Hardware research: what exists, what to buy, what to avoid

USB selection update: the current product uses two Radxa ZERO 3E hardware boxes and targets USB 2 first. See the agreed USB design. Earlier NanoPi/VirtualHere and USB3 recommendations in this dated study are not the selected design.

Status: live-web research 2026-09-24 (four lanes: commercial endpoints, USB, 4K120 compression path, commodity capture). The verdict these support is in FEASIBILITY.md; the endpoint design in ../design/EDGE-DEVICES.md. Prices were read on 2026-09-24 unless flagged, and move fast; re-check before ordering.

Flags: V verified on a primary source; S search snippet only; I inferred. ⚠ marks a claim that is bench-unverified: code that encodes one must isolate it and its tests must state the assumption (AGENTS.md).

1. Commercial HDMI-over-IP endpoints

4K120

Only one vendor claims it: BZBGEAR XLink-ONE XTREME (launched InfoComm June 2026). HDMI 2.1 FRL5/6 in, 4K120 4:4:4 over 1GbE with a proprietary codec (≥26:1, so not transparent), "<1 frame", HDCP 2.3, USB 3.1 claimed, PoE, 132×75×20 mm, 12 W, $899, "Shipping Q4 2026", unavailable at retail; no public device API (V, product, retail). Watch item: bench one pair when it ships ⚠.

Everything else checked tops out at 4K60 (HDMI 2.0 class): Semtech BlueRiver/SDVoE (newest AVP1000/2000/2000T; the 2026 roadmap has no 4K120), Just Add Power, Crestron NVX, Extron NAV, Lightware, Kramer, Atlona, WyreStorm, AVPro MXNet, Aurora, Matrox, Magewell, Blackmagic 2110, Adder, IHSE Draco, Black Box Emerald (V/S). G&D reaches 4K120 only over dedicated fiber, not switched IP (S). "HDMI 2.1" on a spec sheet (e.g. ASPEED AST1530) does not imply 4K120 (S).

Smallest 10G endpoints (4K60 4:4:4, SDVoE)

SDVoE compresses 4K60 4:4:4 at ~1.3:1 (visually lossless) and sends 4K30 4:4:4 and 4K60 4:2:0 uncompressed; latency <100 µs (V, sdvoe.org). USB is 2.0.

Sources: ZyPer4K-XS datasheet, ZyPer4K-XSE datasheet, AVPro, Aurora.

Control model: who can be routed per device?

Just Add Power (the jap driver's target, ⚠ bench-unverified): every encoder streams continuously to a multicast group derived from its channel number (video 239.92.0.0 base; channel 0100 → 239.92.1.0; audio 239.93.x.x; USB 239.94.x.x). A decoder re-pairs by being told a channel: it IGMP-leaves the old group and joins the new one. API on each device: Telnet port 23 (channel 0100) or HTTP justAPI (POST /cgi-bin/api/command/channel); state persists on the decoder. 802.3af PoE, HDCP 2.2/2.3, USB 2.0 KVM in-band. MaxColor tops out at 4K60 4:4:4 over 1G only (the earlier "10G SFP option" claim is refuted, V, MC-TX1 spec). Sources: multicast switching, single-VLAN switching, Telnet. Risks: command vocabulary per firmware, re-pair glitch duration, in-band USB behavior on channel change ⚠.

2. USB over IP

Hardware extenders and hubs

Sources: Raven manual, Raven price, Aten KX9980T, Digi guide. Black Box Emerald and Adder carry USB 2.0 only.

Software

• VirtualHere (the mesh path). USB3 "including bulk streaming" is fully implemented only for Linux server → Windows client (V, node/3280). Throughput is latency-bound: 347 Mb/s measured by a user; "up to about 700 Mbps" with bulk streaming per the developer (V, node/4243). DisplayLink probably does not use bulk streams, so expect the low end (I). $49 per server, perpetual, hardware-bound, unlimited devices; free trial shares one device (V). arm64 server builds exist. Windows client API: named pipe \\.\pipe\vhclient with LIST, USE,<addr>, STOP USING,<addr>, AUTO USE … (V, client_api). The "Device Client" (network → gadget) is Raspberry Pi 4/5 only, USB2, one device (V).
• Linux usbip. vhci-hcd has supported SuperSpeed since 2017 (V); 10/20 Gb/s devices are rejected pending a June 2026 patch (V); vhci fixes are still landing (I). Windows client: usbip-win2 (V). No modern throughput figures found.
• DisplayLink over USB-over-IP: only a USB2 DisplayLink display (MIMO, 17e9:02ff) is shown working, via VirtualHere in 2019 (V). USB3 DisplayLink over any of these is unproven ⚠.

