Following the release of DeepSeek v3.1， an announcement from the company has significantly boosted attention for domestic AI computing chips.
According to DeepSeek， the upcoming UE8M0 FP8 is a next-generation domestic chip design. This format utilizes an 8-bit exponent and 0-bit mantissa in its FP8 sub-format， specifically optimized for core AI computations like matrix multiplication. This level of optimization is crucial for enhancing AI performance.
Currently， NVIDIA’s AI chips widely support FP8 and even FP4 precision formats. In contrast， many domestic AI chips are still operating with FP16. FP8， while offering lower precision， delivers significantly higher performance. It’s estimated to at least double the performance within the same chip area， drastically reduce power consumption to about 1/4 of FP16， and also lowers bandwidth requirements， presenting substantial advantages.
In a recent announcement， Dongxin Co.， Ltd. stated that their subsidiary， Lichee Technology， has developed the 7G100 series GPU chips， which can support 8-bit integer computations and other computational tasks. This development by Lichee aligns with the industry’s move towards more efficient computational formats.
Dongxin Co.， Ltd. further explained that Lichee Technology is dedicated to researching and developing multi-level (scalable) graphics rendering GPU chips. These products are designed for mainstream graphics rendering and AI acceleration across end-user devices， cloud platforms， and edge computing environments. The versatility of Lichee’s GPUs is a key factor in their potential widespread adoption.
Their 7G100 series GPU chips are capable of supporting various computational precisions， including single-precision floating-point (FP32)， half-precision floating-point (FP16)， and 8-bit integer (INT8) operations. The company emphasizes that each precision level has distinct applicable scenarios based on performance， resource consumption， and efficiency. This nuanced approach allows developers to select the most suitable precision for their specific AI workloads， optimizing both performance and resource utilization.
The 7G100 series GPU is a genuinely self-developed domestic GPU， manufactured using a 6nm process. Its entire architecture， from compute cores to instruction sets， has been designed entirely in-house. Performance tests have indicated that this GPU surpasses the RTX 4060， and it is expected to be available for sale in the fourth quarter of this year. This achievement represents a significant milestone for China’s domestic AI hardware industry， showcasing its growing capabilities in advanced chip design.
