Designed specifically for frontier AI, large-scale inference and base model training, AMD Helios offers a complete rack-scale architecture. . New analysis; Unites AMD Instinct MI455X GPUs, 6th Generation AMD EPYC server processors, AMD Pensando networking technologies and open source AMD ROCm software under one roof .

Summary Points of the News

• All Hardware in One Architecture:AMD Helios; It offers Instinct MI455X GPU, 6th Generation EPYC CPU, Pensando network infrastructure and ROCm software as an integrated cabinet solution .

• Record Performance Values:A single Helios cabinet provides 2.9 exaflops of FP4 process power, 31 TB of HBM4 memory, and 1.7 petabytes per second of memory bandwidth .

• Advantage over Competitors:Helios architecture promises 15% more FP4 performance, 50% more HBM memory capacity and up to 30% cost advantage compared to closest competing solutions .

Architecture Specially Designed for Artificial Intelligence Factories

Traditional server systems have difficulty meeting the heavy loads brought by contemporary artificial intelligence models with trillions of parameters. This situation is driving technology giants to develop cabinet-scale integrated systems instead of single servers. .

AMD Helios is being developed to meet this need. . This architecture, which can be scaled seamlessly from a single cabinet to gigawatt-scale giant artificial intelligence clusters, meets the performance and power efficiency expectations of institutions. .

The 18 open cabinet-aligned 4-GPU trays at the heart of the AMD Helios system enable a total of 72 GPUs to work in full harmony. . This architecture completely eliminates information bottlenecks in distributed inference and model training processes .

AMD Helios in Numbers: Exaflops Power

According to the technical information shared by AMD, Helios sets the standards again in terms of processing capacity and memory bandwidth. . The basic performance costs offered by a single AMD Helios cabinet are listed as follows:

• Processing Power:2.9 exaflops AI performance at Peak FP4 and 1.4 exaflops at FP8 .

• Memory Capacity:31 Terabytes of HBM4 memory .

• Memory Bandwidth:A tremendous information transfer rate of 1.7 Petabytes per second .

Compared to competing solutions in the industry (such as NVIDIA Vera Rubin NVL72), AMD Helios; Delivers 15% more peak FP4 performance, 50% more high-bandwidth memory (HBM) capacity, and 6% more memory bandwidth . It also accelerates communication in large cluster installations by providing 50% more scale-out bandwidth. .

Open Standards and Cost Advantage

Today, one of the most critical factors in artificial intelligence infrastructure investments is the process power obtained per dollar spent, that is, token efficiency. . AMD Helios generates up to 30% more tokens per dollar invested compared to competing solutions . This significantly reduces artificial intelligence data center operating costs. .

Another prominent point of the system is that it is based entirely on open standards instead of closed ecosystems. . Helios uses UALink over Ethernet (UALOE) technology for high-speed in-cabinet connections, and uses AMD Pensando Vulcano 800 AI network cards compatible with Ultra Ethernet Consortium standards for large-scale network connections. . In this way, companies can build flexible and reliable artificial intelligence infrastructures without being subject to hardware lock. .