Tiny corp is shipping its Tinybox AI hardware line, led by the Green v2 configuration rated at 3,086 TFLOPS of FP16 compute with 384 GB of GPU VRAM, according to specifications listed on tinygrad.org.
Hardware Specifications and Lineup
Tiny corp currently offers two shipping desktop configurations:
- Tinybox Red v2: 4x AMD 9070XT GPUs, 64 GB GPU VRAM, 2,560 GB/s bandwidth, 778 TFLOPS FP16 compute, 32-core AMD EPYC CPU, 128 GB system RAM, single 1600W power supply.
- Tinybox Green v2: 4x RTX PRO 6000 GPUs, 384 GB GPU VRAM, 7,168 GB/s bandwidth, 3,086 TFLOPS FP16 compute, 32-core AMD Genoa CPU, 192 GB system RAM, dual 1600W power supplies.
Both shipping units run Ubuntu 24.04 LTS, with the Green v2 offering 59.3 GB/s disk read bandwidth over PCIe 5.0 x16.
Tiny corp also lists an "Exabox" as a 2027 preorder item — a multi-rack system targeting roughly 1 exaFLOP of FP16 compute using 720 RDNA5 AT0 XL GPUs, 25,920 GB of GPU VRAM, and 120 AMD Genoa CPUs, drawing about 600 kW. This product has not shipped and is listed only as a future preorder.
Software: tinygrad
The hardware runs on tinygrad, tiny corp's open-source neural network framework, which reduces neural network operations to three core types — ElementwiseOps, ReduceOps, and MovementOps. Tinygrad uses lazy tensor evaluation to fuse operations before compilation and generates a custom kernel for each tensor shape, aiming for high hardware utilization without vendor-specific deep learning libraries.
According to tinygrad.org, the framework remains in alpha and tiny corp has said it will exit alpha once it can reproduce a set of benchmark papers at twice the speed of PyTorch on a single Nvidia GPU.
Purchasing
Per tiny corp's site, the company has no sales representatives, does not complete custom vendor onboarding paperwork, and accepts payment only by wire transfer through its website.
Why it matters
Tiny corp is pitching bare-metal GPU boxes paired with a minimalist compiler as a lower-cost alternative to enterprise AI infrastructure. Whether that holds up depends largely on tinygrad clearing its own stated performance bar against PyTorch, which by the company's own account it has not yet done.