Tinybox – A powerful computer for deep learning
Tiny Corp’s new “Tinybox” deep learning machines — ranging from a $12k desktop-style rig to a hypothetical $10m container-scale “exabox” — are prompting debate over who actually needs local AI hardware at these prices when cloud inference is cheap and often easier. Commenters weigh the appeal of prebuilt, AMD‑based systems and tight integration with the tinygrad stack against concerns about value for money, driver maturity, power and cooling requirements, and the fact that similar configurations can be self-built for less. The project’s uncompromising sales model, strong opinions on democracy and governance, and refusal to accommodate typical enterprise procurement add a layer of skepticism about how well this approach will fit serious B2B buyers despite clear enthusiasm from some local‑AI enthusiasts.
Overall impressions
- Many see Tinybox as a nicely packaged, well-balanced prebuilt version of what enthusiasts already assemble for local LLMs, not revolutionary but convenient.
- Aesthetics and “human-written” website copy are praised; some like the clear, opinionated tone, others find it off-putting or arrogant.
Hardware, performance, and use cases
- Red/Green Tinyboxes are viewed as solid inference machines, but several doubt they can run 120B-parameter models at “comfortable” speeds without aggressive quantization or offloading.
- Some want advertised tokens/sec on specific open-source models to reduce buying risk.
- Others argue modern laptops (M-series, Strix Halo) or consumer multi-GPU rigs can reach acceptable performance for far less.
Pricing, value, and alternatives
- Strong split: some call $12k–$65k “insane” for commodity components; others say it’s cheap relative to engineer salaries or enterprise GPU pricing.
- Comparisons: DGX Spark, Mac Studio/Mini, DIY RTX/AMD builds, used data-center GPUs, and cheap cloud inference (e.g., $/M tok) often look more cost-effective.
- Markup is seen by some as a way to fund the tinygrad project rather than pure hardware value.
AMD vs NVIDIA and software
- Surprise at heavy AMD usage given past criticism, but several say ROCm has improved a lot recently.
- Others report ongoing ecosystem gaps, CUDA-only tools, and rough edges on consumer AMD cards.
Exabox concept
- The $10M exabox (massive AMD-based cluster in a container-like form factor) is read by some as semi-jokey “vaporware,” by others as a probe for hyperscale-style interest (e.g., startups, privacy-sensitive sectors).
- Specs, weight, and power draw (≈600 kW) spark debate about practicality; comparisons made to NVIDIA rack-scale systems and TPU pods.
Power, cooling, and deployment
- 3.2 kW draw raises home/office power questions (dual 120V vs 240V circuits), leading to a long code/safety discussion.
- Several say serious buyers will just colocate where 208–240V and cooling are standard.
Business model, ordering, and trust
- Wire-transfer-only payment and “no customization, just order through the site” policies worry some as scam-like or incompatible with normal B2B procurement.
- Others defend wire transfers as standard for large hardware and see the anti-onboarding-form stance as a deliberate rejection of enterprise bureaucracy.
Founder politics and hiring practices
- A linked anti-democracy blog post and past political statements cause some to lose interest or describe the broader ideology (great-man, anti-democratic, “authoritarian techno-libertarian”).
- Hiring requirement to have prior tinygrad contributions (with small, discretionary bounties) is criticized as asking for unpaid spec work; defenders point to bounties as a fit filter.