JumpStation targets a continuous spectrum of AI compute, from sub-milliwatt microcontrollers to 45-TOPS production edge systems. Every device in the spectrum is a first-class citizen: the same toolchain and JumpBundle format that targets a Pico also programs a JumpModeler Turbo. This is the integral ecosystem — you could prototype a JumpModeler Turbo application on an UNO Q.
The device spectrum, ordered by compute class:
| Device | On-chip AI | RAM | Role |
|---|---|---|---|
| Arduino UNO (ATmega328P) | — | 2 KB | Ultra-constrained embedded, TFLite Micro |
| Raspberry Pi Pico | — | 264 KB | Sensor inference, TFLite Micro |
| ESP32 | — | ~520 KB | Wireless edge AI nodes |
| Arduino UNO Q | Adreno GPU + QRB2210 AI | 4 GB | Minimum Linux/UI platform, shield-compatible |
| JumpStation (CM5 / Pi5) | CPU only | 4–8 GB | Development host, GPIO testbed |
| JumpModeler Junior (RK3588S2) | 6 TOPS Rockchip NPU | 4–8 GB | RK dev host, native NPU testbed |
| JumpStation Turbo (CM5 + DX-M1) | 25 TOPS | 4–8 GB | Primary profiling & distillation engine |
| JumpStation Turbo RK (RK3588S2 + DX-M1) | ~31 TOPS | 4–8 GB | Max-throughput profiling & distillation |
| JumpModeler (Orion O6) | 29 TOPS NPU | 8 GB | Production edge AI |
| JumpModeler Turbo (Orion O6 + DX-M1) | 45 TOPS | 16 GB | High-performance production edge AI |
The JumpStation chassis uses a swappable compute module. The same WaveShare carrier board accepts either a Raspberry Pi CM5 or a Rockchip RK3588S2 module. Combined with the optional DX-M1 M.2 accelerator, this yields four benchmarkable configurations:
| Configuration | On-chip AI | M.2 Slot |
|---|---|---|
| JumpStation (CM5) | CPU only | Free for NVMe |
| JumpModeler Junior (RK3588S2) | 6 TOPS | Free for NVMe |
| JumpStation Turbo (CM5 + DX-M1) | 25 TOPS | DX-M1 |
| JumpStation Turbo RK (RK3588S2 + DX-M1) | ~31 TOPS | DX-M1 |
Running the same profiling job across all four configurations quantifies the acceleration benefit of each tier for a given workload — informing both hardware purchasing decisions and application targeting.
The base development platform. Raspberry Pi CM5 in a WaveShare carrier.
| Attribute | Value |
|---|---|
| SoC | Raspberry Pi CM5 (quad-core Cortex-A76, 2.4GHz) |
| RAM | 4 GB or 8 GB LPDDR4X |
| Storage | eMMC 32GB + M.2 NVMe slot |
| OS | Raspberry Pi OS (64-bit) + JumpStation stack |
| GPIO | 40-pin HAT-compatible + CM5 expansion |
| Connectivity | Gigabit Ethernet, WiFi 5, BT 5.0, USB 3.0 |
| Accelerator | None (CPU inference only) |
| Power | WaveShare PSU, USB-C PD |
| Form factor | WaveShare CM5 carrier board + enclosure |
The JumpStation base is the prototyping and testbed platform. Its GPIO connects directly to every device in the spectrum below it (UNO Q, Pico, ESP32, UNO), allowing the developer to wire real sensors and actuators while developing in a full Linux environment.
