AMD has introduced its new Helios rack-scale AI platform, featuring high-performance MI455X accelerators, while simultaneously securing a partnership with Cerebras to integrate wafer-scale chips into its data center systems.
The Launch of Helios Rack Systems
AMD has officially entered the rack-scale AI competition with the launch of Helios, a platform designed to challenge Nvidia’s data center market share. The system is notable for its physical scale, measuring 1.2 meters wide and 44OUs high, which makes it significantly larger than Nvidia's NVL72 architecture. Within the Helios rack, AMD utilizes 18 liquid-cooled compute blades, each housing four MI455X GPUs and a single 96-core Venice Epyc processor. By utilizing an Ultra Accelerator Link over Ethernet (UALoE) fabric, the system avoids the need for proprietary switches, allowing builders to leverage industry-standard merchant silicon from companies like Broadcom. With each MI455X accelerator supported by three 800 Gbps Pensando Vulcano network cards, AMD claims the platform offers superior scale-out bandwidth compared to Nvidia's existing offerings, positioning it as a powerful contender for large-scale training and inference workloads.
Technical Architecture of the MI455X
The power behind the Helios platform is AMD's latest data center accelerator, the MI455X, which is built on the 5th-gen CDNA compute architecture. This chip represents a massive engineering effort, utilizing a 'silicon sandwich' design that incorporates 24 chiplets and 2nm process technology for compute dies. To prioritize AI performance, AMD has moved away from FP64 support in this specific SKU, reallocating die area to focus on low-precision formats like MXFP4 and MXFP8. The architecture includes a significant increase in shared L2 cache and a new direct memory access (DMA) engine aimed at reducing data latency. These improvements allow the chip to support spatial partitioning, enabling it to function as a single large GPU or be divided into multiple virtual instances, providing hyperscalers with flexible deployment options that were previously limited in earlier generations.
Strategic Partnership with Cerebras
Alongside its internal hardware advancements, AMD has deepened its ecosystem influence through a new partnership with Cerebras. At an industry conference, Cerebras CEO Andrew Feldman confirmed that his company’s unique 'wafer-scale' chips will be utilized within AMD’s Helios AI systems. This integration aims to address the critical industry demand for ultra-low latency, which is essential for generative AI applications requiring rapid response times. Cerebras hardware will begin appearing in Cerebras-managed data centers later this year, and server buyers will have the option to configure their own AMD-based systems with this wafer-scale technology. The companies have pitched this collaboration as a way to achieve significantly higher tokens-per-second-per-watt efficiency, creating a distinct niche in a market otherwise dominated by standard GPU-based clusters.
Market Impact and Competitive Landscape
AMD is positioning Helios as a direct rival to Nvidia’s Blackwell and Vera Rubin platforms, claiming performance leads in memory bandwidth and AI training speeds. By matching the launch timeline of Nvidia’s newest platforms, AMD appears to be moving away from its historical 'catch-up' model. The company has also secured significant commitments from major players: OpenAI plans to deploy large-scale compute power using MI455X GPUs, and Anthropic has committed to significant deployments in exchange for investment deals. Despite the competitive posturing, AMD has adopted a more transparent marketing approach regarding performance metrics. By publicly admitting that peak theoretical FLOPS are rarely achievable in real-world scenarios, the company hopes to build long-term credibility with enterprise customers who prioritize sustained, measurable performance over potentially inflated marketing figures.
⚖ The Balanced View
Supporting view
Supporters note that Helios offers a superior performance-per-dollar ratio, higher HBM4 memory capacity, and faster scale-out bandwidth than current Nvidia equivalents.
Concerns & criticism
Critics point out that real-world AI performance is heavily dependent on software optimization and workload shape, meaning that 'peak theoretical' FLOPS are often functionally impossible to reach.
→What's next
AMD plans to expand its CDNA 5-based GPU portfolio to include variants specifically optimized for enterprise deployments. Meanwhile, Cerebras-equipped Helios systems are expected to begin rolling out in data centers starting later this year.