"Tinyboxes finally have a buy it now button"

Hardware & Specs

  • Two main variants: “red” with 6×7900 XTX GPUs ($15k) and “green” with 6×4090 GPUs ($25k).
  • Air‑cooled chassis, roughly 3200W max power draw, large under‑desk style case rather than dense rackmount.
  • CPU is described only generically (EPYC‑class); several commenters find the CPU/RAM specs vague and possibly underpowered relative to the GPU count.
  • PCIe 4.0 x16 links for each GPU; discussion around bandwidth marketing (32 GB/s per direction vs “64 GB/s” bidirectional spec).

Performance, Bottlenecks & Workloads

  • Some see it as strong GPU‑per‑dollar compared with high‑end multi‑4090 workstations.
  • Concerns that limited CPU and RAM could bottleneck workloads with heavy preprocessing.
  • Debate over PCIe and interconnect: fine for many workloads on a single box, but much slower than InfiniBand for large distributed training.
  • Several comments say it cannot realistically fine‑tune a 70B model; VRAM is borderline and interconnect is too weak for multi‑box scaling at that size.
  • One view: OCP 3.0 slot with ~200 Gbps NIC is adequate up to a few‑billion‑parameter models, but not beyond; 70B remains “unclear/likely no.”

Power, Electrical & Cooling Concerns

  • 3200W draw triggers extensive discussion of residential power: circuit derating to 80%, AFCI nuisance trips, transient GPU spikes, and code limits on continuous loads.
  • US users may need two separate 120V circuits or a dedicated 240V/20–50A feed; EU users note standard 230V outlets are not rated for 3.2kW 24/7.
  • Suggestions: dedicated circuits, mini‑split AC, separate shed, or basement placement; some argue this extra infrastructure undercuts the “cheap compute” story.
  • Significant concern about dumping 3kW of heat into typical homes, especially in hot climates.

Value, Pricing & Alternatives

  • Raw GPU cost of the red model is ~⅓ of total; the rest is attributed to CPU, motherboard, RAM, storage, PSUs, custom case, assembly, and support.
  • Some feel the box is overpriced versus DIY builds (e.g., ~$4k dual‑GPU rigs) or professional alternatives (Supermicro + H100 in colo, or cloud).
  • Others argue hobbyists undervalue engineering, integration, and “works out of the box” convenience.

Target Use Cases & Scalability

  • Seen as appealing for individuals or small teams wanting lots of local GPU without cloud.
  • Skepticism about using many of them as a cluster: networking is limited, form factor is bulky (≈15U each), and operations costs rise quickly.
  • Question over long‑term business viability and support; unclear how often hardware will be refreshed.