Deployed at full-rack scale via Vultr Cloud GPU and Vultr Bare Metal, the AMD Helios™ rackscale solution integrates 72 AMD Instinct™ MI455X GPUs and 18 AMD EPYC™ CPUs. This open foundation optimizes the entire agentic stack—driving swift task orchestration, fleet-scale model inference, and high-capacity memory/state retention. Providing up to 432 GB of HBM4 memory per GPU (31 TB per rack), the Helios solution accelerates complex reasoning pipelines under a predictable, transparent pricing model.
Capacity available soon
Reserve capacity at Vultr now to access the latest in AMD acceleration.
Key features
Integrates 72 AMD Instinct™ MI455X GPUs and EPYC™ CPUs. Deployed via Vultr for exaflop-class peak AI compute.
Features up to 432 GB HBM4 and 23.3 TB/s bandwidth per GPU, providing 31 TB of total high-speed GPU memory.
Powered by low-latency UALink™ over Ethernet scale-up links inside the rack paired with standards-based UEC-aligned Ethernet scale-out clustering.
Vultr delivers the AMD Helios™ rackscale solution at full-rack or dedicated data-center scale. This deployment unifies your agent orchestration, high-volume model inference, and memory/state layers using 72 AMD Instinct™ MI455X GPUs. Achieve predictable financial modeling across your entire AI footprint without proprietary stack dependencies.
Streamline the agentic stack using automated deployment tools on Vultr. We wrap dedicated AMD Helios™ rackscale solutions in turnkey Vultr Clusters featuring pre-integrated Slurm or Kubernetes schedulers. This unifies orchestration, inference, and memory pipelines globally, eliminating multi-vendor integration complexity.
Vultr hardens dedicated full-rack deployments by shielding the orchestration, inference, and memory layers through isolated cloud networking. The AMD Helios™ rackscale solution adds hardware-level runtime attestation and encrypted link communication directly inside the rack. This multi-layered architecture enforces rigorous data sovereignty within physically secure data centers.
The AMD Helios™ rackscale solution delivers dedicated full-rack or data center capacity for complex multi-agent workflows
Deploying the Helios solution via Vultr Cloud GPU and Vultr Bare Metal seamlessly unites the orchestration, inference, and memory/state layers of the agentic stack. Leverage full-rack AMD Instinct™ MI455X GPUs on Vultr's global platform to scale agentic AI and inference applications with predictable, transparent pricing and no proprietary cloud lock-in.
An AMD Helios™ rackscale solution with 72 AMD Instinct™ MI455X GPUs using the peak Matrix/Vector FP32 datatypes offers up to 2.0X or 100% better peak theoretical precision performance per GPU, compared to the 72x AMD Instinct™ MI355X GPU configuration.
An AMD Helios™ rackscale solution with 72 AMD Instinct™ MI455X GPUs offers up to 1.5X the memory capacity, or 50% more memory and up to 2.91X or 191% higher peak memory bandwidth compared to a 72x AMD Instinct™ MI355X GPU configuration.
Powered by the 5th Gen AMD CDNA™ architecture, the AMD Instinct™ MI455X GPU (432GB) is equipped with sufficient memory to run the Generative Pre-trained Transformer 3 (GPT-3) Large Language Model (LLM) at 175B parameters on one (1) GPU using FP16 datatype, while the AMD Instinct™ MI355X GPU (288GB) requires two (2) GPUs.
*Based on calculations by AMD Performance Labs in June 2026
Get AI pilots into production with control and speed via the AMD Helios™ rackscale solution, featuring GPU acceleration, CPU orchestration, open software, and scale-up and scale-out networking. See our datasheet to learn how.
AMD Helios™ Rackscale Solution
AMD Instinct™ MI455X GPUs
72
AMD EPYC™ CPUs
18
GPU Memory
Up to 432 GB HBM4 per GPU; up to 31 TB HBM4 per rack
System Memory
18.4 TB
Total Memory Bandwidth
Up to 23.3 TB/s per GPU; up to 1.67 PB/s per rack
Peak Open Compute Project MXFP8 Performace (TFLOPS)
20133
Peak Matrix FP16 Performance with Structured and Sparsity (TFLOPS)
10066
Scale-up/Scale-out
260 TB/s UALink™ over Ethernet/43 TB/s UEC-aligned Ethernet
Interconnect
AMD Pensando™ networking / AMD Pensando™ AI-NICs
AMD Helios™ is an integrated full-rack- or data center-scale infrastructure solution comprising 72 AMD Instinct™ MI455X GPUs, 18 AMD EPYC™ CPUs, up to 31 TB of HBM4 memory capacity, scale-up and scale-out AMD Pensando™ networking, open ROCm™ software, and direct liquid cooling (DLC) options.
AMD Instinct™ MI455X GPUs are accelerators of the AMD Instinct™ MI400 Series. They’re built on AMD CDNA™ 5 architecture, which is designed for reduced data movement overhead and improved power efficiency.
AMD Instinct™ MI455X GPUs are designed specifically for the AMD Helios™ rackscale solution. They carry a peak engine clock of 2400 MHz, a dedicated memory size of 432 GB, and a peak memory bandwidth of 23.3 TB/s.
The AMD Helios™ rackscale solution enables enterprises to deploy and scale their agentic AI, fine-tuning, training, and inference workloads with ease at massive scale. AMD Instinct™ MI455X GPUs deliver advanced performance and efficiency for demanding workloads, backed by high HBM4 memory capacity and exceptional memory bandwidth. Direct liquid cooling provides outstanding thermal management and energy efficiency, while ROCm™ software optimizes throughput, scalability, and AI and HPC performance.
Scale-up and scale-out AMD Pensando™ networking features 260 TB/s UALink™ over Ethernet and 43 TB/s UEC-aligned Ethernet to meet high scaling demands.
Vultr enables teams to deploy the AMD Helios™ rackscale solution with top-tier flexibility, control, and price-to-performance globally, with:
Vultr is currently taking pre-orders for reserved capacity from 2027 through 2028. Contact our sales team.
AMD Helios™ is an integrated full-rack- or data center-scale infrastructure solution comprising 72 AMD Instinct™ MI455X GPUs, 18 AMD EPYC™ CPUs, up to 31 TB of HBM4 memory capacity, scale-up and scale-out AMD Pensando™ networking, open ROCm™ software, and direct liquid cooling (DLC) options.
