Innosilicon Fenghua 2
Product Overview
Fenghua 2 (Fenghua No.2) is Innosilicon's second-generation self-developed GPU, released in Wuhan on 2022-08-03. Unlike the 2021 Fenghua 1, which targeted high-performance data centers, Fenghua 2's core goals shifted to versatility and ultra-low power: the chip adopts a full set of Innosilicon's own IPs, with FP32 floating-point compute of 1.5 TFLOPS, texture/pixel fill rate of 48 GTex/s (48 GPixel/s), and AI compute of 12.5 TOPS (INT8), while typical operating power is only 4~15W, as low as about 4W at minimum. It can be designed as a passively cooled (fanless) structure and can also be used in laptops and MXM all-in-ones.
Fenghua 2 is a 4K-class high-performance rendering GPU that integrates ultra-low power, HD triple-display output, and intelligent AI edge computing, with built-in AI and floating-point compute cores supporting scientific and edge computing. The chip supports Innosilicon's own LPDDR5X memory technology (also compatible with LPDDR5/4X/4), can drive 4K@60Hz triple independent displays, and integrates physical unclonable function (PUF) security technology and hardware virtualization.
On the domestic ecosystem front, Fenghua 2 has completed mutual certification with domestic operating systems such as UOS, Kylin (Galaxy Kylin, NeoKylin), and supports domestic CPU platforms including Phytium, Loongson, Sunway, Hygon, Zhaoxin, and Kunpeng, while remaining compatible with mainstream Intel / AMD x86 CPUs and systems such as Windows and Ubuntu. It scored over 6500 in the GLmark2 benchmark on Kylin OS and passed the heavy Unigine Heaven benchmark. Launch demonstrations covered scenarios such as Kingsoft WPS, ZWCAD, GIS, glmark, a racing game, and Windows 10 office and entertainment use.
Commercial traction: on launch day, Innosilicon signed procurement agreements worth a total of RMB 500 million with companies including China Energy Engineering Investment and UnionTech (UOS), covering "Eastern Data, Western Computing" project orders and Xinchuang industry demand; in September 2023 it formed a strategic partnership with SAHC to advance Fenghua 2 in industrial applications. Fenghua 2 won the "2022 Hardcore China Chip — Best Processor Chip" award.
⚠️ Positioning Clarification Some sources classify Fenghua 2 as a "server/data center-class GPU". According to Innosilicon's official announcements and multiple media reports, Fenghua 2's primary positioning is an ultra-low-power desktop / laptop / industrial PC / embedded GPU (typical power 4~15W, PCIe 3.0 x8, LPDDR5X memory up to 8GB). The official materials also mention it can "provide 4K-class rendering support to aid data center computing", but the server/data center-class high-performance positioning belongs to Fenghua 1 (5 TFLOPS FP32, GDDR6X, up to 16GB, 304 GB/s), not Fenghua 2. This page follows the official low-power desktop positioning.
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | Innosilicon proprietary GPU architecture (full set of self-developed IPs, self-developed controllable core evolution) |
| Process Node | Not disclosed |
| FP16 / BF16 Compute | Not disclosed |
| INT8 Compute | 12.5 TOPS |
| FP32 Compute | 1.5 TFLOPS |
| Texture/Pixel Fill Rate | 48 GTex/s / 48 GPixel/s |
| Memory Capacity | 2GB / 4GB / 8GB |
| Memory Type | LPDDR5X / LPDDR5 / LPDDR4X / LPDDR4 (Innosilicon's own LPDDR5X technology), GDDR6 option also available |
| Memory Bandwidth | Not disclosed |
| TDP | Typical power 4~15 W (measured; supports high-performance and low-power modes; passive cooling possible) |
| Interconnect | Not disclosed (no multi-card interconnect solution) |
| Interface | PCIe 3.0 x8 |
| Display Outputs | HDMI 2.0 ×2, DP 1.4 ×1, VGA ×1, LVDS ×1; supports 4K@60Hz triple independent displays |
| Graphics APIs | OpenGL 4.2, OpenGL ES 3.2, Vulkan 1.2, DirectX 11 |
| Compute APIs | OpenCL 3.0 |
| AI Frameworks | Caffe, TensorFlow, PyTorch |
| Video Capability | 4K hardware video decoding |
| Virtualization | Hardware virtualization, multiple operating systems running simultaneously |
| Security | Built-in physical unclonable function (PUF) security technology |
| Benchmark Results | GLmark2 > 6500 on Kylin OS; passed the heavy Unigine Heaven test |
| Launch | 2022-08-03 (Wuhan launch event) |
| Mass Production/Availability | In mass production and commercial use; RMB 500 million in procurement agreements signed on launch day (China Energy Engineering Investment, UOS, etc.) |
