RFD 1128 details: the apparatus, and what each number would mean

The instrument is the card we have, and it is enough for this

An RTX 3090: 24 GB, sm_86, 2020. It is not a 4090, and there is no high-performance inference pipeline on this desk. The machine also carries a Radeon 780M, which is an integrated RDNA3 part sharing system memory with no dedicated VRAM, so it is not a second candidate.

Neither fact blocks this measurement, because tolerance is not a throughput question. It needs enough memory to hold two forms of one stage at once, and it needs identical inputs. 24 GB gives the first and a deterministic renderer gives the second.

Windows reports the 3090 as 4 GB through Win32_VideoController. That is the 32-bit AdapterRAM field overflowing; nvidia-smi reports 24576 MiB. The wrong number is recorded here so nobody re-derives it.

What is compared, and against what

run precision purpose
A bf16 the baseline, on this card, not a paper
B four-bit, DFC host emulation what the device would compute

Same image, same seed, same fov in radians. render_view.py is bit-reproducible at one thread: two runs, identical sha256, measured.

The quantity

Surface deviation between the two meshes, in millimetres, paired with a household object. A penny is 1.52 mm and a credit card is 0.76 mm, so half a penny of error is a sentence somebody can act on, and “the mesh looks fine” is not.

Report a distribution rather than one number: median, 95th percentile and maximum. A mean hides the case that matters, which is a hand or an ear moving while the torso stays put.

What would make this a failure

No threshold is set here, on purpose. The RFD that consumes the mesh should set it. What this produces is the number a threshold can be written against.

Why the DFC and not bitsandbytes

A general quantiser measures a four-bit. The device runs the Dataflow Compiler’s four-bit, with its own calibration and layer rules, and the DFC ships a host emulator that runs that graph. The emulator is slow and exact, which is the right trade when the question is numerics rather than speed. weftspun-hailo-dfc:5.3.0 is the image, built and verified importable on 2026-08-22.