NVIDIA has squeezed a data-center-class AI system into a box the size of a paperback. The NVIDIA DGX Spark, formerly known as Project DIGITS, puts a Grace Blackwell superchip and 128GB of coherent unified memory on your desk for $4,699, promising to run models up to 200 billion parameters without the cloud. Shipping since October 2025 and selling out through early 2026 after an 18 percent price hike, it raises an obvious question for developers, researchers, and creators: is a palm-sized DGX box worth it when an RTX 5090 or a Mac Studio costs less and benchmarks faster?
This guide answers every top search question about the DGX Spark, from what it is for and how NVFP4 memory math works to real tokens-per-second benchmarks, model capacity limits, and how it stacks up against the RTX 5090, Mac Studio M3 Ultra, and AMD Strix Halo. You will get verified specs, independent LMSYS results, a practical model-fit table, and buying advice grounded in 2026 street pricing.
What Is the NVIDIA DGX Spark For?
The NVIDIA DGX Spark is billed as the world’s smallest AI supercomputer, a personal AI development platform built around the GB10 Grace Blackwell Superchip. NVIDIA positions it for developers, researchers, data scientists, robotics teams, and students who need to prototype, fine-tune, and inference large models locally before scaling to DGX Cloud or larger clusters.
Unlike a traditional gaming GPU that you drop into a PC, the DGX Spark is a complete system. It ships as a 150 x 150 x 50.5 mm, 1.2 kg desktop unit running DGX OS (Ubuntu-based Linux with NVIDIA’s AI stack preinstalled), including PyTorch, RAPIDS, NGC containers, and support for every major open-source LLM framework via full CUDA 13. It is designed for local-first workflows where privacy, latency, or repeated cloud spin-up costs matter.
Common use cases NVIDIA highlights for the DGX Spark include:
- Prototyping and fine-tuning large language models up to 70B parameters on-device
- Running inference on 120B to 200B models with NVFP4 quantization without model sharding
- Building AI agents, retrieval-augmented generation pipelines, and computer vision or robotics prototypes
- Validating workloads locally before migrating to DGX Cloud or data center Blackwell systems
The Spark is not a gaming PC, though it can render graphics via its Blackwell GPU. Learn more from the official NVIDIA DGX Spark product page. Its networking (ConnectX-7 with up to 200 Gbps) even lets you link two Sparks into a single 256GB pool for 405B-parameter models, a trick demonstrated by EXO Labs with a hybrid Spark plus Apple M3 Ultra cluster.
Inside the GB10 Grace Blackwell Superchip: Key Specs
At the heart of every DGX Spark is the GB10, co-developed with MediaTek on TSMC’s 3nm-class process. It fuses a 20-core Arm Grace CPU and a Blackwell GPU in one package, connected by NVLink-C2C at roughly five times the bandwidth of PCIe Gen 5.
| Specification | NVIDIA DGX Spark (GB10) |
|---|---|
| CPU | 20-core Arm: 10x Cortex-X925 at 4 GHz + 10x Cortex-A725 at 2.8 GHz |
| GPU | Blackwell architecture, 6,144 CUDA cores, 5th-gen Tensor Cores, 4th-gen RT Cores |
| AI Compute | Up to 1 petaFLOP FP4 with sparsity, native NVFP4 support |
| Memory | 128GB LPDDR5X unified, coherent CPU+GPU pool, 273 GB/s bandwidth, 256-bit bus |
| Storage | 1TB or 4TB self-encrypting NVMe M.2 SSD |
| Networking | ConnectX-7 Smart NIC, 10GbE, Wi-Fi 7, Bluetooth 5.4, 2x QSFP 200 Gbps |
| Power | 240W external adapter, GPU TDP about 140W, about 195W measured at wall |
| OS and Software | DGX OS 7.4.0 (Ubuntu 24.04, kernel 6.17), CUDA 13.0.2, NVIDIA AI Enterprise (90-day trial) |
| Form Factor | 150 x 150 x 50.5 mm, 1.2 kg |
| Local Model Capacity | Up to 200B parameters (single unit), up to 405B with two clustered units |
The single most important number is 128GB unified. Unlike a discrete GPU, where system RAM and VRAM are separate pools that must copy data over PCIe, the GB10 shares one coherent address space. Both CPU and GPU see the same memory, which is why the Spark can load models that a 24GB or 32GB card cannot, even if its 273 GB/s bandwidth is far lower than the RTX 5090’s 1,792 GB/s.
