AMD Unveils Helios AI Rack System at Advancing AI 2026, Claiming 30% Better Price-Performance Over NVIDIA Vera Rubin NVL72At its Advancing AI 2026 event, AMD officially launched AMD Helios, a fully integrated, liquid-cooled rack-scale server system designed specifically for large-scale AI "token factories" running next-generation frontier models. During the announcement, AMD directly pitted the Helios architecture against NVIDIA’s Vera Rubin NVL72 platform, boasting a 30% advantage in price-performance efficiency.
Following initial teasers revealed at CES earlier this year, the Helios system is now in full-scale volume production. Major hyperscalers and frontier AI labs including Microsoft, Meta, OpenAI, and Oracle have already committed to deploying Helios infrastructure within their global data center footprints.
Helios Architecture & Hardware Specifications
The Helios rack platform combines AMD’s full data center stack, integrating 6th Gen AMD EPYC "Venice" CPUs, 5th Gen AMD Instinct MI455X GPUs, and AMD Pensando network acceleration processors.
Cluster Density: Houses 18 compute trays, each containing 4 MI455X GPUs, totaling 72 GPUs per rack.
Interconnect: Interlinked via high-throughput UALink over Ethernet (UALoE) networking.
Compute Power: Delivers 2.9 Exaflops of FP4 performance and 1.4 Exaflops of FP8 compute capability.
Memory Footprint: Features 31TB of unified memory with a massive combined memory bandwidth of 1.7 Petabytes per second (PB/s).
Data Types Supported: Native support for MXFP4, MXFP6, MXFP8, and FP8 formats (competing directly against NVIDIA’s NVFP4 standard).
Thermal & Power Design: Full direct-to-chip liquid cooling with a maximum power envelope of 250kW per rack.
Industry Endorsement: Anthropic on Porting Ease
Early testing feedback has been remarkably positive. Tom Brown, co-founder of Anthropic, shared that his team conducted extensive pre-production testing on Helios hardware. He highlighted that porting and optimizing large-scale models to the Helios platform proved unexpectedly straightforward, noting that a single engineer leveraging AI-assisted developer tools—was able to continuously optimize model throughput over a single weekend to unlock near-full hardware performance.
The sheer power density of modern AI racks, such as the Helios, which operates at 250kW per rack, represents a significant advance in data center thermal engineering. Traditional air-cooled data centers have a peak power output of around 15–20kW per rack. Achieving 250kW requires 100% direct-to-chip liquid cooling, forcing large cloud providers like Azure and Oracle to completely redesign their data center power and cooling infrastructure to accommodate these next-generation token manufacturing plants.
AMD's commitment to UALink over Ethernet (UALoE) is noteworthy, as while NVIDIA heavily relies on its proprietary NVLink/NVSwitch ecosystem to connect 72 GPUs, AMD takes an open standards alliance approach using UALink and Pensando Ethernet. This open architecture appeals to cloud providers seeking to avoid complete vendor monopoly at the physical network level.
The competition in FP4 and FP8 processing has shifted to microscaled models (MX Specs). With direct support for MXFP4 and MXFP8, AMD aligns with the Open Compute Project (OCP) standard for low-precision inference and training. Low-precision quantification allows models with hundreds of billions of parameters to be stored in memory without losing contextual precision, significantly reducing the cost per generated token.
AMD Unveils Helios AI Rack System at Advancing AI 2026, Claiming 30% Better Price-Performance Over NVIDIA Vera Rubin NVL72At its Advancing AI 2026 event, AMD officially launched AMD Helios, a fully integrated, liquid-cooled rack-scale server system designed specifically for large-scale AI "token factories" running next-generation frontier models. During the announcement, AMD directly pitted the Helios architecture against NVIDIA’s Vera Rubin NVL72 platform, boasting a 30% advantage in price-performance efficiency.
Following initial teasers revealed at CES earlier this year, the Helios system is now in full-scale volume production. Major hyperscalers and frontier AI labs including Microsoft, Meta, OpenAI, and Oracle have already committed to deploying Helios infrastructure within their global data center footprints.
Helios Architecture & Hardware Specifications
The Helios rack platform combines AMD’s full data center stack, integrating 6th Gen AMD EPYC "Venice" CPUs, 5th Gen AMD Instinct MI455X GPUs, and AMD Pensando network acceleration processors.
Cluster Density: Houses 18 compute trays, each containing 4 MI455X GPUs, totaling 72 GPUs per rack.
Interconnect: Interlinked via high-throughput UALink over Ethernet (UALoE) networking.
Compute Power: Delivers 2.9 Exaflops of FP4 performance and 1.4 Exaflops of FP8 compute capability.
Memory Footprint: Features 31TB of unified memory with a massive combined memory bandwidth of 1.7 Petabytes per second (PB/s).
Data Types Supported: Native support for MXFP4, MXFP6, MXFP8, and FP8 formats (competing directly against NVIDIA’s NVFP4 standard).
Thermal & Power Design: Full direct-to-chip liquid cooling with a maximum power envelope of 250kW per rack.
Industry Endorsement: Anthropic on Porting Ease
Early testing feedback has been remarkably positive. Tom Brown, co-founder of Anthropic, shared that his team conducted extensive pre-production testing on Helios hardware. He highlighted that porting and optimizing large-scale models to the Helios platform proved unexpectedly straightforward, noting that a single engineer leveraging AI-assisted developer tools—was able to continuously optimize model throughput over a single weekend to unlock near-full hardware performance.
The sheer power density of modern AI racks, such as the Helios, which operates at 250kW per rack, represents a significant advance in data center thermal engineering. Traditional air-cooled data centers have a peak power output of around 15–20kW per rack. Achieving 250kW requires 100% direct-to-chip liquid cooling, forcing large cloud providers like Azure and Oracle to completely redesign their data center power and cooling infrastructure to accommodate these next-generation token manufacturing plants.
AMD's commitment to UALink over Ethernet (UALoE) is noteworthy, as while NVIDIA heavily relies on its proprietary NVLink/NVSwitch ecosystem to connect 72 GPUs, AMD takes an open standards alliance approach using UALink and Pensando Ethernet. This open architecture appeals to cloud providers seeking to avoid complete vendor monopoly at the physical network level.
The competition in FP4 and FP8 processing has shifted to microscaled models (MX Specs). With direct support for MXFP4 and MXFP8, AMD aligns with the Open Compute Project (OCP) standard for low-precision inference and training. Low-precision quantification allows models with hundreds of billions of parameters to be stored in memory without losing contextual precision, significantly reducing the cost per generated token.
Comments
Post a Comment