NVIDIA debuts world’s smallest personal supercomputer to push boundaries of generative AI

NVIDIA founder and CEO, Jensen Huang revealed to the world the DGX Spark AI supercomputer during the keynote of the GTC 2025 conference in California. Formerly codenamed Project DIGITS, the high-performance desktop supercomputer is powered by the NVIDIA Blackwell Ultra platform. This model can run locally or on an accelerated cloud data center, allowing AI developers, data scientists, robotic developers, researchers, and students to develop applications or conduct generative AI studies without any computing limitations.

At the keynote, Jensen said that AI has “transformed every layer of the computing stack. It stands to reason a new class of computers would emerge—designed for AI-native developers and to run AI-native applications.” According to him, this personal AI computer will take forward AI on the cloud, desktop, or edge applications. Along with the DGX Spark supercomputer, the company also revealed the DGX Station to leverage the capabilities of the Grace Blackwell architecture for more professional applications.

Designer: NVIDIA

The DGX Spark is the world’s smallest AI supercomputer destined to overshadow the current generation of computing machines with its exponentially powerful hardware. This compact machine (about the size of an Apple M4 Mac mini) is powered by the small NVIDIA GB10 Grace Blackwell system-on-chip that has a GPU employing fifth-generation Tensor Cores and boasts FP4 support. Just for reference, the chip paired with 128GB unified system memory can perform 1,000 trillion AI computing operations per second.

This makes the new supercomputer ideal for AI reasoning models that demand such mind-boggling levels of computing power. Both the GB10 Superchip and the onboard GPU are optimized for memory-intensive AI workloads thanks to the coherent memory model having 5 times the capacity of a fifth-generation PCIe. The transfer of created models from a user’s desktop to the cloud is going to be equally seamless, making it easier to prototype or fine-tune the workflows. The DGX Spark can be reserved right away for $3,999 with shipping expected by summer this year.

To support the computing power of the smaller desktop model is the DGX Station which is more of an enterprise solution for professionals. It comes with the GB300 Grace Blackwell Ultra Desktop Superchip that offers 20 petaflops of performance and 784GB of unified system memory. The solution gets a 72-core Grace Neoverse V2 CPU and has support for FP4 precision. This machine will come loaded with a ConnectX-8 SuperNIC network interface card supporting 800Gb/s throughputs. That is an insane amount of power, but it will eventually become the norm in coming years given the requirement of AI-based smart systems. There is no word yet on the pricing or availability of the DGX Station, but it is expected to be released later this year with manufacturing partners like Asus, Dell, HP, and Lenovo.

The post NVIDIA debuts world’s smallest personal supercomputer to push boundaries of generative AI first appeared on Yanko Design.