TinyGrad, the company behind the popular TinyBox system, is aiming to commoditize the PetaFLOP. Its latest powerhouse, the TinyBox Pro, is on display on X. This high-performance system boasts an impressive 1.36 PetaFLOPS of FP16 computing power and is based on commercial GPU. The TinyBox Pro configuration features eight NVIDIA RTX 4090 GPUs, surpassing its predecessors with a combined GPU RAM of 192 GB and a memory bandwidth of 8,064 GB/s. This substantial upgrade is complemented by dual AMD Genoa processors and 384 GB of system RAM, delivering a memory bandwidth of 921.6 GB/s. What sets the TinyBox Pro apart is its enterprise-grade architecture. The system utilizes four 2000 W power supplies requiring 200 V+ input, housed in a 4U form factor that spans 31 inches in depth. Despite its compact size, weighing 88 pounds, the unit comes equipped with Supermicro rails for seamless rack integration.

Connectivity options are equally impressive, featuring two open PCIe 5.0 x16 slots that provide extensive expansion capabilities. Storage is managed through a 1 TB boot drive, though this might seem conservative compared to some competitors' offerings. The system runs on Ubuntu 22.04 and is noted for its superior driver quality (compared to commercial AMD GPUs), addressing a common pain point in high-performance computing on commercial hardware. On social media, TinyGrad was very vocal about its fight with the AMD Radeon GPU drivers. However, potential buyers should be prepared for significant noise levels during operation, a trade-off for the remarkable computing power packed into the 4U chassis. With a pre-order price tag of $40,000, the TinyBox Pro positions itself as a serious contender in the professional AI computing market, where regular GPU boxes can cost 100s of thousands of US Dollars. This pricing reflects its enterprise-grade specifications and positions it as an accessible alternative to larger, more expensive computing clusters.