Another Blackwell advance is NVFP4, a native 4-bit floating point format. It shrinks models by about 3.5x versus BF16 in practice and is executed natively on fifth-gen Tensor Cores, enabling aggressive quantization without the accuracy cliff older 4-bit schemes suffered. GPT-OSS models are even trained with NVFP4 awareness, which is why they run particularly well on Spark.
Unified Memory Explained: Why 128GB Changes Which Models You Can Run
If you are choosing hardware for local LLMs, think in two dimensions: capacity and bandwidth. Capacity determines which models load at all, and bandwidth determines how fast they generate tokens once loaded.
On discrete GPUs, each card has its own VRAM. Two RTX 5090s give you 64GB total, but not a single 64GB pool. Models that do not fit on one card require tensor parallelism and fast interconnects, adding complexity and overhead. On unified systems like the DGX Spark or Apple Silicon, the entire pool is addressable by all compute units.
At 4-bit quantization (Q4 or NVFP4), the rule of thumb is about 0.5 GB per billion parameters. That makes capacity planning simple:
- 70B model at 4-bit: about 35 to 40 GB, leaving room for long context windows and fine-tuning overhead
- 120B model (GPT-OSS 120B class): about 60 to 70 GB, comfortable on a single Spark
- 200B model: about 100 to 110 GB, the practical ceiling for a single 128GB unit once you reserve memory for the OS, context, and KV cache
- 405B model (Llama 3.1 405B): requires two Sparks linked via ConnectX-7 for a 256GB shared pool in FP4
Fine-tuning is hungrier than inference because it stores optimizer states and gradients. In practice, the Spark comfortably fine-tunes up to about 70B-parameter models, which is still far beyond any 32GB consumer card. This is the Spark’s originality edge: it trades raw speed for capacity that simply does not exist in that form factor elsewhere.
A practical note from hands-on users: stop watching nvidia-smi for memory on Spark. There is no separate GPU memory to query. Check free -h instead. The full pool reports as about 122,570 MB usable after firmware reservations, roughly 119 GiB for workloads.
Real-World Performance: Benchmarks vs. RTX 5090 and Other Local Rigs
Marketing says 1 petaFLOP. Independent benchmarks tell a more nuanced story: the DGX Spark is excellent for small to mid-sized models at low concurrency and becomes bandwidth-bound on larger models or heavy batching.
LMSYS Org, the most cited independent source in 2026, ran controlled tests on the same models across Spark and workstation GPUs. Highlights:
| Model and Setup | DGX Spark (GB10) | RTX 5090 (32GB) | RTX Pro 6000 Blackwell Workstation |
|---|---|---|---|
| GPT-OSS 20B MXFP4 (Ollama, batch 1) | 2,053 tokens/sec prefill / 49.7 decode | 8,519 prefill / 205 decode | 10,108 prefill / 215 decode |
| Llama 3.1 8B FP8 (SGLang, batch 1) | 7,991 prefill / 20.5 decode | Not disclosed in same run | Faster by about 3 to 4 times |
| Llama 3.1 8B FP8 (SGLang, batch 32) | 7,949 prefill / 368 decode | Higher per-stream speed | Higher per-stream speed |
| GPT-OSS 120B (NVFP4, SGLang) | 35 to 80+ tokens/sec decode (quantization dependent) | Cannot load (exceeds 32GB) | Cannot load without multi-GPU |
The pattern is consistent across IntuitionLabs, GPUSmith, and Tom’s Hardware reviews: the RTX 5090’s 1,792 GB/s GDDR7 bandwidth delivers roughly 2.7 to 4 times faster single-stream decode on models that fit in 32 GB. About 38.6 tokens/sec on Spark versus 124 tokens/sec on a 3x RTX 3090 rig for the same workload is a typical gap reported. Where the Spark wins is not speed but eligibility. A 70B or 120B model that does not fit on the 5090 at all runs fine on Spark, albeit slower.
CES 2026 brought a meaningful software lift. NVIDIA’s January 2026 update delivered up to 2.5x inference improvements through TensorRT-LLM optimizations and speculative decoding and up to 8x speedups for video generation workloads. Reviewers on Medium and Hacker News noted the update changed the value proposition after early thermal and bandwidth complaints.
