FMQL30TAI FMSH FPAI 8 TOPS AI SoC FCBGA676
- Product Details
- Download
- 125K logic cells
- 9.3 Mb BRAM
- FCBGA676 (27×27 mm) package
- Quad-core processor @ 1 GHz
The FMSH FMQL30TAI is an FPAI heterogeneous AI SoC delivering 8 TOPS of INT8 compute on 1 × ZG330 iCore @ 1 GHz, with 125K logic cells of FPGA fabric in an FCBGA676 package measuring 27×27 mm. A CPU, an NPU and reconfigurable logic share one die, programmed through the ICraft toolchain.
Key Specifications
Programmable Logic
| Logic Cells | 125K |
| LUT Count | 78,600 |
| BRAM Capacity | 9.3 Mb |
| PCIe | Gen2 x4 |
| High-Speed Transceivers | 4 × transceivers, up to 12.5 Gbps |
| XADC | 1 |
Processing System
| Processor | Quad-core processor @ 1 GHz |
| L1 Cache | 32KB I + 32KB D |
| L2 Cache | 256KB |
| On-Chip Memory | 256KB |
| DMA Channels | 8 channels |
| External DDR | DDR3 / DDR3L / LPDDR2 (ECC) |
| External Storage | 2×Quad-SPI, NAND, NOR |
| General Peripherals | CAN, SPI, UART, I2C, GPIO (up to 174 pins) |
| High-Speed Peripherals | 2×USB, 2×GigE |
| Security Features | Dedicated security engine: RSA / AES / SHA-256 / SM4 with secure boot |
AI Engine
| AI Architecture | 1 × ZG330 iCore @ 1 GHz |
| Compute | 8 TOPS @ INT8 |
| FP32 Compute | 4 TFLOPS @ FP16 |
| Numeric Precision | INT8 / FP16 / BF16 / TF32 |
| Model Support | CNN, Transformer and similar networks |
| Power | Dynamic power management |
Package and I/O
| Package | FCBGA676 |
| Package Size | 27 × 27 mm |
| Ball Pitch | 1.0 mm |
| HPIO / HDIO Channels | 150/100 |
Family Overview
FPAI is a heterogeneous reconfigurable chip that puts FPGA fabric, an AI accelerator and a CPU on a single die. The line spans 4 to 128 TOPS and is programmed through the ICraft toolchain, which covers model compilation down to the fabric.
- CPU + NPU + FPGA on one die
- 4 to 128 TOPS
- ICraft toolchain
- Distributed LLM inference
Applications
Edge inference, video analytics, industrial inspection and distributed LLM serving where a CPU, an NPU and reconfigurable logic must share one die.
Design Notes
- TOPS figures are peak INT8 unless stated otherwise. Real throughput depends on the model's operator mix — convolution-heavy networks reach the peak far more readily than attention-heavy ones.
- Model compilation goes through ICraft. Operator coverage is the practical limit, so check that your network's layers are supported before sizing the part.
- Host and device memory are separate. For large models, plan the weight partitioning and the PCIe or Ethernet link budget between cards.
- Power scales with the number of active cores. The headline figure assumes a workload profile; measure against your own model rather than the datasheet ceiling.
Common Questions
How much logic does the FMQL30TAI provide?
125K programmable logic cells. The fabric also carries the BRAM and DSP resources listed above.
What package is the FMQL30TAI in?
FCBGA676, 27×27 mm. Check the pin-out for the transceiver and HPIO assignment before layout.
What is the AI throughput of the FMQL30TAI?
8 TOPS @ INT8. That is the peak figure; sustained throughput depends on the operator mix of the model being compiled.
Which EDA toolchain supports the FMQL30TAI?
Procise for the FM7 family, Vulture for FM9, and ICraft for FPAI. The PSoC line is supported within the same flow as the FPGA family it derives from.
Shenzhen Gaorunxin Technology Co., Ltd
Online message
Shenzhen Gaorunxin Technology Co., Ltd
