DRACO 72× MI455X Rack-Scale AI System

Overview

Rack-scale AI without the interconnect lock-in. Seventy-two AMD Instinct MI455X accelerators in a single liquid-cooled rack, joined by an all-Ethernet, open-standards fabric rather than a proprietary one. Built for organisations that want frontier-class capacity and a second source of silicon — and who would rather their next decade of AI infrastructure not depend on one vendor’s roadmap.

Key highlights

• 72× AMD Instinct MI455X (CDNA 5, 2 nm) — AMD’s frontier-class accelerator.
• ~31 TB unified HBM4 per rack — a memory-capacity advantage AMD leads on.
• 432 GB HBM4 per GPU — materially more per-accelerator memory than the Blackwell generation.
• All-Ethernet open fabric — no proprietary scale-up interconnect; standard optics and switching.
• ~2.9 exaFLOPS FP4 per rack (AMD-published) for low-precision inference and training.
• Liquid-cooled rack-scale integration, delivered and commissioned by RDP.
• Sovereign-capable: deployable on-soil, air-gap possible, DPDP-aligned.

AI workload fit

• Frontier LLM training at trillion-parameter scale.
• Full-parameter fine-tuning of very large models.
• High-throughput inference at FP4/FP8 for national or multi-tenant services.
• HPC & AI convergence — AMD’s traditional strength in FP64 science.
• Sovereign AI programmes seeking supply-chain diversity.
• The compute core of an AI factory built on open standards.

AI workload positioning

This is the credible open-standards answer to proprietary rack-scale AI. The strategic argument is memory capacity and fabric openness: more HBM per accelerator, and Ethernet rather than a closed scale-up fabric, which keeps switching, optics and future upgrades in a competitive market. The honest counter-argument is software maturity — CUDA has a deeper ecosystem than ROCm, and any migration carries real engineering cost. RDP will quantify that cost for your specific stack rather than wave it away. All AMD performance and capacity figures quoted here are AMD-published and are presented as such — they are not RDP measurements. Availability follows AMD’s stated schedule and allocation; RDP confirms silicon availability, delivery window and final configuration in writing at quotation.

Industry use cases

• Public sector & sovereign: national AI capability with deliberate supply-chain diversification.
• Neocloud: differentiated capacity and better memory-per-GPU economics for tenants.
• Research & education: HPC + AI convergence where FP64 matters alongside AI.
• Defence & aerospace: strategic independence from a single accelerator vendor.
• Enterprise & GCCs: hedging a multi-year AI infrastructure commitment.

Performance & how to be sure

AMD’s headline for this generation is memory: 432 GB HBM4 per GPU and ~31 TB per rack, which is what lets very large models stay resident without aggressive sharding. Rather than quote a tokens/sec figure that will not match your workload, RDP offers a “benchmark your model” session: bring the model, precision and context length you actually intend to run, and we will validate it on this exact configuration before you commit.

Series & upgrade path

Below: RDP’s 8× SXM single nodes (NVIDIA) and AMD Instinct servers. Beside: the NVIDIA GB300 NVL72 rack-scale system — the direct comparison, and RDP quotes both honestly. Above: multi-rack AMD and NVIDIA superclusters and the GB300 NVL72 Supercluster containerised node.

On-prem vs cloud (TCO)

For sustained daily AI work this configuration removes per-GPU-hour billing, queueing for scarce instances, and egress charges on your own data — and keeps everything on-premises for DPDP and data-residency obligations. Cloud still wins for burst capacity and one-off very large training runs; on-prem wins on sustained utilisation, control and predictable INR capital cost. We model the crossover with your actual usage rather than assert it.

Software & day-one readiness

Software stack is ROCm-based: ROCm runtime and libraries, PyTorch and JAX support, vLLM and SGLang for serving, Kubernetes/Slurm orchestration, plus cluster management and telemetry. Migration from CUDA is a real engineering exercise and RDP scopes it honestly as part of the proposal rather than treating it as a drop-in swap.

Power, thermal & acoustics

Full liquid-cooled rack with in-row CDU and high-capacity busway feeds. Facility power, cooling capacity, floor loading and water-loop readiness are surveyed before order — this is a datacenter project, not a delivery. Exact wattage, BTU and dB(A) figures come from RDP bench measurement rather than estimates — ask for the site-readiness sheet with your quote.

Deployment, warranty & support

Built to order by RDP Technologies. Supplied with GST invoice (HSN 8471), pan-India onsite support, and availability through GeM for public-sector procurement. Built to order — allocation-dependent; AMD MI400-series systems ship from 2H-2026. Confirmed at quotation.

Why RDP

RDP Technologies is a Make-in-India OEM with 14+ years and 300,000+ units shipped, supplying AI infrastructure from desk-side systems to rack-scale AI factories — with predictable INR pricing, GST invoicing and pan-India onsite service.

Buy with confidence

Use Request a Quote to reach an RDP solution architect for sizing, a benchmark-your-model session, financing options and a delivery plan. No obligation.

Specifications