Skymizer HyperThought & EdgeThought LPU IP for LLM Inference

Skymizer Taiwan Inc. is a software and hardware co-design IP company founded in 2013 and headquartered in Neihu, Taipei, with a decade of AI compiler and runtime work. It licenses Language Processing Unit (LPU) IP and integrated software so chipmakers, device makers and system integrators can run large language models on-device or on-premise without GPU clusters.

HyperThought is an octa-core, multi-chip scalable LPU IP built on the compiler-driven LISA v3 architecture, targeting multimodal and agentic AI at the edge and on-premise. It uses standard LPDDR4/5 memory, weight and KV-cache compression and a decode-first design to support models from 4B to 600B parameters on a mature 28 nm process. EdgeThought is a single-core LPU IP for on-device LLM inference in resource-constrained products. The HTX301 reference chip demonstrates the platform on a single PCIe card. Regional partners operate in South Korea, China and Japan.

• HyperThought LPU IP: octa-core, multi-chip scalable IP on the LISA v3 architecture.
• EdgeThought LPU IP: single-core, compiler-centric IP for on-device LLM inference.
• Model range: supports models from 4B up to 600B parameters with prefill and decode disaggregation.
• Compression: weight and KV-cache compression to reduce DRAM bandwidth needs.
• Mature-node design: built on 28 nm with standard LPDDR4/5 memory.
• HTX301 reference chip: six-chip PCIe card with 384 GB memory for on-premise inference.

A Vietnamese chip design company can license EdgeThought to build an on-device AI chip for smart home or industrial terminals. A system integrator can deploy on-premise LLM servers for banks or government agencies that cannot send data to the cloud. An automotive electronics supplier can integrate LPU IP for in-vehicle voice assistants.

LLM inference runs on the customer's own hardware with predictable cost and no dependence on GPU clusters. The HTX301 reference design runs 700B-parameter models on one PCIe card at about 240 W. Weight compression outperforms llama.cpp by 9% to 17.8%, with KV-cache compression perplexity loss of 0.06% to 3.52%. A mature 28 nm process and standard LPDDR memory keep silicon cost and supply risk low.

Enterprises and device makers want to run generative AI locally for data privacy, latency and cost reasons, creating demand for dedicated LLM inference silicon outside GPU supply chains. Mature-node accelerator IP lowers the entry barrier for regional chipmakers. Vietnam's push for sovereign AI and domestic chip design makes licensable LLM accelerator IP relevant for local projects.

Founded in 2013 in Taiwan; headquartered at Ruiguang Road, Neihu District, Taipei. HyperThought won Best IP/Processor of the Year and Most Promising Product at EE Awards Asia 2025. JFE Shoji Electronics became its official distributor in Japan in September 2025; other regional partners include CIRCLE (South Korea), ADK and LoMicro (China).