NVIDIA DGX Spark (Project Digits)

What it does well

The DGX Spark (formerly Project DIGITS) is NVIDIA's desk-side AI box built on the GB10 Grace Blackwell superchip: a 20-core Arm CPU (10 Cortex-X925 + 10 Cortex-A725) and a Blackwell GPU sharing 128 GB of coherent LPDDR5X, in a 150 × 150 × 50.5 mm case that runs from an external 240 W power supply. The headline is capacity, not speed. CPU and GPU are joined by NVLink-C2C inside the chip, but memory itself runs at 273 GB/s (256-bit LPDDR5X, per NVIDIA's spec sheet). 128 GB is 4× the RTX 5090's 32 GB in one box, and when one box isn't enough you can cluster up to four over ConnectX-7. NVIDIA's product page rates one 128 GB unit for inference on models up to 200B parameters and fine-tuning up to 70B (its user guide says training and fine-tuning up to 200B). We haven't tested either. By our arithmetic the 70B figure can't mean full fine-tuning in BF16: a 70B model's BF16 weights alone are 141 GB, before gradients and optimizer state, so read it as parameter-efficient tuning such as LoRA on a quantized base. The software is first-party. DGX OS is Ubuntu-based, and NVIDIA says it matches the distribution on its other DGX platforms (the Spark has its own recovery process and image). It ships with the CUDA Toolkit preinstalled (13.0.3 in NVIDIA's current release notes), plus cuDNN, TensorRT, Docker with the NVIDIA Container Runtime, NGC access and a JupyterLab built into the DGX Dashboard. NVIDIA lists NeMo, RAPIDS, NIM and PyTorch as preinstalled or available from NGC. The pitch is that you prototype here and move the work to DGX Cloud or other NVIDIA data-center infrastructure with little code change. The Founders Edition MSRP was $3,999 until NVIDIA raised it to $4,699 in February 2026, citing memory supply constraints. On 2026-11-01 NVIDIA's own US Marketplace listed it (128 GB, 4 TB) at $6,950. We found no NVIDIA announcement of that second increase. For that you get a complete system (CPU, 128 GB, up to 4 TB of NVMe, cooling and 200 Gb/s networking), not just a card. A 64 GB configuration also exists, but only through OEM partners; this verdict is about the 128 GB unit.

Where it breaks

• Bandwidth ceiling vs discrete GPUs. 273 GB/s is what limits decode speed on memory-bound workloads, and it's far below an H100 SXM's 3.35 TB/s (about 12× more, our math) or an RTX 5090's 1,792 GB/s (about 6.6×). NVIDIA's NVLink-C2C figure (5× the bandwidth of PCIe Gen 5) describes the CPU-to-GPU link inside the chip, not how fast either side reads memory. Big models will fit long before they run fast. We haven't tested a DGX Spark, so we're not quoting tok/s.
• Arm ecosystem still has friction. NVIDIA's own stack is built for it, but the CPU is Arm (aarch64), so anything that only ships x86 builds (some Python wheels, some Docker images, niche closed-source binaries) won't run natively. Expect occasional pip install failures and images with no arm64 build.
• Small power budget. Per NVIDIA, the whole box runs from a 240 W external supply. The GB10 chip (CPU and GPU together) has a 140 W TDP, and 100 W is left for networking, SSD, USB-C and the rest. For scale, NVIDIA rates the RTX 5090 at 575 W and the H100 SXM at up to 700 W, for the GPU alone. That's what lets it sit on a desk, and it's part of why this is not a throughput machine. NVIDIA also says the included 240 W supply is required for optimal performance: a different or lower-rated one may reduce performance, fail to boot or cause unexpected shutdowns.
• Scaling means more boxes, not a bigger box. Memory is fixed per unit, and NVIDIA's spec table and hardware guide list no expansion slots: four USB-C (one used for power), one HDMI, 10 GbE and two QSFP ports for the ConnectX-7 NIC. Units don't link over NVLink. They cluster over ConnectX-7 Ethernet: two or three with direct QSFP cables, four through a 200 GbE switch, and NVIDIA recommends power-of-2 counts for the best load balance. NVIDIA's product page rates two linked 128 GB units (256 GB) for models up to 400B parameters and four (512 GB) for up to 700B. Each of the two QSFP ports is rated up to 200 Gb/s, which is 25 GB/s (our math) and under a tenth of local memory bandwidth. NVIDIA says clustering also speeds up inference. We haven't tested that, and the link between boxes is far slower than each box's own memory, so treat a cluster mainly as a way to fit a bigger model. And every extra unit is another $6,950 at NVIDIA's US Marketplace (2026-11-01), so two units cost $13,900 and four cost $27,800 (our math).
• First-generation platform. It has been on sale since October 2025, and NVIDIA is still changing a lot underneath it. Its release notes list eight releases between November 2025 and October 2026, with new kernels, drivers and CUDA versions, plus changes to out-of-memory handling and the display memory carveout on GB10's unified memory. That's good for support, but behaviour and performance are still moving targets, and NVIDIA notes that GB10 partner systems may get updates later than its own Founders Edition.
• Not for production serving. NVIDIA now pitches it for always-on local agents, which is fine for one developer or a small team. It's still a dev box: a single 240 W unit with 273 GB/s of memory bandwidth is not a production inference server.

