NVIDIA DGX Spark
by NVIDIA
128GB of unified Grace Blackwell memory on a desk - built for model capacity, not token throughput
DGX Spark is NVIDIA's desktop AI development system built on the GB10 Grace Blackwell Superchip, pairing a 20-core Arm CPU with a Blackwell GPU and 128GB of coherent unified memory. It lets engineers prototype and fine-tune models far larger than a consumer GPU can hold, on the same CUDA and NIM software stack that runs in NVIDIA's data-centre systems.
NVIDIA DGX Spark is a desktop AI development system built around the GB10 Grace Blackwell Superchip, announced at GTC in March 2025 and on sale from 14 October 2025 at $3,999, with NVIDIA raising the MSRP to $4,699 in February 2026 on LPDDR5x supply constraints. The GB10 pairs a 20-core Arm CPU (10 Cortex-X925 and 10 Cortex-A725) with a Blackwell GPU carrying fifth-generation Tensor Cores rated at up to 1 petaFLOP of FP4 compute, sharing 128GB of coherent unified memory across CPU and GPU at 273 GB/s. Storage is a 4TB self-encrypting NVMe M.2; networking is a ConnectX-7 NIC at 200Gbps plus 10GbE, and two units can be linked to hold models up to 405 billion parameters. A single box handles inference on models up to 200B and fine-tuning up to 70B. It ships with DGX OS - an Ubuntu derivative - plus the NVIDIA AI software stack, NIM microservices, NeMo and CUDA libraries, so a workload prototyped here moves to a DGX data-centre system unchanged. The machine is 150mm square, 1.2kg, and runs from a 240W supply on a normal wall outlet. That unified-memory capacity is the entire point: LMSYS measured 803 tokens/sec prefill on Llama 3.1 70B in FP8, a model that will not fit on any single consumer GPU. Decode is the trade-off - 273 GB/s of bandwidth caps generation at roughly 20-50 tokens/sec on mid-size models, several times slower than an RTX 5090 or RTX PRO 6000. OEM variants ship from Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI.
ML engineers and research teams who need to load and fine-tune 70B-200B parameter models locally, on the same CUDA stack their production DGX or cloud cluster runs.
128GB of coherent unified memory on a desk, so models that will not fit any single consumer GPU can be prototyped without renting cluster time.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Flat rate
- Target Market
- ML Engineers, AI Researchers, CTOs, Data Scientists, Academic AI Labs
- Deployment
- Edge-first, Self-hosted
- Founded
- 1993
- Headquarters
- Santa Clara, United States
- Team Size
- 500+
Key Features
- ✓128GB coherent unified memory
CPU and GPU share one 128GB address space, so a 70B-class model loads without the partitioning a multi-GPU rig demands.
- ✓GB10 Grace Blackwell Superchip
20-core Arm CPU plus Blackwell GPU with fifth-generation Tensor Cores delivering up to 1 petaFLOP of FP4 compute.
- ✓Two-unit clustering over ConnectX-7
A 200Gbps NIC lets two Sparks link to hold models up to 405 billion parameters, doubling capacity without a rack.
- ✓Full NVIDIA AI stack preinstalled
DGX OS ships with CUDA, NIM microservices, NeMo and pretrained models, so a prototype moves to data-centre DGX unchanged.
- ✓Desk-scale power and footprint
150mm square, 1.2kg and 240W from a standard wall outlet, so it needs no special power or cooling.
- ✓4TB self-encrypting NVMe storage
Enough local capacity for multiple large checkpoints and datasets, with encryption at rest handled by the drive.
Capabilities
Use Cases
- •Local 70B fine-tuning
Fine-tune models up to 70 billion parameters on-desk, avoiding both cloud GPU rental and any transfer of proprietary training data.
- •Large-model prototyping before cluster spend
Validate an approach on a 200B model locally, then move the identical CUDA and NIM workload to production DGX hardware.
- •Batched small-model inference
LMSYS measured 368 tokens/sec decode on Llama 3.1 8B at batch 32, so throughput workloads with batching perform well.
- •Confidential or air-gapped experimentation
All prompts, weights and datasets stay on a physically controlled device, satisfying policies that block hosted inference entirely.
- •Memory-coherence research
The shared CPU-GPU address space makes it a testbed for architectures that a discrete-GPU system cannot represent.
