AMD has introduced Helios, marking its entry into the rack-scale AI compute market. This platform is designed to rival Nvidia’s systems, particularly the Vera Rubin, with claims of superior specifications and performance.
Specifications and Design
The Helios system features a 72-GPU configuration, significantly larger than Nvidia’s NVL72, with dimensions of 1.2 meters wide and 44OUs high. This allows AMD to incorporate 50% more HBM4 memory and enhanced scale-out bandwidth compared to Vera Rubin, along with a performance increase of 15 to 25% for AI training tasks.
GPU Architecture
At the core of Helios is the Instinct MI455X, built on AMD’s 5th-generation CDNA architecture. This GPU utilizes a chiplet design, comprising 24 chiplets that include eight compute dies fabricated on TSMC’s advanced 2nm process technology. The architecture allows for flexible configurations, enabling the chip to operate as a single large GPU or multiple smaller ones.
Networking and Power Efficiency
The Helios architecture employs Ultra Accelerator Link over Ethernet (UALoE) for inter-GPU communication, utilizing Broadcom’s Tomahawk 6 switch ASICs. Each MI455X GPU is supported by three Vulcano smartNICs, providing a total of 2.4 Tbps of bandwidth per accelerator. The system is powered by a liquid-cooled DC bus bar capable of delivering between 225 and 245 kW, potentially offering better power efficiency than Nvidia’s systems.
Performance Metrics
While AMD claims that Helios can achieve peak theoretical performance close to 40 petaFLOPS, real-world performance is expected to be around 20 petaFLOPS for FP4 workloads. AMD has shifted its marketing focus to emphasize measured performance, acknowledging that theoretical figures often do not reflect actual capabilities.
With Helios, AMD aims to capture a significant share of the AI compute market, already attracting major clients like Meta and Microsoft. The company is also developing additional CDNA 5-based GPUs for various applications, indicating a strategic push into both AI and high-performance computing sectors.
This article was produced by NeonPulse.today using human and AI-assisted editorial processes, based on publicly available information. Content may be edited for clarity and style.
