DRACO 8× B300 HGX GPU Server

Overview

The DRACO 8× B300 HGX GPU Server is a factory-built, ready-to-deploy 8-GPU AI server — 8× NVIDIA B300 (HGX, Blackwell Ultra), 2× Intel Xeon 6 6767P (64-core, 2.4 GHz), 4 TB DDR5-6400 (32× 128 GB RDIMM) — assembled and validated for frontier-scale LLM training and high-throughput inference. It is a specific, in-stock configuration RDP delivers turnkey in India: landed, GST-invoiced, installed and supported, so a CIO gets a running AI node, not a procurement project.

Built on the NVIDIA HGX B300 8-GPU SXM baseboard with all 8 GPUs tightly coupled over NVLink/NVSwitch, it is the class of machine that trains and serves the largest models on-prem — with data sovereignty and no per-hour cloud meter.

Key highlights

• GPUs: 8× NVIDIA B300 (HGX, Blackwell Ultra) — 2,304 GB HBM3e (8× 288 GB), NVLink/NVSwitch-coupled.
• CPU: 2× Intel Xeon 6 6767P (64-core, 2.4 GHz) (128 cores) for data pipeline, orchestration and host workloads.
• Memory: 4 TB DDR5-6400 (32× 128 GB RDIMM).
• Storage: 2× 1.92 TB M.2 + 8× 3.84 TB U.2 NVMe (PCIe 4.0).
• Networking: 8× 800G InfiniBand XDR (ConnectX-8 SuperNIC) + 2× 200G (ConnectX-7) — cluster-ready east-west fabric.
• Form factor: 8U rack, air-cooled · 12× 3000W Titanium PSU (6+6 redundant).
• Availability: made-to-order (vendor 5-wk + delivery) · lead time 8 weeks (delivered, cleared, installed in India).
• Make-in-India delivery — INR price, GST tax invoice, pan-India onsite support, GeM-procurable.

AI workload fit

• Frontier LLM training — 2,304 GB HBM3e (8× 288 GB) of coupled HBM holds large models + optimizer state for data/tensor-parallel training.
• High-throughput inference — serve many concurrent sessions / long-context requests per node.
• Fine-tuning & RAG — full-parameter and PEFT on 70B–400B-class models on a single node.
• HPC & scientific AI — mixed-precision simulation and AI-for-science.

AI workload positioning

An 8× B300 Blackwell-Ultra node carries ~2.3 TB of HBM3e and 800G XDR fabric — next-gen capacity for the largest training and reasoning workloads. A single 8-GPU node is the unit of scale for on-prem AI: cluster several over the 800G InfiniBand XDR fabric to build a pod, or run one as a self-contained training/inference engine. We do not publish fabricated tokens/sec — request a POC for your model and we will benchmark it.

Industry use cases

• Government & sovereign AI — train national/domain models on-prem, data never leaves the country.
• BFSI — private LLMs for risk, fraud and document AI under data-residency rules.
• Neocloud / AI providers — a priced, in-stock node to stand up or expand a GPU cloud fast.
• Healthcare & research — genomics, imaging and foundation-model research under compliance.
• Enterprise & manufacturing — in-house copilots, agentic automation, engineering AI.

Performance & how to be sure

Positioned on real, coupled HBM capacity and fabric bandwidth rather than headline FLOPS. Want certainty? Request a POC / benchmark on your own model and dataset before you commit — we will show real throughput on this exact configuration. No fabricated benchmark numbers.

Series & upgrade path

• DRACO — the flagship tier: 8-GPU servers, multi-node systems, rack-scale AI factories and superclusters.
• This node clusters over 800G InfiniBand XDR into multi-node pods and rack-scale systems without re-platforming.
• Pair with RDP storage (GPUDirect), fabric switches and liquid-ready racks as you scale.

On-prem vs cloud — the TCO case

For sustained training/inference, an owned 8-GPU node beats per-hour cloud within quarters and keeps models and data in-house. RDP pricing is fixed in INR with a GST input-credit-eligible invoice — ask for a 3-year TCO vs cloud for your utilisation.

Software & day-one readiness

Delivered validated for the CUDA/NVIDIA AI stack — drivers, CUDA, NCCL, containers, and orchestration (Slurm/Kubernetes); optional NVIDIA AI Enterprise and an RDP KV-cache / extended-memory tier. Ships racked, burned-in and networked. (Exact software stack confirmed at handover.)

Power, thermal & acoustics

8U rack with air-cooled · 12× 3000W Titanium PSU (6+6 redundant); datacenter-grade airflow and power redundancy. Plan for a high-density rack (three-phase power, adequate cooling). Site power/thermal survey included in deployment.

Deployment, warranty & support

• Deployment: delivered, cleared, racked, cabled and validated in India; optional onsite commissioning.
• Warranty: OEM 3-year, with RDP pan-India onsite SLA options.
• Support: single RDP point of contact; spares and engineers pan-India.
• Add-ons: deployment, 3-yr onsite, AMC and financing available as separate services.

Why RDP

14 years of Make-in-India infrastructure and 300,000+ devices shipped. Indian OEM & integrator — INR pricing, GST tax invoice, pan-India onsite engineers, GeM availability, and DPDP / sovereign-AI-ready delivery. We import, land, integrate and support so you don’t have to.

Buy with confidence

This is a specific, in-stock 8-GPU AI server at a transparent INR price — add to cart or talk to an RDP solution architect. Limited availability (made-to-order (vendor 5-wk + delivery)); confirm your slot early.

Specifications