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

TAIPEI, TAIWAN – May 9, 2025 – Following the success of EdgeThought, the breakthrough LPU that revolutionized on-device LLM inferencing in 2024, Skymizer today announced the launch of its powerful next-generation AI accelerator IP: HyperThought. Engineered to meet the growing demands of real-time, multimodal, and agent-based AI, HyperThought represents a major leap forward in performance, scalability, and deployment flexibility.

Build Your Own AI Chip with Skymizer’s HyperThought LPU IP – Skymizer’s LPU IP is not just an accelerator; it’s a fully integrable subsystem that empowers SoC teams to design and deploy custom AI silicon tailored for their applications. HyperThought gives developers the control to combine domain-specific intelligence with their own host CPU and edge subsystems.

HyperThought is powered by Skymizer’s Language Instruction Set Architecture (LISA) v3, enabling multimodal processing, concurrent multi-model execution, and native support for agentic AI workflows. Paired with a fully updated runtime and compiler stack, the HyperThought platform delivers unmatched efficiency in compute, memory, and bandwidth—making it ideal for a wide range of edge AI deployments from smart devices to autonomous systems.

“EdgeThought brought LLM inference to the edge. HyperThought makes the edge intelligent,” said William Wei, CMO of Skymizer. “With LISA v3, we’re not just accelerating tokens—we’re enabling intelligent agents that can reason, interact multimodally, and adapt in real time. This is a transformative step forward for edge AI.”

Key Highlights of HyperThought

• Built on LISA v3: Skymizer’s third-generation Language ISA introduces dynamic graph control, parallel agent execution, and advanced memory ops—fully hardware-supported.

• Multimodal Capability: HyperThought supports seamless execution of text, vision, and instruction-based models across shared compute resources.
• Concurrent Multi-Model Support: Efficient execution of multiple quantized LLMs in parallel, enabling low-latency response across applications.
• Agentic AI Infrastructure: Designed to support persistent, goal-driven agents capable of interactive and contextual reasoning across tasks.
• Optimized Software Stack: Co-designed compiler-runtime platform balances compute throughput with memory and bandwidth constraints.
• Scalable Integration: IP block is modular and compact for rapid integration into mobile, embedded, and automotive-grade SoCs.

“HyperThought is the most advanced LPU we’ve engineered,” said Luba Tang, CTO of Skymizer. “LISA v3 enables us to support not just language models, but real-time intelligent systems that need to react, infer, and adapt. It’s a foundational step toward truly autonomous, on-device AI.”

Join Skymizer at COMPUTEX 2025

Skymizer will showcase HyperThought at COMPUTEX Taipei, from May 20–23, 2025. Attendees will experience:

• FPGA verified IP powering multimodal inference and agent-based workflows
• Technical sessions on LISA v3 and SoC integration best practices
• Previews of HyperThought’s software tools and developer stack

Visit Skymizer at the following location

Taipei Nangang Exhibition Center Hall 1 – System Integration Solution Area
Our Booths:
1F Power Chip Booth: I-1111 (PSMC Booth)
4F RISC-V Booth: L-0425 (RVTA Booth)

Please note: Visitors must register for a COMPUTEX Member Account in advance:

https://www.computexonline.com.tw/?userlang=en

About Skymizer

Skymizer is a semiconductor IP provider enabling intelligent computing everywhere. With deep expertise in compiler technology and silicon architecture, Skymizer delivers efficient, scalable AI acceleration solutions for the next generation of smart devices.

www.skymizer.com