Thermal sensors such as the FLIR Boson, Lepton, and Tau families stream 14-bit or 16-bit pre-AGC data where each code value is linearly proportional to incident radiance, and therefore to scene temperature once the calibration curve is applied [S2]. A standard 8-bit display path can only resolve 256 greyscale steps, so the same scene must be remapped, equalized, or clipped before it ever reaches a monitor [S2][S4].
For predictive maintenance, R&D test cells, gas-detection overlays, and any quantitative workflow, the 14-bit radiometric stream is the engineering source of truth; the 8-bit AGC stream is the visualization layer that sits on top of it. The two are not interchangeable, and treating them as such is where measurement programs quietly lose accuracy.
Bit depth, code count, and what the numbers actually mean
A 14-bit ADC produces 2^14 = 16,384 distinct code values per pixel; a 16-bit ADC produces 65,536 codes. By comparison, an 8-bit display path holds only 256 codes, which is why a 14-bit radiometric frame, when naively rendered on an 8-bit monitor, looks flat, low-contrast, and often "grey on grey" [S2]. A single 14-bit image therefore carries roughly 64x more amplitude resolution than the 8-bit frame derived from it.
In a still-photography context the same scaling logic applies: a 14-bit RAW file holds up to 4 trillion shades versus 68 billion for 12-bit RAW, and the extra codes primarily buy headroom for heavy shadow recovery and aggressive push-processing in post [S1]. In thermography the analogous "push-processing" is the temperature calibration step, where a 0.1 K or 0.01 K delta must round-trip through the ADC without quantisation-induced drift.
Pre-AGC radiometric vs post-AGC display: the two pipelines
FLIR's OEM documentation states that "pre-AGC data is typically used in applications because the data is linearly proportional to temperature in radiometric cameras," and that in-camera AGC is typically more advanced than the simple linear remap used by viewer software [S2]. A simple linear AGC stretches the lowest and highest measured values across the full 0 to 255 range, which is a non-linear, scene-dependent transform that destroys the radiometric mapping the sensor was calibrated for.
Histogram equalization, plateau equalization, and information-based AGC variants go further by redistributing codes according to the histogram of the current frame, which improves contrast for human eyes but invalidates any temperature lookup that assumes a linear code-to-temperature relationship [S2][S4]. The Axiom Optics primer frames AGC as a necessity, not a feature: "the raw data from a thermal sensor (often 14-bit) represents a wide range of temperatures. Since standard displays (like a monitor or TV) can only show 8-bit..." the dynamic range has to be compressed for viewing [S4]. That compression is the loss channel for any downstream measurement.
Selection criteria: which stream fits which application
For temperature measurement, alarms, trending, and ML training data, specify 14-bit (or 16-bit) pre-AGC radiometric output over USB, Ethernet, or CSI, and budget for a host-side pipeline that ingests Y16 or equivalent monochrome 16-bit frames [S2]. Storage scales with bit depth: a 640 x 480 radiometric stream at 14 bits runs at roughly 640 x 480 x 2 bytes x frame_rate, so a 9 Hz Boson generates about 5.5 MB/s, versus roughly 0.3 MB/s for the 8-bit display stream.
For driver-view, security monitor, search-and-rescue hand-off, or any feed where a human is the consumer, 8-bit AGC is appropriate and is what the camera's on-board processing is tuned to deliver. For mixed use, request both streams simultaneously (e.g. FLIR Boson exposes a radiometric Y16 side-channel alongside the AGC video), so analytics see linear codes and operators see a tunable image [S2].
Comparison matrix: 14-bit raw vs 16-bit raw vs 8-bit AGC
Across the three common thermal data paths, 8-bit AGC wins on bandwidth, hardware cost, and human readability; 14-bit raw wins on the balance of radiometric fidelity, file size, and host compute load; 16-bit raw wins on dynamic range and on systems that need the extra codes for very wide scene temperatures or sub-Kelvin NETD chains. Thermal imaging systems offer a trade between 16-bit and 14-bit pre-AGC data output, but both have too many shades of grey to be displayed directly on a screen and are typically pre-processed and converted into 8-bit greyscale values for display [S2].
On a decision axis, choose 8-bit AGC for: human viewing, embedded displays with no host compute, low-bandwidth radio links, and demonstrations. Choose 14-bit raw for: radiometric thermography, NDT, building envelope audits, condition monitoring, ML training, and any temperature alarm that has to survive post-processing. Choose 16-bit raw for: scientific cameras (LN2-cooled MWIR InSb, nBn), radiometric calibration labs, and high dynamic range scenes with both hot and cryogenic targets in frame [S2].
Limits, failure modes, and integration gotchas
Three traps catch first-time integrators. First, naive file viewers (ImageJ, IrfanView) apply a default linear stretch that visually rescues 14-bit pre-AGC data but discards the radiometric calibration; users see an image, not a temperature [S2]. Second, HART and other analog fieldbus protocols are not in this comparison; thermal sensor digital links are a separate domain from process instrumentation buses, and treating them as substitutes breaks both pipelines. Third, frame rate and bit depth trade against each other on the same USB or Ethernet link, so specifying 16-bit at 60 Hz requires more bandwidth headroom than 14-bit at 9 Hz, and on shared buses this can starve other instruments on the same pressure transmitter network segment.
Fourth, NETD (noise-equivalent differential temperature) of the sensor sets the floor on usable codes; a 50 mK NETD sensor fed into a 14-bit ADC has only the codes that fall above the noise floor carrying real temperature information, so a 16-bit upgrade on a noisy uncooled core yields diminishing returns. Cooled MWIR cameras with sub-10 mK NETD benefit more from the wider code space than entry-level uncooled LWIR cores. Fifth, the radiometric calibration curve (gain/offset, sometimes per-pixel non-uniformity correction) is typically applied on the host or in-camera after the raw stream; if a host discards the calibration metadata the 14-bit stream degrades to a 14-bit image, not a 14-bit temperature map.
Standards, sourcing, and reference points
FLIR's own OEM knowledge base states that the 16-bit or 14-bit pre-AGC data from Boson, Lepton, and Tau is "linearly proportional to temperature in radiometric cameras" and that the camera-side AGC defaults to histogram equalization rather than a linear remap [S2]. The Axiom Optics AGC primer is the cleanest public definition of why 14-bit raw maps to 8-bit display, and it is the most direct non-vendor confirmation that the compression step is the root of the contrast and dynamic-range loss [S4]. The Photography Life RAW article is the generic still-image analogue that documents the 12-bit vs 14-bit code-count math and the post-processing headroom argument, which carries over directly into how much you can push thermal codes in software [S1].
Watch the OEM firmware release notes on Boson and Lepton for any change in pre-AGC bit depth, packet format, or Y16 packing; any of those will break host-side parsers and should trigger a regression test of the radiometric pipeline. For new system designs that will live alongside process instrumentation, plan the flow meter and industrial valve networks on physically separate Ethernet segments from the thermal imaging link, since mixing high-rate 16-bit video with time-critical control traffic on the same unmanaged switch is a common source of dropped radiometric frames. Finally, for adjacent spec-driven comparisons in the SourceBySpec catalog, see the 0.5-4.5 V ratiometric vs 4-20 mA pressure transmitter write-up and the SIL3 2oo3 safety PLC architecture piece; both follow the same source-vs-display separation that distinguishes 14-bit raw from 8-bit AGC video [S1][S2][S4].