How It Compares: DGX Spark vs RTX 5090 vs Mac Studio vs Strix Halo
Buy the RTX 5090 if your models fit in 32GB and you want maximum tokens per second. Buy the DGX Spark or an Apple Silicon Mac when capacity beats per-stream speed, or when unified memory, low power, and zero PCIe copying matter more than raw bandwidth.
| System | Memory for Models | Memory Bandwidth | Peak AI Compute | Power at Wall | Street Price (Mid-2026) |
|---|---|---|---|---|---|
| NVIDIA DGX Spark (GB10) | 128GB unified (256GB with 2 units) | 273 GB/s | 1 PFLOP FP4 sparse | ~195W | $4,699 Founders, ~4,679 on Amazon |
| RTX 5090 PC Build | 32GB GDDR7 (discrete) | 1,792 GB/s | Between 5070 and 5070 Ti class for AI | 575W GPU TGP alone | $1,999 MSRP, $3,658 to $4,329 street |
| Mac Studio M3 Ultra | Up to 192GB unified | ~400 to 546 GB/s (model dependent) | 38 TOPS FP16 (not directly comparable) | ~100 to 150W | $4,499 to $6,000+ configured |
| AMD Strix Halo (128GB) | 128GB unified | Similar LPDDR5 class | Comparable to Spark in FP8/FP16 per tests | Lower than discrete GPU rig | ~2,348 |
The honest 2026 trade-off: the RTX 5090 answers “how fast,” the Spark answers “how big.” AMD’s Strix Halo is the price disruptor. IntuitionLabs found a $2,348 Strix Halo system achieving comparable inference performance to the $4,699 Spark under FP8 or FP16 in early tests, though the CUDA software moat remains NVIDIA’s advantage. The Mac Studio wins on silence, efficiency, and daily driver versatility, and an equivalently spec’d M4 Max Studio with 128GB and 4TB SSD actually costs about $700 more than Spark in one widely cited comparison ($4,699 vs $3,999 at launch). For fine-tuning, the Spark’s full CUDA and NVFP4 stack gives it an edge over Apple. For single-stream gaming or max prefill speed on small models, the discrete GPU tower still leads.
Two Sparks linked via ConnectX-7 remain the only palm-sized path to 405B local inference without a server rack, a niche but real need for teams prototyping agentic loops that must stay air-gapped or private.
Price, Release Date, and Availability in 2026
The DGX Spark was announced as Project DIGITS at CES in January 2025, shown at GTC in March 2025, and launched as a Founders Edition on October 15, 2025, at $3,999 with a 4TB SSD. NVIDIA raised the price to $4,699 on February 23, 2026, an 18 percent increase the company attributed to worldwide memory supply constraints.
As of August 2026:
- Official NVIDIA Marketplace: $4,699 for the Founders Edition
- Amazon: about $4,679, with intermittent stock
- Partner systems (ASUS Ascent GX10, PNY, Dell, HP, Lenovo, Acer, Gigabyte, MSI): around $4,499 for similar GB10 configurations
- Operating system: DGX OS 7.4.0, Ubuntu 24.04-based, kernel 6.17, with CUDA 13.0.2 and ConnectX-7 networking
At Computex 2026, NVIDIA also unveiled the RTX Spark line, bringing the same GB10 silicon to Windows mini PCs and laptops from major OEMs, alongside the Linux-focused DGX Spark. Stock has been tight since launch, with most retailers showing sell-outs and restocks rather than sustained availability.
If you are buying now, factor in that memory pricing volatility is driving the current street price. The launch price of $3,999 is unlikely to return soon.
Is the NVIDIA DGX Spark Worth It? Pros, Cons, and Verdict
The DGX Spark is not for everyone. It is a developer workstation first, a general PC second.
Choose the DGX Spark if you:
- Need to load 70B to 200B models locally without cloud costs or data leaving your desk
- Fine-tune 7B to 70B models with full CUDA compatibility and NVFP4 efficiency
- Value silent-ish, compact, low-power (195W) operation over peak tokens per second
- Want a validated, preconfigured Linux AI stack that works with SGLang, Ollama, TensorRT-LLM, and PyTorch out of the box
- Plan to cluster two units for 405B experiments or hybrid clusters with Macs
Skip it or choose an alternative if you:
- Live mostly in 7B to 32B models, where a single RTX 5090 is 3 to 4 times faster and cheaper per token
- Need Windows-native gaming or maximum prefill speed for high-concurrency serving
- Want the quietest, most efficient daily driver with good-enough local AI, where a Mac Studio or Mac mini M4 Pro fits better
- Are price sensitive and comfortable with AMD Strix Halo’s emerging ROCm stack at roughly half the price
Bottom line: For its intended audience of AI developers who prototype daily and pay for cloud GPUs by the hour, the DGX Spark pays for itself in convenience and privacy once the software update gains are factored in. For gamers or anyone whose models fit comfortably in 32GB, the RTX 5090 remains the faster, cheaper tokens-per-second buy. The DGX Spark answers a different question, and on that question, capacity per watt per kilogram, it has no equal in 2026.
