NVIDIA DGX Spark (Project Digits)

What it does well

The DGX Spark (also marketed as Project Digits) is NVIDIA's first true ARM-based desk-side AI development workstation. The headline feature is unified memory — 128 GB of LPDDR5X shared between the Grace ARM CPU and the Blackwell-generation GPU, accessible via NVLink-C2C at ~600 GB/s, in a small form factor that runs from a 240 W power adapter. For local development of frontier models, this is genuinely unique: it fits Llama 3.3 405B at Q3, DeepSeek V3 671B at Q2 with paged offload, or Qwen 3 235B at FP8 / Q4 with comfortable context — workloads that no other $3,000 device on Earth can host. The full CUDA stack works (with ARM-compiled binaries; CUDA 12.8+ ships first-class ARM/Grace support), so you can dev against the same software stack you'll deploy on H200/B200 clusters. NVIDIA provides the entire NeMo + DGX OS + JupyterLab pre-configured stack out of the box. At $3,000 retail, it's roughly 1/3 the price of a 96 GB workstation card alone — and it's a complete system with CPU + RAM + storage + cooling.

Where it breaks

• Bandwidth ceiling vs discrete GPU. 600 GB/s NVLink-C2C between CPU and GPU is dramatically below an H100's 3.35 TB/s or even an RTX 5090's 1.79 TB/s. Decode speed for memory-bound workloads is meaningfully slower — a 405B Q3 might run at 5–~10 tok/s rather than the ~25–40 tok/s a 4× H100 box delivers.
• ARM ecosystem still has friction. While CUDA on ARM works, many third-party tools (some Python packages, certain Docker images, niche binaries) don't ship native ARM builds. Expect occasional 'pip install' failures, x86 emulation slowdowns, and Docker images that don't run.
• Power envelope limits sustained workloads. 240 W total system power means GPU is never running at "actual H100 / B200" wattage. Sustained inference on big models throttles compared to discrete-GPU equivalents.
• Single-system, no multi-card scale. No NVLink to other Spark units, no PCIe expansion. What you buy is what you get. For workloads that grow beyond 128 GB, you're recommitting to a different platform.
• First-generation product risk. Software ecosystem (drivers, distro support, niche tooling) will mature over 12–24 months. Buying day-one means living through the maturity curve.
• Not for production serving. Single-user dev box, period. Don't deploy this as a production server.

Ideal model range

• Sweet spot: Local development of 200B–671B-class models — fit them in memory at low quant for prompt iteration, prototyping agentic workflows, and validating model behavior before deploying to H200/B200 cluster.
• Sweet spot: 70B–200B-class development at FP16 with comfortable context. For dev-loop work where you don't need maximum throughput, this is the cheapest path.
• Sweet spot: Mixed-model agentic prototyping — fit 70B + 30B + 7B simultaneously for draft → review → summarize loops without offload thrashing.
• Stretch: Light fine-tuning at 7B–13B QLoRA. Bandwidth-limited but functional.
• Bad fit: Production serving (any scale), high-throughput single-model decode, training, multi-card scale-out workloads.

Bad use cases

• Production inference deployment. Wrong tier — pick L40S / H100 PCIe / H200.
• Maximum tok/s on small models. Sub-13B at >~150 tok/s is consumer GPU territory (RTX 4090 / 5090). DGX Spark is bandwidth-limited.
• Training workloads. Training is bandwidth-and-compute-heavy. DGX Spark is for development on top of trained models, not training.
• Anyone whose stack doesn't have ARM support already. Audit your toolchain first. If you have many CUDA-x86-only dependencies, this is friction you don't want.
• Long-horizon production reliance. First-gen platform. Expect software maturity issues for 12–18 months.

Verdict

Buy this if you do local development on frontier (200B+) models that you'll deploy to H200/B200 clusters, you want a desk-side dev box that fits 128 GB of model weights, 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. The DGX Spark hits a unique price point — there is genuinely no other $3,000 device that does what this does.

Skip this if you need maximum tok/s on smaller models (consumer GPU wins), your workloads fit 24–48 GB (RTX 5090 or used 3090 is dramatically cheaper), you're production-serving (wrong tier entirely), you're CUDA-x86-locked and ARM friction is painful, or you want a mature first-day-functional platform (wait 12 months).

How it compares

• vs Mac Studio M3 Ultra (192 GB) → Mac Studio at 192 GB unified memory is the closest comparable: more memory, similar dev-tier positioning, more mature ARM ecosystem (Apple Silicon has 3+ years), but no CUDA. Pick DGX Spark when CUDA-on-ARM is non-negotiable for your stack; Mac Studio when MLX/Metal works and you want more memory. See /compare/nvidia-dgx-spark-vs-mac-studio-m3-ultra.
• vs RTX 5090 (32 GB) → 5090 wins on raw bandwidth (1.79 TB/s vs 600 GB/s), tensor compute, and price-per-performance for everything that fits 32 GB. DGX Spark wins on memory ceiling (4× the VRAM-equivalent) for frontier-scale dev. Pick 5090 for hobbyist / sub-32 GB; DGX Spark for 200B+ dev box.
• vs RTX PRO 6000 Blackwell (96 GB) → PRO 6000 Blackwell wins on bandwidth (1.79 TB/s) and pure tensor compute, at 2.8× the GPU price. DGX Spark at $3,000 includes the full system; PRO 6000 Blackwell at $8,499 is just the card. For cost-conscious frontier dev, DGX Spark wins; for serious workstation-tier prosumer inference, PRO 6000 Blackwell.
• vs renting H200 / B200 on cloud → Renting H200 at $3–$4.50/hr lets you actually run frontier inference at speed. DGX Spark dev 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.