Hardware startup wants to solve billion-dollar bandwidth problem by using age-old technique: AI memory compression technique could save Google and Microsoft billions, but Nvidia won’t be happy
Swedish company ZeroPoint Technologiesa spin-off from Chalmers University of Technology in Gothenburg, was founded by Professor Per Stenström and Dr. Angelos Arelakis with the aim of providing efficient real-time memory compression for the entire memory system. The company aims to maximize server efficiency by addressing memory bottlenecks, helping hyperscalers such as Microsoft, Meta and Google, as well as large enterprises, save significant costs.
ZeroPoint claims its technology eliminates up to 70% of unnecessary data in microchip memory through a combination of ultra-fast compression, real-time data compaction and efficient memory management. This approach maximizes performance per watt and addresses the long-standing challenge of memory bottlenecks that have hindered performance scaling for decades.
With 38 patents already secured, ZeroPoint offers a hardware IP block for data compression and compaction, accompanied by custom memory management software for integration into CPUs or SoCs. Evaluation typically involves analyzing compression ratios, emulating memory management, and running architectural simulations, all of which can be completed within weeks.
Reduce server costs by up to a quarter
ZeroPoint says its compression technology is 1,000 times faster than traditional solutions and can increase memory capacity by 2-4x while increasing performance per watt by up to 50%, potentially reducing data center server costs by up to 25%.
CEO Klas Moreau stated: “Our memory optimization technologies can increase the efficiency, performance and capacity of enterprise and hyperscale computing applications across a wide range of use cases. As an organization, we are driven by the ambitious mission to make this technology the industry standard. “
Blocks and files reports that ZeroPoint expects to reach $110 million in revenue by 2029, aiming to become a major player in the billion-dollar memory market. The site quotes CEO Klas Moreau as saying: “Hyperscalers pay an absolute fortune for their GPUs and can only use half of them for their AI workloads. We meet them and they tell us that.”
While this technology could bring significant benefits to hyperscalers and large enterprises, companies like Nvidia may be less enthusiastic as it could reduce demand for high-memory GPUs.
ZeroPoint isn’t the only company looking to reduce memory usage on servers. In 2022, Meta explained how a smart memory optimization technique saved the company millions of dollars.