Frequently Asked Questions
What is the NVIDIA DGX Spark for?
It is a personal AI supercomputer for local development. Developers use it to prototype, fine-tune, and run inference on large models entirely on-device, keeping data private and avoiding cloud wait times, then scale validated work to DGX Cloud or data-center Blackwell systems.
How much does the NVIDIA DGX Spark cost?
The Founders Edition launched at $3,999 on October 15, 2025, and rose to $4,699 on February 23, 2026, due to memory supply constraints. In mid-2026 it lists at $4,699 on NVIDIA’s marketplace and about $4,679 on Amazon, with partner builds like the ASUS Ascent GX10 around $4,499.
Is the DGX Spark better than an RTX 5090?
It depends on the metric. For raw speed on models that fit in 32 GB, the RTX 5090 is faster by about 3 to 4 times thanks to 1,792 GB/s bandwidth versus 273 GB/s. For capacity, the DGX Spark is better; it can load 70B to 200B models that do not fit on a 5090 at all, and two Sparks can handle 405B. Choose speed or capacity based on your models.
What are the DGX Spark specs?
GB10 Grace Blackwell Superchip (20-core Arm CPU with 10 Cortex-X925 and 10 Cortex-A725, Blackwell GPU with 6,144 CUDA cores and 5th-gen Tensor Cores), 128GB unified LPDDR5x at 273 GB/s (see the NVIDIA DGX Spark datasheet), 1 petaFLOP FP4 sparse, 1 to 4TB NVMe, ConnectX-7 networking with 200 Gbps QSFP, DGX OS (Ubuntu 24.04), CUDA 13, Wi-Fi 7, 1.2 kg chassis.
What operating system does the DGX Spark run?
DGX OS 7.4.0, which is NVIDIA’s Ubuntu 24.04-based Linux distribution with kernel 6.17, preinstalled NVIDIA drivers, CUDA 13.0.2, and the full AI software stack. Partner RTX Spark variants run Windows, but the DGX Spark itself is Linux-native.
Can the DGX Spark run a 70B model?
Yes, easily. A 70B model at 4-bit needs about 35 to 40 GB. It fits comfortably with room for context and even fine-tuning. The Spark’s practical limits are about 200B for inference on one unit and 70B for fine-tuning, or 405B inference when two units are linked.
What is the difference between DGX Spark and RTX Spark?
Both use the GB10 Grace Blackwell silicon and 128GB unified memory. DGX Spark is the Linux developer workstation sold by NVIDIA, and the RTX Spark is the Windows ecosystem brand for GB10 mini PCs and laptops from OEMs like ASUS, Dell, HP, Lenovo, Acer, Gigabyte, and MSI announced at Computex 2026.
Can the DGX Spark be used for gaming?
Technically yes, it has a Blackwell GPU with RT Cores, but it is not optimized or priced for gaming. A similarly priced RTX 5090 gaming rig will deliver far higher frame rates and bandwidth. The Spark’s value is AI capacity and CUDA software, not gaming performance.
Conclusion
The NVIDIA DGX Spark turns the local AI hardware ladder on its head. Where last year’s hardware ladder for local AI was about how many discrete GPUs you could afford to bolt together, the GB10’s 128GB coherent pool and native NVFP4 ask a simpler question: do you want the fastest tokens on small models, or the ability to load the model you actually want to use? Independent 2026 benchmarks make the trade clear. An RTX 5090 wins per-stream speed, a DGX Spark wins capacity, and a second Spark removes the 200B ceiling entirely for 405B local work.
If you fine-tune or serve models above 32GB, value privacy and low power on your desk, and rely on CUDA, the Spark is the most compact path to frontier-scale local AI in 2026. If your world is 7B to 32B models and maximum throughput, a discrete GPU remains the better buy. Either way, the era of waiting 40 minutes for a cloud instance to prototype an idea is ending, and the DGX Spark is why.
