In Development · TSMC 28nm sign-off in progress
DragonX RXP-1

AI-Native Robotics Processor

The first purpose-built SoC for embodied AI — RISC-V real-time cores, a hardware LETKF sensor-fusion engine, and an on-chip NPU fused on a single TSMC 28nm die. Replaces the discrete MCU + GPU + FPGA stack in humanoid robots, autonomous drones, and collaborative arms with one chip at under 4 W.

12 TOPS
On-chip NPU (INT8)
< 80 µs
Sensor fusion loop
< 4 W
Full-system TDP
RISC-V
RV64GCV + extensions
Die Floorplan

17 mm² of co-designed silicon.

Block-level floorplan on TSMC 28nm HPC+ LVT, FCBGA-400. NoC backbone unifies the NPU, LETKF engine, and four lockstep RT cores around a shared 4 MB SRAM with hardware-arbitrated access.

DragonX RXP-1 die floorplan showing NPU, LETKF engine, four RISC-V real-time cores, shared SRAM, sensor fabric, LPDDR4X PHY, CAN-FD, secure enclave, and PMIC interface on a TSMC 28nm FCBGA-400 die.
NPU / LETKF
RT Cores
Shared SRAM
Sensor Fabric
PMIC Interface
I/O & PHY
Architecture

Everything a robot brain needs. Nothing it doesn't.

Six co-designed subsystems share a 4 MB on-die SRAM fabric with hardware-arbitrated access, eliminating the bus contention that makes discrete multi-chip robot stacks unpredictable.

On-chip NPU

  • Systolic INT8/FP16 matrix engine
  • Hardware paged attention (2-cycle TLB hit)
  • Sparse activation skip — 3× eff. TOPS
  • ONNX/TFLite runtime, MLIR compiler

Hardware LETKF Engine

  • Local Ensemble Transform Kalman Filter in silicon
  • Fuses IMU · pressure · optical flow in < 80 µs
  • State-vector width 64 (configurable)
  • Direct path to actuator DMA — zero-copy control loop

Sensor Fabric

  • Native OC-IMU-001 · OC-ACC-001 · OC-PRS-001 bus
  • Hardware timestamping — 10 ns resolution
  • Sensor DMA with zero-latency ring buffers
  • Auto-calibration co-processor (OC-TBIAS-001 aware)

Real-Time Subsystem

  • 4× lockstep RISC-V cores @ 200 MHz
  • Hardware task scheduler (no OS jitter)
  • Deterministic CAN-FD for joint bus control
  • ISO 26262 ASIL-B functional safety path

Memory Hierarchy

  • L1: 32 KB I$ + 32 KB D$ per RT core
  • L2: 512 KB shared (4-way set assoc.)
  • On-die SRAM: 4 MB (sensor scratchpad + model)
  • LPDDR4X interface up to 68 GB/s

Security & Safety

  • TrustZone-equivalent secure enclave
  • Hardware RNG + AES-256 + SHA-3 accelerators
  • ECC SRAM + lockstep error detection
  • Signed firmware boot chain
Specifications

DragonX RXP-1 — Key Parameters

Process NodeTSMC 28nm HPC+ LVT
ISARISC-V RV64GCV + Custom Extensions
Core Count4× real-time + 1× application
Peak TOPS (INT8)12 TOPS (on-chip NPU)
SRAM4 MB on-die (sensor scratchpad + model cache)
Sensor I/OSPI × 8 · I²C × 4 · MIPI CSI-2 × 2 · CAN-FD × 2
IMU interfaceNative OC-IMU-001 protocol (DragonX sensor bus)
Interrupt latency< 80 µs worst-case (hardware LETKF pipeline)
TDP< 4 W (nominal robot payload)
PackageFCBGA-400 · 17 × 17 mm
PMIC companionOC-PMIC-001-MR (co-packaged option)
StatusArchitecture complete · TSMC 28nm sign-off in progress
Performance

Versus today's discrete robot stack

Analytical model. Silicon characterisation in progress on TSMC 28nm.

