FFI8805 built-in CIM Array combined with FPGA Gateway and power electronics core for millisecond-level intelligent anomaly detection and state estimation
This integration brings CIM AI computing to the power grid edge, achieving millisecond-level intelligent anomaly detection and state estimation without compromising hard real-time protection safety.
Core design principles: Clear layering, technology complementarity, safety first — Click each layer to expand details
Station SCADA/EMS + CPU asset health & trend analysis
Model publishing, monitoring telemetry & MMS reporting, non-hard real-time
CIM accelerator for fault identification / state estimation (Batch=1, low latency)
Outputs advice/score/estimation without direct trip action. Weights resident on chip (Model-in-Chip)
Power electronics core PWM/closed-loop control + FPGA/IED fast trip/safety gate
μs-level closed-loop control, OC/OV/short-circuit hardware protection, gate drive & fast shutdown
The 'Fast Real-Time Layer' in power system control protection, executing efficient inference within 0.5–10 ms cycles
| Precision | BF16/FP16 |
| 추론 지연 시간 | WCET ≤ 100 μs |
| End-to-End Latency | ≤ 300 μs |
| Memory Architecture | Model-in-Chip (Cache-less) |
| Interconnect | CXL/PCIe Gen5 |
x[FEATURE_DIM] → CIM Core → y_score / ood_score / confidenceEdge inference module integrating CIM Array, bridging power electronics high-speed I/O with upper control systems
| FFI8805 Mini | 288KB CIM Array (1152×256 bits) |
| FFI8805 Pro | 576KB CIM Array (1152×256 bits) |
| Protocol | IEC 61850 SV/GOOSE |
| Sync | PTP (IEEE 1588) |
Handles μs-level voltage/current sampling and PWM generation, triggers hardware protection on severe faults
End-to-end latency exceeding control period causes closed-loop failure
Cache-less/TCM architecture eliminates cache miss; LOW_JITTER_MODE fixed scheduling
SV packet jitter/loss, timestamp deviation affects phasor measurement
IEEE 1588 PTP hardware timestamping; PRP/HSR network redundancy & Quality Flags
Overheating in sealed enclosures causes throttling or thermal shutdown
10–30W TDP optimized design; S1–S4 degradation state machine active throttling
High data transfer latency or data races between FPGA and CIM
CXL shared memory (Zero-Copy); Ring Buffer with explicit Memory Barrier
AI model misjudgment or hardware failure causes protection malfunction
FPGA safety gate final arbitration; WDT/ECC/model signature verification
fault_type, confidence, ood_scoreestimated_state[N], topology_errorthd_score, transient_event_iddevice_health, predictive_alertBatch-1 architecture with Model-in-Chip design meets 1–10 ms fast real-time requirements, eliminating cache jitter
Degradation state machine, WDT/ECC/CRC protection with quality flags & OOD detection to prevent malfunction
CXL/PCIe standard interface + Ring Buffer protocol, MMIO/IRQ compatible with existing station/IED systems
10–30W CIM power design, suitable for passive cooling in substations or roadside cabinets
Here are the most frequently asked technical questions about FFI8805 in power system applications
Driving heterogeneous packaging of FPGA and CIM to further reduce interconnect latency and power
Complete MLOps pipeline with model signature verification, canary deployment & auto-rollback
Extending to PMU/WAMS edge event summary for asset health assessment & wide-area cross-station defense
Contact our technical team to learn how FFI8805 can provide millisecond-level intelligent inference for your substations, distribution networks, or power electronics