#Taiwan #Taipei #Skymizer #HyperThought #AI Chip

Skymizer Launches HyperThought: Build Your Own AI Chip with Skymizer's LPU IP

Skymizer has officially announced the launch of its most advanced AI accelerator, HyperThought, following the success of its previous product, EdgeThought, which fundamentally shifted on-device LLM (Large Language Model) inferencing in 2024. HyperThought is engineered to address the increasing need for real-time, multimodal, and agent-based AI, marking a substantial advancement in performance, scalability, and flexibility for deployment.

The Essence of HyperThought

HyperThought offers SoC teams (System-on-Chip) the opportunity to craft their own AI silicon tailored explicitly to different applications. With HyperThought, developers can merge their unique domain-specific intelligence with a host CPU and edge subsystems, creating a fully integrable subsystem rather than just an accelerator.

In its workings, HyperThought is powered by Skymizer's third-generation Language Instruction Set Architecture (LISA) v3, which facilitates multimodal processing, concurrent execution of multiple models, and native functionalities for agentic AI workflows. Moreover, the platform is supported by a completely revamped runtime and compiler stack, leading to unprecedented efficiency in compute, memory, and bandwidth, which is crucial for a diverse range of edge AI implementations ranging from smart devices to autonomous systems.

Key Features of HyperThought

1. Built on LISA v3: Skymizer’s innovative third-generation Language ISA introduces advanced capabilities such as dynamic graph control, parallel agent execution, and enhanced memory operations that are fully supported by the hardware.

2. Multimodal Capability: The chip allows seamless processing across different model types, encompassing text, vision, and instruction-based paradigms, sharing computing resources effectively.

3. Concurrent Multi-Model Support: HyperThought can efficiently execute multiple quantized LLMs simultaneously, ensuring low-latency responses across various applications.

4. Agentic AI Infrastructure: The system is designed to enable goal-driven agents that can engage in interactive and contextual reasoning across multiple tasks.

5. Optimized Software Stack: HyperThought comes equipped with a co-designed compiler-runtime platform that optimizes compute throughput while respecting the limitations of memory and bandwidth.

6. Scalable Integration: The IP block is compact and modular, facilitating quick integration into various SoCs, including mobile, embedded, and automotive systems.

A Transformative Leap Forward

William Wei, CMO of Skymizer, commented, “EdgeThought brought LLM inference to the edge. HyperThought makes the edge intelligent. With LISA v3, we’re not just accelerating tokens—we’re empowering intelligent agents capable of reasoning and interacting in real-time. This marks a transformational phase for edge AI technology.”

Join Skymizer at COMPUTEX 2025

Skymizer will showcase the capabilities of HyperThought at COMPUTEX Taipei from May 20 to May 23, 2025. Attendees can look forward to:

• - A live demonstration of FPGA-verified IP that powers multimodal inference and agent-based workflows.
• - Technical sessions that explore LISA v3 and the best practices for SoC integration.
• - Preview sessions highlighting the advanced tools and developer stack available for HyperThought users.

Visitors interested in attending must register for a COMPUTEX Member Account ahead of time through this link.

About Skymizer

Skymizer stands as a leading semiconductor IP provider committed to facilitating intelligent computing across a variety of applications. With rich expertise in compiler technology and silicon architecture, Skymizer delivers efficient and scalable AI acceleration solutions for the next generation of intelligent devices. To learn more, visit www.skymizer.com.