Ideal model range

• Sweet spot: Models up to NVIDIA's ~200B single-unit ceiling, with 200B itself a tight fit. Size math in this section is our arithmetic and counts weights only: parameters × bits per weight, at either a bare 4 bits or the 4.5 bits of NVIDIA's NVFP4 format (4-bit values plus one FP8 scale per 16), before context. The display carveout (2 GB by default, as low as 512 MB with no display attached) also comes out of the 128 GB. By that math a 200B model is ~100 GB at a bare 4 bits, or ~113 GB in NVFP4, which leaves roughly 13–15 GB for KV cache and the OS. That's why 200B is a ceiling rather than a comfortable fit. Good for prompt iteration, prototyping agentic workflows and checking model behaviour before moving to data-center GPUs.
• Sweet spot: 70B-class development at 8-bit (~70 GB of weights) with room to spare for context. Not FP16: Llama 3.3 70B's BF16 checkpoint is 141 GB on Hugging Face, more than the whole box. For dev-loop work where you don't need maximum throughput, this is the comfortable fit.
• Sweet spot: Mixed-model agentic prototyping, such as a 70B, a 30B and a 7B loaded side by side for draft → review → summarize loops. That's 54 GB of weights at 4-bit (60 GB in NVFP4), so all three stay loaded with plenty of room for context instead of being swapped in and out. At 8-bit (~107 GB) they still fit, but that leaves under 20 GB for the OS and three KV caches, so keep contexts short.
• Stretch: Anything bigger than one box. Qwen 3 235B (235B total, 22B active per token, per Qwen's model card) is above NVIDIA's 200B figure. At a bare 4 bits its weights alone are 118 GB, and in NVFP4 ~132 GB, more than the whole 128 GB box. The official FP8 checkpoint is 239 GB, so plan on two units. Llama 3.1 405B (405.85B parameters; ~203 GB at a bare 4 bits, ~228 GB in NVFP4) sits right at the two-unit limit. NVIDIA's user guide still rates a dual-Spark setup for up to 405B, as its Project DIGITS announcement did, while its product page says 400B. DeepSeek V3 (671B parameters in the main model, 685B including the multi-token-prediction module on Hugging Face; the official FP8 checkpoint is 689 GB) is a four-unit model. Its bare 4-bit weights (336 GB) fit in three units' 384 GB on paper, but in NVFP4 they come to ~377 GB, which leaves about 7 GB across three boxes for three OSes, display carveouts and KV cache. NVIDIA also recommends power-of-2 cluster sizes, so plan on four units (512 GB, NVIDIA's up-to-700B tier). Fine-tuning is a stretch too: NVIDIA's product page and launch release rate one unit for fine-tuning up to 70B (its user guide says up to 200B), and we haven't tested it.
• Bad fit: Production serving (any scale), maximum-speed single-model decode, pretraining, and anything that needs more than four units' 512 GB.

Bad use cases

• Production inference deployment. Wrong tier: pick L40S / H100 NVL / H200.
• Maximum speed on small models. If the model fits in 24–32 GB, a consumer GPU is the better tool (RTX 4090 / 5090). The 5090 has 6.6× the Spark's memory bandwidth (1,792 vs 273 GB/s, our math). DGX Spark is bandwidth-limited.
• Pretraining. Pretraining needs a lot of bandwidth and compute. DGX Spark is for development on top of trained models. NVIDIA's product page and launch release rate it for fine-tuning up to 70B (its user guide says training and fine-tuning up to 200B). We haven't tested either.
• Anyone whose stack doesn't have Arm support already. Audit your toolchain first. If you have many x86-only dependencies, this is friction you don't want.
• Long-horizon production reliance. First-gen platform, with software still changing from release to release. NVIDIA's DGX OS page says scheduled updates come about twice a year (around February and August) for the first two years after initial release, with more updates and security patches in between. In practice the Spark release notes list eight releases between November 2025 and October 2026, and those apply only to the Founders Edition; partner GB10 systems may get them later. NVIDIA doesn't publish an end-of-support date.

