12:42Machine-translated from Azerbaijani by AIRead in Azerbaijani
AMD Unveils Helion Server Rack With 32,768 Zen 6 Cores
AMD has developed a 21-inch standard server rack called Helion for data centers and autonomous AI agents. The processor-only version can operate with 32,768 Zen 6 cores. It uses 128 EPYC Venice 9996 processors, each with 256 cores. Helion will be released in the first half of 2027.
The design is based on the Open Compute Project organization's ORv4 standard. The rack is slightly wider and taller than conventional IT racks and contains 44 OpenU modules. The system can be cooled either by air or by delivering liquid directly to the chips. Helion also complies with the DC-MHS standard, which was developed to enable different types of equipment to work together in data centers. Helion will not replace the Helios rack shown last year but will complement it. Helios is designed for training AI models and processing queries. In the new design, the processors handle most of the work. For example, the system can perform tasks such as software development and send intensive AI computations to a separate server. AI agents in a shared system, where each user is allocated a separate computing environment, can also benefit from this configuration.
The design also has a hybrid version. This variant uses 16 MI350P GPUs and 102 Tbps networking, but the number of main processors is reduced. AMD developed it for companies that want to combine data management, networking and AI workloads in the same rack. This will allow enterprises to use a single hardware system instead of separate systems. AMD will not sell finished server racks itself. The company will share the reference design through the OCP, while hardware partners will handle manufacturing and sales. The partners were not named during the initial presentation. Their names are expected to be announced during a presentation by company CEO Lisa Su. The new rack is scheduled to reach the market after the Zen 6 server platform, which is expected to become widely available later this year.
The design is based on the Open Compute Project organization's ORv4 standard. The rack is slightly wider and taller than conventional IT racks and contains 44 OpenU modules. The system can be cooled either by air or by delivering liquid directly to the chips. Helion also complies with the DC-MHS standard, which was developed to enable different types of equipment to work together in data centers. Helion will not replace the Helios rack shown last year but will complement it. Helios is designed for training AI models and processing queries. In the new design, the processors handle most of the work. For example, the system can perform tasks such as software development and send intensive AI computations to a separate server. AI agents in a shared system, where each user is allocated a separate computing environment, can also benefit from this configuration. The design also has a hybrid version. This variant uses 16 MI350P GPUs and 102 Tbps networking, but the number of main processors is reduced. AMD developed it for companies that want to combine data management, networking and AI workloads in the same rack. This will allow enterprises to use a single hardware system instead of separate systems. AMD will not sell finished server racks itself. The company will share the reference design through the OCP, while hardware partners will handle manufacturing and sales. The partners were not named during the initial presentation. Their names are expected to be announced during a presentation by company CEO Lisa Su. The new rack is scheduled to reach the market after the Zen 6 server platform, which is expected to become widely available later this year.