AI compilers service provider, Skymizer has announced a strategic pivot by launching the ET2 LPU chip hardware IP and the Skymizer LLM System platform, entering the chip IP solutions market.
Skymizer says they are targeting consumer electronics and automotive sectors in this initial stage with product prices ranging from US$100 to US$500. The company is also positioning itself within the AI PC market, though with a unique approach distinct from existing NPUs on the market.
The focus is on using LLM technology as an interface for human-machine interaction.
Vision for the technology
Skymizer's founder and CEO, Luba Tang, explained that LLM represents a new interaction interface for human-to-machine and machine-to-machine communication. Based on the company's observations of AI development trends, Tang predicts that every computing device may eventually need two LLM chips: one for human interaction and another for machine interaction.
Edge AI trends
Recognizing the trend of generative AI moving towards edge computing, Skymizer is focusing on developing small models for specific applications. This requires new technical approaches to address the challenges of deploying generative AI extensively at the edge.
Tang highlighted that the absence of corresponding IC products in the market is a key reason why generative AI applications are not yet widespread. He expects that by the first quarter of 2025, many generative AI chip products will be available.
Overcoming development challenges
The inconsistency between chip development timelines and the rapid pace of AI technology evolution remains an issue. This is an area where Skymizer excels, leveraging its LLM System platform to bridge the gap. In addition, Skymizer is addressing the high costs and power consumption challenges associated with LLM deployment by developing low-cost, low-power hardware IP. This comprehensive hardware and software solution enables IC design companies to quickly introduce LLM chip products.
Innovative product design
The ET2 LPU, developed with an LLM-centric approach, is designed around memory, differing significantly from the traditional Convolutional Neural Network (CNN) approach, which focuses on computational power. Tang noted that LLMs do not require high computational power and the critical bottleneck is memory bandwidth, which necessitates new DRAM design architectures. Many DRAM companies with specialized architectures are based in Taiwan, facilitating collaboration.
Competitive advantage
Compared to other major IC design firms, the ET2 LPU can achieve superior performance with just 0.4 TOPs of computational power, processing 32 tokens per second, and can be manufactured using TSMC's 28nm process.
Ultimately, the computational power requirements depend on customer applications. For example, cheaper children's toys requiring basic generative AI dialogues can use lower-spec LPUs, which are more cost-effective and power-efficient than NPUs.