Technology
Every platform below is built on our own heterogeneous processors — general-purpose ARM computing fused with the patented ManyCore™ deep-learning architecture, and shipped with a complete software toolchain.
Designed for “general-purpose computing + AI inference” in cloud and edge scenarios. Heterogeneous computing and a compute-storage coupled architecture push throughput per watt and per dollar. Supports all mainstream Linux distributions and virtualization solutions.
| CPU | 40-core ARMv8.2 64-bit @ 2.5–3.0 GHz |
| AI Engine | Built-in inference NPU, patented ManyCore™ architecture |
| Memory | 4× DDR4-3200 / DDR5-4800 channels, 1DPC / 2DPC |
| I/O | 64-lane PCIe 4.0 bus |
| Video Decode | 64× 1080p30 H.264 / H.265 hardware decoder |
| Video Encode | 10× 1080p30 H.264 / H.265 hardware encoder |
| Scaling | Up to 4× NUMA interconnection |
Standard single-socket boards that drop into any rack or tower chassis, plus fully customized dual-socket designs adapted to the client's enclosure.
| Form Factor | Standard Micro-ATX / ATX, or customized to client chassis |
| Socket | Single TF7000 (standard) · Dual TF7000 (custom) |
| Memory | 4×/8× DDR4 RDIMM (single) · 16× DDR4 RDIMM (dual) |
| Expansion | Multiple PCIe 4.0 slots for GPU and peripherals; 1× M.2 PCIe 4.0 x4 NVMe |
| Network | Integrated dual GbE data ports |
| Management | Server-grade BMC, dedicated GbE port, IPMI protocol |
| Power | ATX socket + PMBus telemetry (single) · CRPS redundant PSU (dual) |
Purpose-built for “general-purpose computing + visual analytics” at the cloud edge. Multiple computing clusters run diverse AI algorithms in parallel.
| CPU | 12 / 20-core ARMv8 64-bit @ 1.8–2.4 GHz |
| AI Engine | Built-in inference NPU, patented ManyCore™ architecture |
| Memory | 5× LPDDR4X-4266 channels |
| I/O | 8-lane PCIe 3.0 bus |
| Video Decode | 64× 1080p25 H.264 / H.265 hardware decoder |
| Interfaces | RGMII, USB 2.0, UART, SPI, I2C, GPIO |
A compact computing cell for scalable cloud inference. Integrates the TF16110 heterogeneous processor and supports mainstream deep-learning frameworks and virtualization — elastic AI server clusters, built module by module.
| Processor | TF16110 · Quad-core ARMv8 64-bit @ 2.4 GHz |
| AI Compute | 4× ManyCore™ clusters, 12.8 TOPS @ INT8 |
| Memory | 18 GB LPDDR4X-4266 + 64 / 128 GB eMMC flash |
| Video Decode | 64× 1080p25 / 16× 4K25 hardware decoding |
| Connectivity | High-speed board connector, PCIe 3.0 + GbE |
| Power / Temp | 25 W peak · 0–55 °C operating |
The TF16110 on a standard half-height, half-length PCIe card — compatible with any scale-out rack server and exposing application-layer service APIs.
| AI Compute | 4× ManyCore™ heterogeneous computing clusters |
| Memory | 18 GB LPDDR4X-4266 + 64 GB eMMC flash |
| Video Decode | 64× 1080p25 / 16× 4K25 hardware decoding |
| Form Factor | PCIe 3.0 x8, half-height half-length, single slot |
| Host OS | Windows and Linux |
| Power / Temp | 25 W peak · 0–55 °C operating |
A dedicated 1U video analytics server built on NNM100 modules — ultimate analysis throughput and power efficiency, deployable in data centers or edge nodes alike.
| Chassis | Standard 19-inch 1U rack server |
| Compute | Up to 6× built-in NNM100 modules |
| Network | 2× GbE data + 1× GbE management |
| Capacity | Up to 144 HD video streams, real-time facial recognition |
| Scaling | Seamless upgrade and expansion via virtualization |
| Power | Dual PSU hot backup · 200 W peak |
Built on the versatile TF16005 processor for mixed transaction-processing and deep-learning workloads. Rich peripheral interfaces, strong environmental adaptability, industry-customized models on request.
| Processor | TF16005 · Quad-core ARMv8 64-bit @ 1.8 GHz |
| AI Compute | 2× ManyCore™ heterogeneous computing clusters |
| Memory | 10 GB LPDDR4X-3200 + 64 / 128 GB flash |
| Video Decode | 16× 1080p25 / 4× 4K25 hardware decoding |
| Interfaces | 1× GbE, USB 2.0 / 3.0, RS232 / RS485, GPIO, HDMI out |
| Power / Temp | 15–25 W peak · −20–85 °C operating |