ZeroStream uses hardware-accelerated compression to move more data across existing memory and streaming interfaces, improving performance for bandwidth-bound applications. According to the company’s announcement, it offers lossless compression with no quantization or accuracy loss, up to 1.5x compression on large language model (LLM) weights, and 20-35% effective bandwidth improvement (up to 50 percent for some workloads).
‘What sets an AI system apart today is how much performance and context you can pull from the memory you already have,’ said Brett Cline, chief executive of ZeroPoint. ‘ZeroStream lets chip designers do more with that same memory, increasing tokens per second, enabling more context held in memory, and fitting larger foundation models in a smaller footprint.’
ZeroPoint was founded in 2016 by HiPEAC founding member Per Stenström and HiPEAC member Angelos Arelakis, building on years of research at Chalmers University of Technology. The commercialization of the underlying technology was recognized by a HiPEAC Technology Transfer Award in 2016.