Taiwan, China-based Skymizer Launches HTX301 Inference Chip to Enable Efficient Local Large Model Operation

Skymizer, an AI IP company based in Taiwan, China, recently released the HTX301 accelerator chip for the inference decoding phase. This chip can work with GPUs responsible for the pre-filling stage to accelerate inference result output, improving the utilization rate and overall energy efficiency of computing systems.

HTX301 is built on the HyperThought hardware-software co-design platform using the LISA instruction set architecture, focusing on bandwidth-intensive loads. It supports SoC and PCIe AIC deployment forms; the PCIe version can integrate 6 HTX301 chips and 384GB memory on a single card, meeting the demand for local operation of 700B high-parameter large models while consuming only 240W of power.