d-Matrix today announced d-Matrix Squadron, a fleet of multi-rack systems delivering megawatts of efficient inference compute for AI models at data center scale. Built on d-Matrix's memory-centric architecture, Squadron clusters serve AI labs, enterprises, and sovereign organizations running latency-sensitive applications like coding, personal agents, cybersecurity, and high-frequency trading. Developed with Arista, Broadcom, and Supermicro, Squadron is deployed in d-Matrix's Arizona data center and now shipping to priority customers. The Corsair servers and Squad Racks are also generally available.
"The journey to enable customers with our memory-centric compute advantage coincides with d-Matrix's evolution from a specialized chip company to a data center compute infrastructure business," said Sid Sheth, Founder and CEO of d-Matrix, who unveiled Squadron at the OCP Global Summit in San Jose. "With Squadron shipping, and our Corsair servers and Squad Racks now generally available, the wait is over."
Squadron scales up to 16 Squad Racks—single eight-server racks—supporting up to 1,024 d-Matrix Corsair XPUs. Clusters use Jet Stream NICs and Aviator software, and are air-cooled on industry standards, integrating into existing data centers without liquid cooling.
During his keynote, Sheth revealed d-Matrix successfully implemented Attention-FFN Disaggregation (AFD) using Squadron paired with GPU servers. Offloading Experts/FFN to Corsair racks yielded 845 tokens per second per user, with 64 concurrent users reaching 20,415 tokens per second total throughput.
Source: TechPowerUp News · Summarized by HeadlinesBriefing