BITECH Engineering Insights

TOPS is a peak-throughput number, not a system spec. What actually governs real-world performance is what the compute engine talks to and how — camera ingress (MIPI vs USB vs GMSL2), memory bandwidth, the toolchain your team already uses, and the thermal envelope you can build around it.

An RK3588 platform quoted at 32 TOPS is actually 6 TOPS onboard plus a 26 TOPS M.2 accelerator — two engines, two toolchains. A Jetson Orin NX at 157 TOPS is a single unified device. Neither number tells you whether the platform fits your project.

Match architecture to workload first. TOPS is the last number you check, not the first.

Full framework: Jetson Orin vs RK3588 vs x86 →

Three specific triggers, not one:

• Distributed cameras — more than two cameras cabled across a moving chassis. GMSL2 ingress on Jetson (4 lanes over coax, ~15 m reach) solves what MIPI or USB cannot.
• Toolchain match — your perception team already lives in CUDA, TensorRT or Isaac ROS. RKNN conversion for RK3588 adds real engineering time your Jetson project doesn't pay.
• Model iteration frequency — if you'll update the model monthly, Jetson's toolchain overhead is lower than RK3588's quantization loop.

If none of the three apply, RK3588 (AE-3588LBT or AE-3588BT) delivers the same job at lower unit cost and lower power.

Full comparison: RK3588 vs Jetson Orin for Multi-Camera Vision and AMR →

Match four things in this order — before you look at the processor:

1. IP rating — IP65 for indoor washdown or factory floor; IP66 for outdoor with driven rain; IP69K only for high-pressure cleaning (food, pharma).
2. Touch technology — PCAP for glove-free multi-touch and glass overlay; resistive when operators wear thick gloves or the interface sees liquid pooling.
3. Mounting — panel-cut, VESA, or open-frame. This determines your enclosure design, not the panel PC's.
4. Screen size — driven by operator distance and content density.

CPU class only matters after the above are locked. A wrong IP rating kills a deployment; a slightly slower CPU is a firmware problem.

Deep dive: Industrial Panel PC Selection Guide →

SOM + carrier (AX series: AX-130BT / AX-134BT) wins when:

• Your program runs 5+ years and you want to survive CPU generation changes without re-certifying the enclosure
• You need a custom I/O layout that doesn't exist on any catalog SBC
• You're integrating for a customer who runs "Copy Exactly" audits

Single-board (AE series: AE-3588 line) wins when:

• Space and unit cost drive the design (AGV fleets, consumer-adjacent robotics)
• The compute workload is stable and won't need a mid-program upgrade

Both are backed by Fixed BOM Control and formal PCN. The difference isn't reliability — it's what happens when the silicon roadmap moves under you.

Related: Lifecycle Management →

Isolated I/O is required when signals cross ground references or touch the outside world — motor drives, long cable runs, vehicle power buses, high-EMI floors with variable-frequency drives, or any deployment where your device shares ground with something you don't control.

Non-isolated I/O is fine when the deployment lives in a controlled cabinet with a single grounded PSU, cables under 3 m, and no switching loads nearby. Isolation costs money and adds propagation delay — don't pay for what you don't need.

BITECH platforms treat isolation as a configured option, not a fixed feature: AE-3588BT ships isolated as standard, AE-3588LBT non-isolated, AX-130BT configurable per project.

Full guide: Isolated vs Non-Isolated I/O →

4000V isolation is a withstand voltage — the level a 1-minute dielectric test verifies without breakdown. It is NOT a continuous working voltage. Working voltage on the same barrier is typically well below 1000V.

What it protects against: transient surges, ground potential differences between devices, capacitive coupling from switching supplies, and ESD from field wiring. It does not turn your industrial PC into a medical isolator, and it doesn't replace a proper grounding scheme.

Ask any supplier quoting isolation whether they mean withstand or continuous — the answer separates real engineering from spec-sheet inflation.