Device profile: devices/jumpstation/
RK module variant. Rockchip RK3588S2 in the same WaveShare carrier — adds a 6 TOPS on-chip NPU.
| Attribute | Value |
|---|---|
| SoC | Rockchip RK3588S2 |
| CPU | 4× Cortex-A76 @ 2.4 GHz + 4× Cortex-A55 @ 1.8 GHz |
| GPU | Mali-G610 MP4 |
| NPU | 6 TOPS INT8 (on-chip Rockchip NPU) |
| RAM | 4 GB or 8 GB LPDDR4X / LPDDR5 |
| Storage | eMMC 32 GB + M.2 NVMe slot |
| OS | Debian/Ubuntu ARM64 + JumpStation stack |
| GPIO | 40-pin HAT-compatible + expansion header |
| Connectivity | Gigabit Ethernet, WiFi 6, BT 5.0, USB 3.0 |
| Power | WaveShare carrier PSU, USB-C PD |
| Form factor | WaveShare carrier board + enclosure (shared chassis) |
The JumpStation RK is the RK3588S2 native development and testbed platform. Its on-chip 6 TOPS NPU enables validated INT8 inference without an external accelerator — important when the production deployment target is also RK-silicon. Swap the module from CM5 to RK3588S2 in the same chassis for direct performance comparison.
Device profile: devices/jumpstation_rk/
Raspberry Pi CM5 + DX-M1 M.2 neural accelerator. The primary profiling and distillation platform.
| Attribute | Value |
|---|---|
| SoC | Raspberry Pi CM5 (quad-core Cortex-A76, 2.4GHz) |
| RAM | 4 GB or 8 GB LPDDR4X |
| Storage | eMMC 32GB + M.2 slot occupied by DX-M1 |
| OS | Raspberry Pi OS (64-bit) + JumpStation stack |
| GPIO | 40-pin HAT-compatible + CM5 expansion |
| Connectivity | Gigabit Ethernet, WiFi 5, BT 5.0, USB 3.0 |
| Accelerator | DX-M1 M.2 — 25 TOPS INT8 |
| Power | WaveShare PSU, USB-C PD |
| Form factor | WaveShare CM5 carrier board + enclosure |
The Turbo is the targeting and distillation engine for the CM5 module configuration. The DX-M1 accelerates profiling, quantization simulation, and the INT8 distillation training loop.
Device profile: devices/turbo/
Maximum-throughput variant. Rockchip RK3588S2 (6 TOPS) + DX-M1 M.2 (25 TOPS) in the same chassis.
| Attribute | Value |
|---|---|
| SoC | Rockchip RK3588S2 |
| CPU | 4× Cortex-A76 @ 2.4 GHz + 4× Cortex-A55 @ 1.8 GHz |
| On-chip NPU | 6 TOPS INT8 (Rockchip NPU) |
| M.2 Accelerator | DX-M1 — 25 TOPS INT8 |
| Combined AI | ~31 TOPS INT8 |
| RAM | 4 GB or 8 GB LPDDR4X / LPDDR5 |
| Storage | eMMC 32 GB (M.2 slot occupied by DX-M1) |
| OS | Debian/Ubuntu ARM64 + JumpStation stack |
| GPIO | 40-pin HAT-compatible + expansion header |
| Power | WaveShare carrier PSU, USB-C PD |
| Form factor | WaveShare carrier board + enclosure (shared chassis) |
The Turbo RK combines native RK3588S2 silicon profiling with DX-M1 supplemental acceleration. The on-chip NPU and DX-M1 can be scheduled for different pipeline stages simultaneously — NPU for inference benchmarking, DX-M1 for the distillation training loop — yielding the highest combined throughput in the JumpStation lineup.
Device profile: devices/turbo_rk/
A child-first computing experience.
| Attribute | Value |
|---|---|
| Display | 320×240, 16-bit color, TFT LCD (impact-resistant) |
| Input | Large physical buttons, simple D-pad, no touchscreen |
| CPU Class | ARM Cortex-A (Linux-capable) |
| RAM | 256 MB |
| Storage | Built-in eMMC, 4GB |
| Power | 2000mAh Li-Po battery + durable USB-C |
| Connectivity | USB-C only (no wireless by default) |
| OS Layer | Linux + JumpStation launcher (restricted mode) |
The Kidputer runs a restricted launcher profile. Only content rated E is surfaced by default. Parental oversight tools are built in at the OS level.
