NVIDIA DGX Spark Personal AI Supercomputer

Overview

A petaflop of AI on your desk — no data centre, no cloud bill, no queue. The NVIDIA DGX Spark puts the Grace Blackwell architecture in a 150 mm box you can carry in one hand, with 128 GB of unified memory that both the CPU and GPU address at once. It runs models locally that normally demand a server — and it is the most accessible way into serious AI development that RDP sells.

Key highlights

• NVIDIA GB10 Grace Blackwell Superchip — up to 1 petaFLOP of AI performance at FP4.
• 128 GB unified LPDDR5x — CPU and GPU share the full pool over NVLink-C2C at ~5× PCIe Gen5 bandwidth. No copying tensors across a bus.
• Runs up to ~200B-parameter models locally for inference; fine-tune up to ~70B — work that normally needs rented cluster time.
• Blackwell GPU with 5th-gen Tensor Cores + RT Cores, paired with a 20-core Arm Grace CPU.
• 4 TB NVMe onboard for datasets, checkpoints and model weights.
• ConnectX-7 200GbE — link two Sparks to reach roughly 405B-parameter class workloads.
• 150 × 150 mm, ~1.2 kg — desk-side, plugs into a normal wall socket, no rack or facility work.
• Ships with NVIDIA DGX OS and the full NVIDIA AI software stack — CUDA, containers and frameworks ready on power-up.

AI workload fit

• Local inference and model serving up to ~200B params.
• Fine-tuning / LoRA / SFT on models up to ~70B.
• RAG and retrieval pipelines against private data that never leaves the desk.
• Agentic AI development — build and iterate on agent fleets locally.
• Generative AI, computer vision and NLP & speech prototyping.
• Not for large-scale pretraining — that is cluster work (see the upgrade path below).

AI workload positioning

DGX Spark is where an AI programme starts. The unified 128 GB pool is the differentiator: memory capacity, not raw FLOPS, is what usually stops a developer running a large model locally, and Spark removes that wall at desk scale. Develop, fine-tune and validate here on the same CUDA stack you will later run in production — then scale the identical software to an RDP GPU server or rack-scale system without a rewrite.

Industry use cases

• Enterprises & GCCs: give each AI engineer a private development box instead of contended cloud credits.
• Research & education: per-lab or per-researcher AI compute within a departmental budget.
• BFSI: prototype on sensitive data that must never leave the premises.
• Healthcare: local work on patient data under data-residency obligations.
• Media & entertainment: desk-side generative image, video and audio iteration.

Performance & how to be sure

The headline figure — 1 petaFLOP FP4 — is NVIDIA’s published peak for the GB10 Superchip, and peak is not the same as sustained throughput on your model. Rather than quote a tokens/sec number that will not match your workload, RDP offers a “benchmark your model” session: bring the model and precision you actually intend to run and we will validate it on this hardware before you commit.

Series & upgrade path

Spark is the CARINA (entry) tier of the GPU Mart ladder and the natural first rung. When a single desk-side unit is no longer enough: link a second Spark over ConnectX-7 (~405B class) → step up to a QUASAR multi-GPU workstation → an RDP GPU server (8× H200 / B300) → rack-scale GB300 NVL72. The software stack is identical at every rung, so nothing is thrown away when you scale.

On-prem vs cloud (TCO)

At a fixed capital cost, Spark removes per-hour GPU billing, queueing for scarce instances, and egress charges on your own data — and keeps everything on-premises for DPDP / data-residency purposes. Cloud still wins for burst capacity and very large training runs; Spark wins for the daily, sustained development cycle that would otherwise meter continuously. For a team of engineers billing cloud GPU hours every working day, the payback arithmetic is usually short — we will model it with your actual usage rather than assert it.

Software & day-one readiness

Ships with NVIDIA DGX OS and the NVIDIA AI stack — CUDA, cuDNN, containers, and the common serving and fine-tuning frameworks (vLLM, PyTorch and the wider NVIDIA NGC catalogue). It is the same CUDA environment as every larger NVIDIA system in this catalogue, which is precisely why work moves upward without a rewrite.

Power, thermal & acoustics

Runs from a standard wall socket with an external power adapter — no rack, no dedicated circuit, no facility cooling. Air-cooled and designed for an office desk rather than a machine room. {Exact wattage, BTU and dB(A) figures: to be published from RDP bench measurement rather than estimated.}

Deployment, warranty & support

In stock and ready to ship. Supplied by RDP Technologies with GST invoice (HSN 8471), pan-India onsite support, and availability through GeM for public-sector procurement. Unbox, plug in, and you are developing the same day.

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

Priced at ₹4,80,000 + applicable taxes. Add to cart, or talk to an RDP solution architect first if you want your model benchmarked on the unit before purchase.

Specifications