SHANGHAI -- VeriSilicon (688521.SH) announced the latest advancements in its high-performance and scalable GPGPU-AI computing IPs, which are now empowering next-generation automotive electronics and edge server applications. Combining programmable parallel computing with a dedicated Artificial Intelligence (AI) accelerator, these IPs offer exceptional computing density for demanding AI workloads such as Large Language Model (LLM) inference, multimodal perception, and real-time decision-making in thermally and power-constrained environments.

VeriSilicon’s GPGPU-AI computing IPs are based on a high-performance General Purpose Graphics Processing Unit (GPGPU) architecture with an integrated dedicated AI accelerator, delivering outstanding computing capabilities to AI applications. The programmable AI accelerator and sparsity-aware computing engine accelerate transformer-based and matrix-intensive models through advanced scheduling techniques. These IPs also support a broad range of data formats for mixed-precision computing, including INT4/8, FP4/8, BF16, FP16/32/64, and TF32, and are designed with high-bandwidth interfaces of 3D-stacked memory, LPDDR5X, HBM, as well as PCIe Gen5/Gen6 and CXL. They are also capable of multi-chip and multi-card scale-out expansion, offering system-level scalability for large-scale AI application deployments.

VeriSilicon’s GPGPU-AI computing IPs provide native support for popular AI frameworks for both training and inference, such as PyTorch, TensorFlow, ONNX, and TVM. These IPs also support General Purpose Computing Language (GPCL) which is compatible with mainstream GPGPU programming languages, and widely used compilers. These capabilities are well aligned with the computing and scalability requirements of today’s leading LLMs, including models such as DeepSeek.

“The demand for AI computing on edge servers, both for inference and incremental training, is growing exponentially. This surge requires not only high efficiency but also strong programmability. VeriSilicon’s GPGPU-AI computing processors are architected to tightly integrate GPGPU computing with AI accelerator at fine-grained levels. The advantages of this architecture have already been validated in multiple high-performance AI computing systems,” said Weijin Dai, Chief Strategy Officer, Executive Vice President, and General Manager of the IP Division at VeriSilicon. “The recent breakthroughs from DeepSeek further amplify the need for maximized AI computing efficiency to address increasingly demanding workloads. Our latest GPGPU-AI computing IPs have been enhanced to efficiently support Mixture-of-Experts (MoE) models and optimize inter-core communication. Through close collaboration with multiple leading AI computing customers, we have extended our architecture to fully leverage the abundant bandwidth offered by 3D-stacked memory technologies. VeriSilicon continues to work hand-in-hand with ecosystem partners to drive real-world mass adoption of these advanced capabilities.”