What the NVIDIA DGX Spark Is
The DGX Spark is NVIDIA's small desktop AI computer, built around the GB10 Grace Blackwell Superchip: a 20-core Arm CPU and a Blackwell GPU sharing one memory pool. NVIDIA sells a Founders Edition through its marketplace and through retail partners. OEM versions of the same platform come from partners.
On 2026-10-02 NVIDIA announced a 64 GB configuration sold only through Acer, ASUS, Dell, Gigabyte, HP and MSI, available from 2026-10-23. It uses the same GB10 chip, DGX OS and software stack as the 128 GB model.
Specifications
Processor
- Detail:
- 20-core Arm: 10 Cortex-X925 + 10 Cortex-A725
GPU
- Detail:
- Blackwell, 5th-gen Tensor Cores, 4th-gen RT Cores
AI performance
- Detail:
- Up to 1 PFLOP FP4 (theoretical, with sparsity)
Memory
- Detail:
- 128 GB LPDDR5x unified; 64 GB only via OEM partners
Memory bus and bandwidth
- Detail:
- 256-bit, 273 GB/s
Storage
- Detail:
- Up to 4 TB NVMe M.2, self-encrypting
Networking
- Detail:
- ConnectX-7 200 Gbps (2x QSFP), 10 GbE RJ-45, Wi-Fi 7, Bluetooth 5.4
Ports and display
- Detail:
- 4x USB-C, 1x HDMI 2.1a, up to 3x DisplayPort over USB-C
Power
- Detail:
- 240 W power supply; GB10 TDP 140 W
Size and weight
- Detail:
- 150 x 150 x 50.5 mm, 1.2 kg
Operating system
- Detail:
- NVIDIA DGX OS (Ubuntu-based)
Noise (ECMA-109)
- Detail:
- 35 dB sound power at maximum GPU load, 19 dB idle
NVIDIA model-size claims
- Detail:
- Inference up to 200B (128 GB) or 100B (64 GB); fine-tuning up to 70B; 400B across two 128 GB boxes; 700B across four
All rows come from NVIDIA's DGX Spark product page. NVIDIA states that clusters of two or three systems can use a direct QSFP cable and that four need an external 200 GbE switch. The page does not state a CUDA core count.
Price and Availability
US prices have risen since launch because of memory supply, and they differ by seller, so check the live listing. All figures are US dollars from the stated source.
Launch price, 128 GB
- US price:
- $3,999
- Source and date:
- NVIDIA developer forum, 2026-02-25
- Status:
- Replaced by the price below
Founders Edition, 128 GB
- US price:
- $4,699
- Source and date:
- NVIDIA developer forum, 2026-02-25
- Status:
- MSRP set by NVIDIA; OEM systems not covered
128 GB / 4 TB
- US price:
- $6,950
- Source and date:
- ServeTheHome, 2026-10-03
- Status:
- Reported; OEM prices vary
64 GB (OEM only)
- US price:
- From $4,999
- Source and date:
- NVIDIA blog, 2026-10-02
- Status:
- Available from 2026-10-23 via six OEMs
NVIDIA's product page shows no price and sends buyers to the NVIDIA Marketplace and retail partners. StorageReview's review, published 2026-07-06, put the Founders Edition at $4,699 and noted that some OEM systems still sold near $4,000. Our sources conflict on what a 128 GB unit costs today, which is why each line is dated.
What Independent Reports Show
The head-to-head with published numbers that we found is StorageReview's Ryzen AI Halo review, published 2026-07-06, where the Spark led in all 18 vLLM scenarios. The test used online serving on Linux at batch 64 with six models, including GPT OSS 120B. The Spark was ahead by about 1.9x to 13.4x in 16 scenarios and by about 11-12% in two decode-heavy runs. The same review notes the Ryzen AI Halo has a single 10GbE port and no 200G fabric. HotHardware's review of the Halo instead described the two machines as roughly on par in many models, so the gap depends on the test setup.
ServeTheHome's launch review, published 2025-10-14, measured untuned Ollama at 14.48 tokens per second for gpt-oss 120B and often over 49 for gpt-oss 20B. It describes an early sample, reports display and idle-power-state problems, and tested Linux only. Readers in its comments questioned the 120B figure, so treat it as an out-of-the-box number, not a ceiling.
Tom's Hardware published a DGX Spark review whose headline says the Spark beats the Ryzen AI Max+ 395; we could read only the headline, so we cite no figures from it. Published results also move with runtime and software version.
Who It Suits and Who Should Skip It
- Choose it if your workflow depends on CUDA or NVIDIA's AI software, and you can work in Linux.
- Choose it if you plan to link two boxes over the 200 Gbps ports for larger models.
- Choose it for fine-tuning experiments; NVIDIA states support up to 70B parameters.
- Skip it if you need Windows or more than 128 GB; several alternatives below list Windows, and three list 192 GB models.
- Skip it if price per gigabyte of memory decides the purchase; the sourced price table shows how far it has moved.
Alternatives in the Same Class
- Ryzen AI Halo: AMD's 128 GB developer platform with Windows 11 Pro or Linux, but a single 10GbE port and slower in StorageReview's serving test.
- Framework Desktop: offers a 192 GB configuration on pre-order in a 4.5 L chassis, per Framework and Phoronix; batch status is on its page.
- Minisforum MS-S1 MAX: 128 GB with dual 10GbE and a PCIe x16 slot (x4 electrically, per reviews), per the Minisforum store.
- Local AI Workstations: Strix Halo vs. DGX Spark: the category comparison of all machines in this class.
Frequently Asked Questions
Can the DGX Spark run Windows?
NVIDIA lists DGX OS, which it describes as Ubuntu-based, as the operating system. StorageReview describes the Spark as running only Linux-based DGX OS.
Can two DGX Sparks be combined?
Yes. NVIDIA states clusters of two or three systems can connect with a direct QSFP cable, and a four-system cluster needs an external 200 GbE switch.
Who sells OEM versions of the DGX Spark?
NVIDIA names Acer, ASUS, Dell, Gigabyte, HP and MSI as partners for the 64 GB model. NVIDIA's 2025 launch announcement also listed Lenovo; we did not verify a Lenovo 64 GB model.
How loud is it?
NVIDIA states a mean A-weighted sound power of 35 dB under maximum GPU stress and 19 dB at idle, measured to ECMA-109.
Does the 64 GB model run the same software?
Yes, per NVIDIA. It uses the same GB10 chip, DGX OS and AI software stack, with a stated single-unit limit of 100B parameters for inference.