Gadget side (presenting devices to a PC from a Linux box)

• raw-gadget (mainline since 5.7): no SuperSpeed (issue #61, open since 2023; the uapi takes no SS companion descriptor, V); limited isochronous; cannot emulate a hub, so one UDC = one device identity. usb-proxy (raw-gadget + libusb) is local-only.
• FunctionFS supports SuperSpeed descriptors, so a hand-built configfs/FFS gadget on dwc3 can run at SS, but each cloned device's descriptors must be written by hand (V header, I feasibility).
• SuperSpeed UDCs: RK3588/RK3588S dwc3 OTG0 (USB-C, Gen1) — treat RK3588 boards as having one practical SS UDC; RK3576 Type-A device mode (documented for Android ADB only); Raspberry Pi 5 is USB2 only; CIX P1 is host-only.

Small Linux boards (USB endpoint candidates)

No small ARM board has both native 10GbE and a SuperSpeed UDC.

Sources: R76S, Sige7, NanoPC-T6, ROCK 5C OTG.

3. 4K120 ingest, capture, and compute

Lontium FRL receivers (the only merchant 4K120 ingest)

Merchant FRL-to-pixel silicon is essentially Lontium alone; datasheets, init sequences, firmware, and demo boards are NDA-gated via L2Tek (UK), and no public Linux driver exists. Chips are ~$10–25 (Aug 2026 estimate).

Fallback ingest: Bitec FMC HDMI 2.1 card (€595) + Bitec or AMD HDMI 2.1 IP (quote-only). Bitec's IP offers an optional DSC data path and supports Efinix Titanium (V).

Capture devices

Sources: Elgato 4K X, GC575, SC750N1-L, DeckLink 8K Pro G2. The Elgato 4K X is also the cheapest FRL 4K120 sink instrument.

FPGA boards and codec IP

Zynq US+ EV VCU tops out at an aggregate 4K60, so it cannot do 4K120 (S).

Small x86 boxes (commodity route D)

SVT-JPEG-XS is x86-only (no NEON), so there is no open JPEG XS decoder for ARM or Apple silicon (V). Intel's Media Transport Library does ST 2110-22 with JPEG XS at 4K120 and is validated on E810/E830 NICs (V).

4. The network (UniFi)

Facts from Aug 2026 research that still hold; a 4K120 DSC stream is ~8 Gb/s, so the 10G rules below are tight, not generous.

1. Dedicated AV VLAN; IGMP snooping on; exactly one pinned querier (without periodic queries snooping ages out in ~4 min and streams flood); fast-leave; jumbo frames on every AV port and trunk (one mismatched trunk silently breaks video).
2. Unicast for v1 high-rate flows. A mis-snooped ~8 Gb/s multicast flood takes down every port it reaches. Multicast returns with the many-endpoint mesh, after a flood-containment drill. Unicast must be consumer-initiated at the media layer (the consumer's driver requests the stream from the producer endpoint, as usbip does). Having the producer agent pick a destination would contradict PROTOCOL.md ("producers are passive") and needs an ADR first.
3. Capacity: one 4K120 DSC stream fills ~80% of a 10G link. Keep producer and consumer on one switch, or give inter-switch links more than 10G. Single flows cannot be split across LACP members. Video and D1 USB traffic both flow closet→desk (~8 + ~1 Gb/s), so they cannot share one 10G uplink comfortably.
4. Buffers are shallow (4–8 MB on aggregation switches), so the producer must pace.
5. UniFi Gen 2 switches have documented IGMP snooping defects ⚠; the Netgear M4250 (AV line, IGMP preconfigured) is the escape hatch.
6. UniFi 25G SFP28 exists only on aggregation switches (Pro-Aggregation ~$970 in runout, Pro-XG-Aggregation $2,696). The Enterprise-AV line has 10G access + 100G uplinks, no 25G access (Aug 2026).
7. 10GBASE-R has no mandatory FEC, so the 25G FEC risk flagged in Aug 2026 does not apply at 10G.

5. Shortlists (pending the v1 trade decision in HANDOFF.md)

6. yoke drivers