Key Features
- Industry-leading energy efficiency: runs office, gaming, CAD, and glmark smoothly at operating power as low as about 4W, with energy efficiency far better than products of similar compute; fanless designs, laptops, and MXM all-in-ones supported under appropriate conditions
- 4K triple independent display output: HDMI 2.0×2 + DP 1.4 + VGA + LVDS, meeting industrial and multi-screen office needs
- Self-developed LPDDR5X memory technology: low-power high-bandwidth memory IP, extending Innosilicon's core advantage in memory interface IP
- Fully domestic board BOM: memory, power management, and other components on the board are all domestically sourced; cross-process, cross-package design ensures supply chain security
- Among the broadest Xinchuang ecosystem support: mutually certified with UOS and Kylin; supports six domestic CPU platforms — Phytium, Loongson, Sunway, Hygon, Zhaoxin, and Kunpeng
- Built-in AI + floating-point compute cores: 12.5 TOPS (INT8) AI compute, supporting scientific computing and edge AI inference
- Hardware virtualization: multiple operating systems running simultaneously
- PUF hardware security: physical unclonable function technology, anti-cloning and anti-tampering
- Well-validated commercially: RMB 500 million in orders on launch day, covering "Eastern Data, Western Computing" and Xinchuang projects
- Limitations: 1.5 TFLOPS FP32 is entry-level; PCIe 3.0 x8 interface bandwidth is limited; max 8GB memory; memory bandwidth not disclosed; no large-model training/inference capability
Vendor Information
| Parameter | Value |
|---|---|
| Company | Innosilicon |
| Headquarters | Wuhan, Hubei Province (East Lake High-Tech Development Zone); R&D centers in Zhuhai, Suzhou, Xi'an, Beijing, Shanghai, Shenzhen, Dalian, and Chengdu |
| Founded | 2006 (founder Ao Hai) |
| R&D Scale | R&D team of over one thousand; thousands of core IPs; 200+ global core patents |
| Mass Production Accumulation | Over 300 advanced-process tape-outs; supported mass production of over 10 billion high-end SoC chips |
| Process Coverage | 55nm — 3nm, covering TSMC / Samsung / SMIC / GlobalFoundries / UMC / Intel / Huali |
| Fenghua 2 Development Lead | Led by Innosilicon VP Mao Mingming; algorithm team led by chief algorithm scientist Dr. Yang Xile (former Imagination architect) |
| Business Status | Profitable for consecutive years, not reliant on external financing; preparing for a STAR Market IPO |
| Fenghua Series Evolution | Fenghua 1 (2021-11, data center class) → Fenghua 2 (2022-08, low-power desktop class) → Fenghua 3 (2025-09, full-function GPU integrating a RISC-V CPU with a CUDA-compatible GPU) |
| Official Website | https://www.innosilicon.com.cn |
Use Cases
- ✅ Xinchuang office desktop graphics (supports WPS, ZWCAD, and other domestic office and engineering software)
- ✅ Industrial equipment (financial terminals, rail transit terminals, and other industrial environments)
- ✅ Embedded display and compute (smart terminals, in-vehicle systems)
- ✅ Fanless / passively cooled low-power systems, laptops, MXM all-in-ones
- ✅ Edge AI inference (12.5 TOPS INT8)
- ✅ 4K triple-display multi-monitor office and surveillance walls
- ✅ Desktop virtualization (hardware virtualization with multiple parallel systems)
- ❌ Large-model AI training (1.5 TFLOPS FP32, no HBM, no interconnect)
- ❌ Large-model inference (max 8GB memory)
- ❌ High-density data center cloud rendering (choose Fenghua 1 instead)
- ❌ High-end AAA gaming and professional real-time rendering
- ❌ Direct migration from the CUDA ecosystem
Related Cards
- Innosilicon Fenghua 1 — Previous generation in the same series, with a data center/server-class high-performance positioning
- Jingjia Micro JM9 Series — The closest-positioned domestic competitor (Xinchuang desktop, <30W, 8GB, 1.5 TFLOPS FP32)
- Jingjia Micro JM9271 — Domestic high-end GPU comparison
- Moore Threads MTT S5000 — Domestic full-function GPU, high-compute route comparison
- MetaX C600 — Domestic high-end GPU comparison
References
- Fenghua 2 (Baidu Baike, with launch date, compute, power, and order information)
- Innosilicon releases "Fenghua 2" GPU, capable of fanless design with triple 4K display support (EXPreview, 2022-08-03, with full interface and API specs)
- Bringing Fenghua GPUs into millions of households, Innosilicon empowers smart living (EEWorld, official positioning of product performance and technical innovation)