The Pulse
d-Matrix Starts Shipping Squadron Clusters With Up to 1,024 XPUs
d-Matrix says its new Squadron inference clusters are shipping to priority customers and can scale across 16 racks to 1,024 Corsair XPUs. The air-cooled systems pair d-Matrix accelerators with JetStream networking and Aviator software, and
AI.info Team · Written with AI ·
d-Matrix is pitching Squadron as a data-center-scale inference system, but its first deployments are going to priority customers rather than an unrestricted market. The company announced the multi-rack clusters on October 12 at the OCP Global Summit in San Jose, saying they are shipping to customers while Corsair servers and SquadRacks are generally available. The largest configuration connects up to 1,024 Corsair XPUs.
Sixteen racks, up to 1,024 XPUs
A Squadron cluster can combine as many as 16 SquadRacks, each holding eight servers, according to d-Matrix. JetStream network interface cards connect the systems, while the company’s Aviator software is intended to support deployment. Customers can start with an individual Corsair server or a single SquadRack before expanding to a multi-rack cluster.
d-Matrix says the systems are air-cooled and use industry-standard designs, so data-center operators can install them in existing facilities without adding liquid cooling. The company describes Squadron as megawatt-scale inference infrastructure for latency-sensitive uses such as coding, personal agents, cybersecurity and high-frequency trading. It did not provide a price or a customer name in its announcement.
From Corsair-only models to GPU-XPU workloads
The announcement describes two deployment approaches. Squadron can run models entirely on Corsair XPUs, or work alongside GPU servers in a disaggregated setup that assigns different parts of the workload to different processors. At his OCP keynote, d-Matrix founder and CEO Sid Sheth said the company had implemented Attention-FFN Disaggregation, or AFD, on Squadron paired with GPU servers.
“With Squadron shipping, and our Corsair servers and SquadRacks now generally available, the wait is over.”
Sid Sheth, founder and CEO of d-Matrix
In the AFD demonstration described by d-Matrix, the system offloaded the experts or feed-forward network portion of the workload to Corsair racks. The company reported 845 tokens per second per user and total throughput of 20,415 tokens per second with 64 concurrent users. The announcement does not identify the model or provide a comparison baseline, so those figures describe d-Matrix’s demonstration rather than an independently comparable performance result.
Priority shipments put the rollout to the test
d-Matrix says it has already installed Squadron in its Arizona data center for internal research and development, and that systems are now going to priority customers for deployment in their own facilities. Arista, Broadcom and Supermicro worked with the company on the system. Supermicro president and CEO Charles Liang said the collaboration draws on the companies’ engineering and systems-integration work to bring the infrastructure to market.
The distinction between availability and shipment matters: d-Matrix says the Corsair servers and SquadRacks are generally available, while Squadron shipments are initially directed to priority customers. The announcement gives no customer identities, prices or delivery volumes. Its next public demonstration is scheduled at d-Matrix’s OCP booth on October 14 and 15, where company researchers plan to show the AFD setup.