Huawei Announces Peerium AI Computing Architecture; Atlas 950 Cluster Is in Deployment
Редакция ЦОД:Главное
Huawei has introduced Peerium, an AI computing architecture it says can coordinate processors at million-scale as a single computing system. The company said a 256,000-card Atlas 950 cluster is being deployed, while an NPO-based Atlas 960 system is under testing.
- Huawei says Peerium can coordinate processors at million-scale as one computing system.
- Atlas 950 is Huawei's first-generation product on the Peerium architecture.
- A 256,000-card Atlas 950 cluster is in deployment, not confirmed as operational.
- Huawei identifies Lingqu as the key interconnect technology for Peerium.
- Atlas 960, based on NPO, remains in testing.
Huawei introduced its Peerium AI computing architecture in Shanghai on September 17. The company positions the architecture for large-scale AI computing and says it can enable processors at million-scale to operate as one computing system.
According to Huawei, Peerium combines nested parallelism, unified memory addressing and a peer-to-peer interconnect. Huawei identified the Lingqu interconnect technology as a core element of the architecture. Atlas 950 is the first-generation product based on Peerium.
Huawei said a 256,000-card Atlas 950 supernode cluster is in deployment. This is an IT-scale metric; Huawei did not disclose data-center power capacity, energy consumption, PUE, physical location, CAPEX or a completion schedule for the deployment. The cluster should therefore not be treated as operational.
The company also said that its NPO-based Atlas 960 system is under testing. Huawei has not described it as commercially deployed or in service.
For data-center operators and infrastructure suppliers, the significance is in the direction of travel rather than a disclosed facility build-out. If the stated capabilities are confirmed in real-world deployments, Peerium-based systems could increase demand for high-performance interconnect fabrics, tightly integrated server platforms, power infrastructure and cooling designed for exceptionally dense AI workloads.
If validated in practical deployments, Peerium could raise infrastructure requirements for very dense AI training and inference environments. Architectures built around large-scale, peer-to-peer interconnects place greater demands on the data-center network fabric, power delivery, cooling and system integration.
- IT之家 (IT Home) / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.