ZeroPoint Technologies Introduces ZeroStream for AI Memory Compression

ZeroPoint Technologies AB announced in a press release the launch of ZeroStream, a hardware IP that improves effective memory bandwidth and performance across AI, data center, and edge computing applications.

ZeroStream uses hardware-accelerated compression to move more data through existing memory and streaming interfaces. It offers lossless compression without quantization or accuracy loss, achieving up to 1.5 times compression on large language model weights and double on activations and KV cache. The company reports a 20 to 35 percent effective bandwidth gain, with some workloads reaching 50 percent improvement. The compression operates with deterministic latency and a fully pipelined structure.

ZeroPoint stated that ZeroStream enables chip designers to deliver more tokens per second from AI accelerators and fit larger foundation models in limited memory. The launch coincides with an update to ZeroPoint’s product range, now organized under four main areas: ZeroStream, ZeroConnect, ZeroStorage, and ZeroAI.

ZeroPoint Technologies is based in Gothenburg, Sweden, and develops hardware acceleration solutions to expand existing memory bandwidth for AI and data-intensive environments.