Ideal For
Best For
- ✓Local prototyping and fine-tuning of 70B-class models where cloud GPU rental costs or data policy rule out a hosted cluster
- ✓Teams already standardised on CUDA, NIM and NeMo who need dev-to-production parity with data-centre DGX systems
- ✓Air-gapped or data-sensitive research where model inputs cannot leave the building
- ✓Memory-coherent architecture research that depends on CPU and GPU sharing one address space
- ✓University and corporate AI labs needing per-researcher capacity without allocating shared cluster time
Not Ideal For
- ✗Latency-sensitive or high-throughput production inference - 273 GB/s of memory bandwidth caps decode at roughly 20-50 tokens/sec, and LMSYS measured an RTX PRO 6000 at 215 tokens/sec versus the Spark's 49.7 on the same 20B model, four to five times faster
- ✗Buyers optimising raw performance per dollar; independent reviewers put it at roughly $37 per GB of memory after the February 2026 price rise and concluded that for the money it is simply slow on standard inference
- ✗General-purpose desktop use, gaming, or content and video workloads - this is an Arm64 Linux development appliance, not a workstation replacement
- ✗Teams without Linux and CUDA depth: early adopters hit missing Arm64 PyTorch wheels, x86 assumptions across common tooling, and CUDA 12-versus-13 confusion that took real debugging time
Integrations
Deployment
Market Analysis
Pros
- ✓128GB of unified memory lets a single desktop hold models that need a multi-GPU server otherwise - LMSYS ran Llama 3.1 70B FP8 at 803 tokens/sec prefill
- ✓Software parity with data-centre DGX: the same CUDA, NIM, NeMo and DGX OS stack, so nothing is rewritten on the way to production
- ✓Prefill and batched throughput are genuinely strong - 7,991 tokens/sec prefill and 368 tokens/sec decode at batch 32 on Llama 3.1 8B
- ✓Runs at 240W from a standard wall outlet in a 150mm box, needing no special power, cooling or rack space
Cons
- ✗Memory bandwidth of 273 GB/s is the hard bottleneck - independent benchmarks put decode at 38.6 tokens/sec on GPT-OSS 120B and 2.7 tokens/sec on Llama 3.1 70B, and one reviewer's blunt verdict is that for the money it is slow
- ✗Four to five times slower than discrete alternatives on decode: LMSYS measured RTX PRO 6000 at 215 and RTX 5090 at 205 tokens/sec against the Spark's 49.7 on the same 20B model
- ✗The price rose 18% from $3,999 to $4,699 in February 2026 on memory supply constraints, working out to roughly $37 per GB and weakening the value case
- ✗Arm64 software friction was real at launch - Simon Willison's hands-on found missing PyTorch CUDA wheels for Arm, tooling that assumed x86, and thin documentation until days before embargo lift
- ✗A three-way RTX 3090 rig measured 124 tokens/sec decode versus the Spark's 38.6 for around half the price, so anyone optimising throughput per dollar has better options
Pricing
DGX Spark
From $3,999
- ✓GB10 Grace Blackwell Superchip
- ✓128GB unified memory at 273 GB/s
- ✓4TB self-encrypting NVMe
- ✓DGX OS with the NVIDIA AI stack
- ✓MSRP raised to $4,699 in February 2026
OEM variants
Contact for pricing
- ✓Same GB10 superchip from Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI
- ✓Vendor-specific chassis, storage and support terms
This is a one-time hardware purchase, not a subscription: it launched at $3,999 in October 2025 and NVIDIA raised the MSRP to $4,699 in February 2026 - an 18% increase attributed to LPDDR5x memory supply constraints - so street pricing now varies widely by channel, with Newegg around $4,399 and some retailers well above list. Independent reviewers work out roughly $37 per GB of unified memory at the new price and note that a used multi-GPU rig beats it on raw throughput per dollar; you are paying for memory capacity, coherence and CUDA parity, not tokens per second. Every unit includes a 90-day NVIDIA AI Enterprise licence, after which continued access to that software stack is a separate commercial item.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 4 on domains other than nvidia.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
WEKA NeuralMesh
Microservices storage and memory fabric built for AI training and inference at exabyte scale
Tsuga
Bring-your-own-cloud observability that keeps telemetry, and its cost, inside your own AWS account
OpenObserve
Open-source, Rust-based observability on object storage — logs, metrics, traces and LLM telemetry in one binary
Coralogix
AI-native observability that queries logs, metrics and traces in place — index-free, in your own S3 bucket