Nvidia's Project Digits is a 'personal AI supercomputer'

Original Article ↗ Hacker News Discussion ↗

Nvidia’s new Project Digits aims to put a compact “AI supercomputer” — an ARM‑based Linux box with a Grace Blackwell GB10 chip, 128GB of unified memory and up to 4TB NVMe storage — on the desks of startups, researchers and enthusiasts for around $3,000. Commenters weigh its trade‑offs against RTX 50‑series GPUs, Apple’s high‑RAM Macs and upcoming AMD Strix Halo systems, seeing strong value in local, privacy‑preserving LLM inference but raising questions about memory bandwidth, power limits and Nvidia’s long‑term Linux/Jetson‑style software support.

Hardware & Architecture

  • Compact ARM-based Linux workstation built around the GB10 “Grace Blackwell” superchip.
  • ~1 PFLOP of FP4 AI compute, 128 GB unified LPDDR5X memory, up to 4 TB NVMe storage, 20 CPU cores (10 Cortex‑X925 + 10 Cortex‑A725), ConnectX NIC with two QSFP ports for stacking two units.
  • Unified memory shared by CPU/GPU is a core design point; bandwidth is speculated around ~500 GB/s but not confirmed. FP32/FP16 support level is unclear.

Price, Configurations & Value

  • Announced “starting at $3,000”.
  • Nvidia materials say every unit has 128 GB unified memory; only storage and possibly networking/clock/binning are expected to vary, but that’s not fully confirmed.
  • Some call $3k “cheap” versus Mac Studio / MacBook Pro with 128 GB or multi‑GPU PCs; others find it steep and wish for a sub‑$1k/Jetson‑like option.

Performance vs GPUs, Macs & Alternatives

  • Raw GPU compute is well below RTX 5090/4090; estimates place it around 4070–5070 class in TOPS, far lower memory bandwidth than high‑end gaming cards.
  • Strength is capacity and efficiency: 128 GB addressable by the GPU in a small, relatively low‑power box vs 24–32 GB on consumer GPUs.
  • Seen as a direct challenger to Apple Silicon for local LLMs (M2/M4 Max/Ultra) and to AMD Strix Halo / Ryzen AI Max+ designs, with higher AI throughput but uncertain CPU competitiveness.

Use Cases & Target Users

  • Positioned for AI researchers, startups, labs, and “serious enthusiasts” doing local LLM inference, fine‑tuning, RAG, and experimentation, not as a living‑room PC.
  • At least some commenters see it as a modern Jetson‑style dev kit and “micro‑DGX” rather than a mass consumer product.
  • Stacking two units (via ConnectX) is advertised for ~400B‑parameter‑class models at low‑precision inference.

OS, Tooling & Ecosystem

  • Ships with Nvidia’s DGX OS (Ubuntu 22.04–based, Nvidia‑optimized kernel).
  • Nvidia is pushing Linux/WSL2 as the primary developer environment; Win32 is de‑emphasized for new AI tooling.
  • Many view it as an “onboarding path” that further entrenches the CUDA/Nvidia AI ecosystem, similar to what GeForce did for gaming.

Concerns & Skepticism

  • Unclear longevity and upstream support, given Nvidia’s history with Jetson boards (short lifecycles, outdated Ubuntu, awkward toolchains).
  • Worries about opaque, vendor‑locked software stack and future kernel/driver updates.
  • Real‑world tokens/sec heavily depend on actual memory bandwidth; some fear it may feel slow on very large models despite fitting them.
  • Gaming suitability, exact power draw, and ability to train (not just infer) at higher precision remain unclear.