Target audience: Children ages 5–12.
Device profile: devices/kidputer/
A retro-gaming-focused form factor.
| Attribute | Value |
|---|---|
| Display | 320×240, 16-bit color, TFT LCD |
| Input | Arcade-style buttons, 8-way joystick, 6-button layout |
| CPU Class | ARM Cortex-A (Linux-capable) |
| RAM | 512 MB |
| Storage | MicroSD, 8GB+ |
| Power | Internal battery + USB-C |
| Connectivity | WiFi, USB-C |
| OS Layer | Linux + JumpStation launcher (game-optimized mode) |
The Jumpcade launcher prioritizes game bundles and low-latency input handling. UI chrome is minimized. Input polling is handled at high frequency.
Target audience: Retro gaming enthusiasts, hackers.
Device profile: devices/jumpcade/
Microcontroller-class embedded devices.
| Attribute | Value |
|---|---|
| MCU | RP2040 (dual-core ARM Cortex-M0+) |
| Display | Varies by configuration (OLED, e-ink, none) |
| Input | GPIO pins, physical buttons (configuration-dependent) |
| RAM | 264 KB SRAM |
| Storage | 2MB flash (on-chip), optional external SPI flash |
| Power | USB (5V) or 3.3V regulated supply |
| Connectivity | USB, UART, I2C, SPI, PIO |
| OS Layer | MicroPython or bare-metal C |
Pico devices run headless or near-headless workloads. JumpBundles for Pico are firmware images and MicroPython scripts rather than Python packages. The flashing workflow is handled by core/flashing/pico_flasher.py.
Device profile: devices/pico/
Wireless embedded AI node.
| Attribute | Value |
|---|---|
| MCU | ESP32 (dual-core Xtensa LX6, up to 240MHz) |
| RAM | ~520 KB SRAM |
| Flash | 4MB (typical) |
| Connectivity | WiFi 802.11 b/g/n, BT 4.2, UART, I2C, SPI |
| Power | USB (5V) or 3.3V |
| OS Layer | MicroPython or ESP-IDF (bare-metal C) |
| Inference | TFLite Micro, ESP-IDF ML extensions |
The ESP32 extends the Pico class with WiFi, enabling distributed edge AI nodes that report to a JumpStation host. Bundles targeting ESP32 are firmware images flashed via core/flashing/esp32_flasher.py.
Device profile: devices/esp32/
Minimum Linux/UI platform in the JumpStation spectrum. Dual-brain: QRB2210 MPU + STM32U585 MCU.
| Attribute | Value |
|---|---|
| MPU | Qualcomm Dragonwing QRB2210 (quad-core 2.0 GHz) |
| MCU | STM32U585 (real-time ARM Cortex-M33) |
| GPU | Qualcomm Adreno (integrated) |
| AI | On-chip AI/GPU acceleration (QRB2210) |
| RAM | 4 GB LPDDR4 |
| Storage | 32 GB eMMC (built-in) |
| OS | Linux Debian (MPU) + Arduino sketch (MCU) |
| Connectivity | WiFi 5 dual-band (2.4/5 GHz), Bluetooth 5.1 |
| Headers | Classic UNO shields + high-speed headers |
| Expansion | Qwiic connector, USB-C (PD, video) |
| Inference | TFLite, ONNX Runtime, PyTorch Mobile |
The UNO Q is the first Linux-and-AI-capable device in the Arduino form factor. Its shield compatibility allows the full UNO hardware ecosystem to run on it, while the MPU side runs JumpBundles, Python, and on-device AI inference. The MCU handles real-time I/O in parallel.
User-facing applications (applications with a UI) floor at the UNO Q — it is the minimum platform that can run the JumpBundle launcher and display output.