Verdict

Buy this if you develop against 70B–200B-class models (or 400B- and 700B-class with two or four linked units) that you'll later run on NVIDIA data-center hardware, you want 128 GB of unified memory (shared by the CPU, OS and GPU, minus a small display carveout) and NVIDIA's CUDA stack in a box that sits on a desk, your software stack is Arm-clean (or you're willing to sort out the long tail), and you understand this is dev/prototyping hardware, not production. NVIDIA's US Marketplace lists it at $6,950 (2026-11-01), up from a $3,999 MSRP that NVIDIA raised to $4,699 in February 2026. It isn't cheap, but it's a complete system. You're paying for capacity, not speed.

Skip this if you need maximum speed on smaller models (a discrete GPU wins), or your workloads fit in 24–32 GB: a used 3090 has 24 GB, and an RTX 5090 has 32 GB and is listed from $1,999 at NVIDIA (the Founders Edition was out of stock on NVIDIA's Marketplace on 2026-11-01), about 29% of the Spark's $6,950 (our math). Also skip it if you're production-serving (wrong tier entirely), you're locked into x86-only tooling and Arm friction is painful, or you want a platform whose software has stopped changing.

How it compares

• vs Mac Studio M3 Ultra (96 GB / 256 GB) → The closest comparable: unified memory in a small desktop. Apple's tech specs list the M3 Ultra at 819 GB/s of memory bandwidth, 3× the Spark's 273 GB/s (our math), with 96 GB configurable to 256 GB (it launched with a 512 GB option too), so the 256 GB configuration holds twice the Spark's memory in one box. Apple's current Mac Studio has since moved to M5 Max / M5 Ultra. The M5 Ultra lists 1.2 TB/s (about 4.4× the Spark, our math) and 96 GB configurable to 256 GB or 512 GB, so one current Mac Studio can hold as much memory as a four-unit Spark cluster (our math). The catch is no CUDA. NVIDIA dropped macOS after CUDA 10.2, so you're on Metal/MLX. Pick DGX Spark when CUDA is non-negotiable for your stack; pick a Mac Studio when MLX/Metal works and you want more bandwidth and more memory in one box.
• vs RTX 5090 (32 GB) → The 5090 wins on raw bandwidth (1,792 GB/s vs 273 GB/s) and on NVIDIA's headline AI compute figures: 3,352 AI TOPS vs the Spark's up to 1,000 TOPS (1 PFLOP FP4 with sparsity), about 3.4× (our math; the 5090 page doesn't say whether its figure uses sparsity). It also wins on price for everything that fits in 32 GB: NVIDIA lists it from $1,999, so the Spark at $6,950 (NVIDIA US Marketplace, 2026-11-01) costs about 3.5× as much (our math). The 5090 does still need a PC around it, and NVIDIA rates it at 575 W on its own. DGX Spark wins on memory ceiling (4× the capacity) for big-model dev. Pick the 5090 for hobbyist use and anything under 32 GB; pick DGX Spark for a 70B–200B dev box.
• vs RTX PRO 6000 Blackwell (96 GB) → The RTX PRO 6000 Blackwell Workstation Edition wins on bandwidth (1,792 GB/s, 6.6× the Spark) and on tensor compute. NVIDIA lists 4,000 AI TOPS for it; for the Spark, its spec table says up to 1 PFLOP FP4 and its hardware guide up to 1,000 TOPS, both theoretical sparse FP4, so 4× (our math). But the Workstation Edition is a 600 W card (NVIDIA's max power rating) that still needs a workstation around it, and its 96 GB is less than the Spark's 128 GB. The Max-Q version has the same 96 GB and 1,792 GB/s at 300 W (3,511 AI TOPS). NVIDIA's Marketplace lists either one at $16,000 for the card alone (2026-11-01), about 2.3× a complete DGX Spark at $6,950 (our math). For capacity in a complete, desk-sized system, pick DGX Spark. For serious workstation-tier inference speed on models that fit in 96 GB, pick the PRO 6000 Blackwell.
• vs renting H200 / B200 on cloud → Renting lets you run frontier inference at real speed. DGX Spark means slower iteration but no rental clock. Pick DGX Spark for the "no clock" benefit; rent for the "actual production speed" benefit. Most teams should do both.