Deep dive: Industrial PC Galvanic Isolation →

Classic CAN (up to 1 Mbps, 8-byte payload) is enough for legacy motion control, most AGV drivetrain networks, and any deployment where the bus was designed pre-2015.

CAN FD (up to 8 Mbps, 64-byte payload) becomes worth specifying when:

• You need to move sensor packets (LIDAR, IMU fusion) over the same bus as motion commands
• Your gateway aggregates 3+ subsystems and classic CAN starts to saturate
• Your customer's roadmap includes automotive-grade equipment (most new vehicle buses assume CAN FD)

BITECH platforms use native hardware CAN / CAN FD controllers rather than USB bridges — determinism matters more than raw throughput on a control network.

Related: The 60Ω CAN Bus Check →

GMSL2 wins when cameras are far from the compute (up to ~15 m over coax), the environment is electrically noisy (motor drives, switching supplies), and you need deterministic sync across channels — the AMR case, essentially.

GigE Vision wins in a fixed inspection cell: one or two cameras near the compute box, controlled ground, off-the-shelf interoperability with existing camera libraries.

Neither is universally better. If your cameras are corner-mounted on a moving chassis, GMSL2 (available on Jetson-based AE-NJ60BT) is the answer. If your cameras are on a static machine within a metre of the box, GigE Vision on any x86 platform does the job at lower cost.

Full comparison: GMSL2 vs GigE Vision →

Turnkey OEM and private-label programs start at 100 units minimum. Non-recurring engineering (NRE) is quoted upfront, before you commit — so you know what custom I/O, enclosure and BIOS work costs before the project starts.

Lead times depend on customization depth:

• Standard platforms (AE / AP series) — samples in 10 days
• Custom carrier board — approximately 4 weeks to sample and validate
• Fully custom mainboard engineered to your mechanical envelope — 1 to 3 months from requirements review to first articles

The MOQ is deliberately low. Most catalog-only suppliers won't quote below 500 units; BITECH's SOM + carrier approach makes 100 units economically viable.

Availability depends on the underlying silicon, not marketing claims:

• AE-3588 series (RK3588) — 10+ years, aligned to Rockchip's published production commitment
• AX series (modular x86 SOM) — 7+ years, with CPU generation swap possible without changing the carrier or enclosure
• AE-NJ60BT (Jetson Orin NX) — tracks NVIDIA's published Jetson product lifecycle
• AE-760EBT (x86 + MXM GPU) — shortest window in our range; MXM generations turn over faster than embedded CPUs, and we quote the window explicitly at design-in

All platforms sit on Fixed BOM Control with 6-month PCN notice and a last-time-buy window before any change.

Full policy: Lifecycle Management →

Silent component substitution is the classic failure mode of cheap industrial PC suppliers: the schematic stays the same, but a capacitor, a PHY chip, or a voltage regulator quietly changes between production batches. Your system that was validated six months ago now fails BIOS POST intermittently — and no one can tell you why.

Fixed BOM Control locks the exact component list after DVT. No substitutions without formal PCN. Every unit's build BOM is logged against its serial number, so if a component supplier lot turns out to have a defect, we can identify every unit shipped with that lot and warn the customer before failures start.

This is invisible until you need it — then it decides whether a warranty case takes 2 hours or 2 weeks.

Statistical sampling (5% AQL) catches obvious defects but misses the 1–2% of units that pass QC and die in month three of deployment — the specific failure mode that costs integrators service margin, customer relationships, and weekend investigations.

Every BITECH unit passes an 18-hour bare-board burn-in, a full interface and function test across CAN, serial, Ethernet, display, storage and I/O with real tools (not self-test scripts), then another 18-hour burn-in in its assembled enclosure. Roughly two working days of testing per unit, serial-number traceable.

The second burn-in — after enclosure assembly — catches thermal path errors that a bench test cannot see. Skipping this step is why "cheap industrial PCs" work in the QC bay and die in the field.

Full process: How We Test Every Industrial PC Before Shipment →