Device profile: devices/uno_q/
Ultra-constrained embedded inference. The engineering floor of the ecosystem.
| Attribute | Value |
|---|---|
| MCU | ATmega328P (8-bit AVR, 16MHz) |
| RAM | 2 KB SRAM |
| Flash | 32 KB program flash |
| Connectivity | UART, I2C, SPI, GPIO |
| Power | USB (5V) or 7–12V barrel |
| OS Layer | Bare-metal C (Arduino framework) |
| Inference | TFLite Micro (extremely constrained models only) |
The classic UNO targets the most constrained inference workloads: keyword spotters, threshold classifiers, and signal filters that fit in under 1 KB of RAM. It has no OS, no display, and no networking — pure deterministic embedded execution. Bundles targeting uno are Intel HEX images flashed via core/flashing/uno_flasher.py using avrdude.
Device profile: devices/uno/
Production edge AI platform. 12-core ARM + 29 TOPS NPU.
| Attribute | Value |
|---|---|
| SoC | Cix CD8180 (Cortex-X925 + A725 + A520) |
| CPU | 12-core ARM (4× Cortex-X925 + 4× A725 + 4× A520) |
| NPU | 29 TOPS INT8 (Cix NPU) |
| GPU | Immortalis-G720 |
| RAM | 8 GB LPDDR5 |
| Storage | NVMe M.2 2280 (256 GB+) |
| OS | Linux (64-bit) |
| Connectivity | 2.5GbE, WiFi 7, BT 5.4, USB4, PCIe 4.0 |
| Inference | ONNX Runtime + Cix NPU SDK |
The JumpModeler is the primary production edge AI target. 29 TOPS handles most vision models, small language models, and real-time multi-modal workloads. For workloads that exceed its envelope, the JumpModeler Turbo adds the DX-M1 to reach 45 TOPS.
Device profile: devices/orion_o6/
Flagship production edge AI platform. JumpModeler + DX-M1 expansion = 45 TOPS.
| Attribute | Value |
|---|---|
| SoC | Cix CD8180 (Cortex-X925 + A725 + A520) |
| CPU | 12-core ARM (4× Cortex-X925 + 4× A725 + 4× A520) |
| On-chip NPU | 29 TOPS INT8 (Cix NPU) |
| M.2 Accelerator | DX-M1 — 16 TOPS INT8 |
| Combined AI | 45 TOPS INT8 |
| GPU | Immortalis-G720 |
| RAM | 16 GB LPDDR5 |
| Storage | NVMe M.2 2280 (256 GB+) |
| OS | Linux (64-bit) |
| Connectivity | 2.5GbE, WiFi 7, BT 5.4, USB4, PCIe 4.0 |
| Inference | ONNX Runtime + Cix NPU SDK + DX-M1 SDK |
The JumpModeler Turbo is the highest-capability target in the JumpStation catalog: 45 TOPS of combined on-device neural compute. It handles large vision models, 7B–70B language models (at appropriate quantization), and multi-modal inference workloads. JumpBundles targeting the Turbo are standard Python packages with ONNX or NPU-SDK weights; deployment is a package install.
Device profile: devices/orion/
Each device directory contains:
devices/<name>/
├── profile.json # Machine-readable capability declaration
└── README.md # Human-readable device notes
The profile.json is consumed by the target selector during hardware matching, by the JumpBundle runtime for compatibility checking, and by the Silhouette pipeline for asset normalization. Key fields:
{
"id": "turbo",
"class": "linux",
"cpu_cores": 4,
"cpu_arch": "cortex-a76",
"ram_mb": 4096,
"accelerator": {
"name": "DX-M1",
"tops_int8": 25,
"precision": ["int8", "fp16"]
},
"inference_backends": ["tflite", "onnxruntime", "dx-m1-sdk"],
"storage_mb": 32768,
"connectivity": ["wifi", "ethernet", "bluetooth", "usb"],
"gpio": true,
"os": "linux"
}
A formal JSON Schema for device profiles will be added in a future iteration alongside the targeting suite.