Metric
RXP-1
Discrete SoTA
Advantage
Sensor fusion latency
< 80 µs
1–5 ms (SW)
12–60×
Robot BERT (int8)
420 fps
38 fps (Jetson Orin)
11×
Idle power (RT cores only)
380 mW
2.1 W
5.5×
CAN-FD joint bus latency
< 10 µs
50–200 µs
5–20×
Board real-estate (full stack)
1 SoC
MCU + GPU + FPGA
3× smaller
Competitive Landscape

Versus the chips robotics teams actually buy today.

We do not compete on raw TOPS — Jetson Orin AGX wins that fight 23 ×. We compete on the metrics that determine whether a robot's control loop closes in time: sensor-fusion latency, real-time determinism, sensor I/O depth, and watts. Honest table below.

MetricRXP-1NVIDIA Jetson Orin AGXQualcomm RB6Renesas RZ/V2HNXP i.MX 8M Plus
AI Peak (INT8 TOPS)12275 (sparse)15802.3
TDP (full system)< 4 W15–60 W5–15 W12 W5 W
Sensor-fusion latency< 80 µs (HW LETKF)2–5 ms (SW)3–8 ms (SW)1–3 ms (SW)5–10 ms (SW)
Real-time cores4× lockstep RV64— (A78AE only)— (Kryo only)2× Cortex-R521× Cortex-M7
Native CAN-FD2× on-dievia ext PHY
Functional safetyASIL-B pathASIL-D (Drive variant)ASIL-DASIL-B
AI / ML ecosystemONNX · TFLite · MLIRCUDA · TensorRT (vast)SNPE · TFLiteDRP-AI compilereIQ · TFLite
Process nodeTSMC 28 nmSamsung 8 nmTSMC 4 nmTSMC 16 nmTSMC 14 nm
Unit price (rough)~$45 target$1,999 module~$200 SoC~$80~$40
India local-content / TPCR✓ eligible
Status (mid-2026)Place+CTS done · routing WIPMass productionMass productionMass productionMass production

Competitor specs from public datasheets & product briefs (NVIDIA Jetson Orin AGX 64 GB module, Qualcomm RB6 dev platform, Renesas RZ/V2H, NXP i.MX 8M Plus). RXP-1 numbers are pre-silicon design targets — characterised values follow first tape-out.

Where we win

  • • Sensor-fusion latency — 25–60× lower than any SW solution
  • • Real-time determinism — only RZ/V2H comes close
  • • TDP at full system load — fits a 5000 mAh LiPo
  • • Native sensor bus — no glue ICs, no level shifters
  • • Made-in-India / TPCR-eligible

Where we lose (honest)

  • • Raw TOPS — Orin is 23× ahead on peak inference
  • • No CUDA / TensorRT — ONNX + MLIR only
  • • Older node (28 nm vs Orin 8 nm, RB6 4 nm)
  • • No on-die GPU for graphics / vision
  • • Zero shipped silicon yet — they are in mass production

The right framing

RXP-1 is not the chip you replace Jetson with — it's the chip you put next to Jetson to run the real-time control loop. Today's robots use Jetson (perception) + STM32 (control) + FPGA (sensor glue). RXP-1 collapses the last two into one die with better fusion than either could do alone — at 4 W, $45, and ASIL-B.

Applications

Built for the machines that will be everywhere.

Humanoid Robots

Full-body proprioception + vision inference on a single SoC. Replaces discrete MCU + GPU + FPGA stack in arms like Figure 02 and Apptronik Apollo.

Figure · Apptronik · 1X

Autonomous Drones

Sub-100 µs LETKF sensor fusion loop enables aggressive manoeuvering in GPS-denied environments. 4 W TDP fits a 5000 mAh LiPo mission profile.

ideaForge · Garuda · Asteria

Industrial Arms & AMRs

CAN-FD joint bus + deterministic RT core replace separate safety PLC + motion controller. ISO 26262 ASIL-B path for collaborative robot cells.

Warehousing · Manufacturing · Logistics

Space & Defence Payloads

TSMC 28nm die shrink + radiation-tolerant layout option (in design). TPCR-eligible as Indian-designed IP. Qualified to MIL-STD-883 screen on request.

ISRO · DRDO · BEL

Evaluating RXP-1 for your robot platform?

We work with robotics teams on architecture evaluation, design-on-contract sensor integration, and custom IP development. DragonX Systems — Stanford · IIT alumni founding team.

First conversation free — send us a 1-page spec and get a fixed-price design quote within 5 working days.