Huawei unveils architecture for operating up to 1 million AI chips — Live Science
Huawei has unveiled the Peerium architecture, which, according to the company, could eventually make it possible to combine up to one million artificial intelligence chips into a single computing system. The development was presented on September 17 at the annual Huawei CONNECT conference, Live Science reports.
Typically, chips used to train foundation AI models are placed on separate servers that transmit data over a network and distribute computing tasks. According to Huawei, Peerium is intended to support fast data exchange between chips across several server racks even after the system is scaled to hundreds of thousands of devices.
Three components of Peerium
Huawei said the architecture is based on nested parallelism, unified memory addressing, and direct connections between components. Nested parallelism makes it possible to split large computing tasks into smaller ones and distribute them among hierarchical groups of chips operating in parallel.
More current news is available on the UA.News Telegram channel Telegram.
Unified memory addressing is intended to allow processors to access the memory of other servers rather than only their own. Direct connections between components, in turn, are intended to enable data exchange without intermediate links. Huawei says these approaches should help use computing resources more efficiently for training and inference of AI models.
UnifiedBus technology
For Peerium, Huawei developed its own interconnection technology, UnifiedBus. The company's rotating chairman, Eric Xu, told the Associated Press that its bandwidth between server racks could reach 800 GB per second. For comparison, he cited 200 GB per second for a configuration with NVLink within racks and InfiniBand between them.
Huawei representatives also said that Peerium-based systems will be able to train models with up to 40 trillion parameters. The company simultaneously announced the Atlas 950 SuperPoD, a system with more than 8,000 Ascend 950DT AI accelerators. Huawei expects to begin delivering the SuperPoD no later than the beginning of next year. One SuperCluster with 256,000 computing cards is already being deployed and tested, according to the company.
