. Datasheet graph values are stated with their read precision; tabulated values are exact as printed. No factory drawings are reproduced here — only facts read from them. Models and pipeline are CC0; see models/METHODOLOGY.md.