Glenfly Samples Its First Arm Chip, With an In-House GPU

On the floor of the 26th China International Industry Fair in Shanghai, Glenfly Tech showed the chip itself, mounted on its boards. The GF800 is the first system-on-chip, or SoC, that the company has built around its own graphics processor, and it has finished debugging, the stage at which faults in the design are found and fixed before a product is sent out. Samples are now going to end customers.

For Glenfly, the change is as much about business as engineering. Its executives described a move from supplying graphics cards to offering a complete system-level platform. An SoC combines the main processor, graphics and other blocks on one piece of silicon, so a device maker can build a product around one chip instead of assembling several. The GF800 pairs eight Arm Cortex-A78 processor cores with Glenfly's GPU, a dedicated NPU for AI calculations, and a unified memory architecture shared by the CPU, GPU, NPU and video units. Glenfly says its GPU supports DirectX 12, OpenGL 4.6, OpenGL ES 3.2, OpenCL 3.0 and Vulkan 1.3, and that it has passed the conformance tests for each standard, meaning the software standards are implemented as specified. The chip drives up to four independent displays, outputs up to 8K, and its new image signal processor accepts inputs of up to 48 megapixels. It also carries PCIe 4.0, USB 3.2, SATA and Gigabit Ethernet connections.

The practical question is who benefits and when. Glenfly names AI terminals, edge servers, industrial computing equipment and robots as its targets, and says the aim is to bring computing closer to the data and the machines that use it, within real limits of power and cost, rather than simply moving cloud computing onto devices. The chip is optimized for China's domestic operating systems, UnionTech's UOS and Kylin, and also supports Linux and Android and is compatible with Windows. For device makers, that means a chip they could build on without leaving those systems. But the timing is still open: the chip is going through adaptation and solution validation for graphics, display, system compatibility and industry scenarios, and no date for volume production has been given.

Several numbers that matter for an AI chip are missing. Glenfly has not disclosed the NPU's performance rating, which would show how much AI work the block can do per second, nor the manufacturing process used to make the chip, which determines how small and efficient its transistors can be. Those figures, along with customer results from the sampling phase, will be what separates a working platform from a product that can compete on performance.