AMD Helios: 72 GPUs and 31 TB of HBM4 to Challenge Nvidia

What Happened?

AMD has revealed Helios, a rack-scale system that integrates 72 Instinct MI455X GPUs based on the CDNA 5 architecture, equipped with 31 TB of HBM4 memory. According to The Next Web, the system delivers 2.9 exaflops of FP4 inference performance. Helios is AMD's direct response to Nvidia's NVL72, a similar rack with 72 Blackwell GPUs. The unveiling took place during AMD's Advancing AI 2025 event, where the company also announced the CDNA 5 architecture and the MI400 product line. Helios uses 18 compute trays, each with 4 MI455X GPUs, interconnected via fourth-generation Infinity Fabric, providing 1.6 TB/s of bandwidth between GPUs. The entire system consumes approximately 120 kW, according to analyst estimates, placing it in the range of Nvidia's high-density racks.

Why It Matters

Helios marks AMD's first foray into rack-scale systems for AI, a segment dominated by Nvidia. By offering an alternative with HBM4 memory and high bandwidth, AMD aims to attract customers who need scalability without relying on Nvidia. This could accelerate the adoption of AMD hardware in hyperscale data centers. HBM4 memory, offering 2 TB/s per GPU, is a novelty in the industry: while Nvidia uses HBM3e in its Blackwell GPUs, AMD is leapfrogging with the next generation of high-bandwidth memory. According to JEDEC data, HBM4 doubles the bandwidth per stack compared to HBM3e, enabling Helios to achieve an aggregate bandwidth of 144 TB/s (2 TB/s x 72 GPUs). This is critical for large-scale inference workloads, such as language models with hundreds of billions of parameters.

Market Implications

• Direct competition: Helios directly competes with Nvidia's NVL72, which uses 72 Blackwell GPUs. AMD claims its system offers higher memory capacity (31 TB vs. 24 TB in Nvidia). However, Nvidia holds a significant software advantage with CUDA and its optimized libraries (TensorRT, Triton Inference Server). AMD has improved ROCm to compete, but according to public benchmarks, inference performance on models like Llama 3.1 405B is 15-30% lower on AMD hardware, depending on optimization. Helios could incentivize developers to optimize for AMD, especially if the total cost of ownership (TCO) is lower.
• Software ecosystem: AMD has released ROCm 6.3 with FP4 support and optimizations for large language models. However, ROCm's data center market share is below 5%, compared to CUDA's 90%. Helios could be a catalyst if AMD offers incentives for ISVs and hyperscalers to port their workloads. Companies like Microsoft have already adopted MI300X in their Azure ND H100 v5 instances, but transitioning to MI455X will require engineering investment.
• Price and availability: No pricing has been revealed, but Helios is expected to be available in the second half of 2026. If AMD achieves a competitive price, it could erode Nvidia's market share. According to SemiAnalysis estimates, the cost of an NVL72 rack is around $3 million, with a 60% profit margin for Nvidia. If AMD offers Helios at a 20% lower price, it could attract price-sensitive customers like AI startups and regional data centers. However, volume production of HBM4 will not mature until 2026, which could limit initial availability.

What Readers Should Know

Helios uses 18 compute trays, each with 4 MI455X GPUs. HBM4 memory offers 2 TB/s bandwidth per GPU, according to AMD sources. However, actual performance will depend on workloads and software optimization. Speculation: AMD may be preparing a 144-GPU version for 2027, though unconfirmed. Additionally, Helios includes an integrated network switch based on 800 Gbps Ethernet, allowing scaling to multiple racks without proprietary interconnects like Nvidia's NVLink. This could be a key differentiator for customers preferring open standards. In terms of energy efficiency, AMD claims Helios delivers 24 exaflops per kW in FP4, 15% better than NVL72, though these are lab figures and need validation in real-world environments.

"Helios represents a paradigm shift in AMD's strategy, moving from selling chips to offering complete turnkey systems." — Analyst at TheVortiq

Historical Context

AMD has been closing the gap with Nvidia in AI hardware since the MI300 series launch. With Helios, AMD offers an integrated solution that competes not only on specs but also on ease of deployment. The adoption of HBM4, still in its early stages, could give AMD a temporary advantage if Nvidia does not quickly migrate to this memory. Historically, AMD has been a follower in the AI GPU market, but with the acquisition of Xilinx and integration of FPGAs into its solutions, the company has improved its heterogeneous computing capabilities. Helios also benefits from the CDNA 5 architecture, which introduces native support for FP4 and FP6, optimizing inference for quantized models. In comparison, Nvidia introduced FP4 with Blackwell, but its implementation requires specific hardware. If AMD achieves early adoption of HBM4, it could pressure Nvidia to accelerate its memory roadmap. However, Nvidia has already announced Rubin with HBM4 for 2026, so the window of opportunity is narrow. Helios's success will depend on AMD's ability to deliver a robust software ecosystem and competitive pricing, as well as the maturity of the HBM4 supply chain.