NVIDIA Launches DGX Spark, Mini AI Supercomputer with Great Power
Introduction to NVIDIA DGX Spark
The giant semiconductor company from the United States, Nvidia, announced the presence of the DGX Spark which is claimed to be the world’s smallest AI supercomputer. With a very small size, this device is able to train and run advanced AI models directly on the desk without the need to hire cloud servers such as Amazon Web Services or Google Cloud.
The DGX Spark measures about 15 x 15 x 5 cm and weighs about 1.2 kilograms, similar to the Apple Mac Mini. Although small, its performance is equivalent to a data center (data center). With the ability to run a large AI model of up to 200 billion parameters, the DGX Spark can be regarded as a “private workstation” for researchers, developers, and students.
NVIDIA DGX Spark Technical Specifications
Technically, the DGX Spark brings the performance of the data center class into a device as small as a desk computer. The device is equipped with the Grace Blackwell Superchip GB10 chip, the latest combined CPU and GPU processor made by NVIDIA specially designed for artificial intelligence.
In terms of performance, the DGX Spark is capable of up to 1 computing petaflop, or about 1,000 trillion operations per second. This figure used to only be achieved by large-scale supercomputers at the research center. In addition, this device has a unified memory of 128 GB which is used together between the CPU and GPU to speed up data processing without a hitch.
For storage, the DGX Spark provides space up to 4 TB NVMe SSDs, large enough to accommodate large language models or large amounts of AI training data. In the connectivity sector, Nvidia embeds a 200 gigabit Ethernet network per second (GB/s) and NVLink-C2C technology, which has five times more bandwidth than PCIe Gen 5. fast.
For the operating system, Spark uses DGX OS, which is a derivative of Ubuntu Linux which has been optimized for GPU. This system is equipped with CUDA Library and NVIDIA NIM Microservices to be ready for use for AI research or experiments.
power consumption and excellence
Interestingly, all the performance only requires 240 watts of power, aka enough to be plugged into an ordinary home outlet, without the need for a complicated electrical system like an AI server in the data center. That is, with a small size and low power consumption, the DGX Spark allows anyone to have a personal AI supercomputer ready to use on the desk.
price and availability
In terms of price, the NVIDIA DGX Spark is priced starting at 3,999 US dollars (around Rp. 66.3 million). This price tag at first glance sounds expensive for the size of a desktop computer. However, when compared to devices in its class, the price is relatively affordable.
For comparison, professional-grade GPUs such as the RTX Pro 6000 are sold for around 9,000 US dollars (around Rp. 149.3 million). Meanwhile, GPU AI for servers such as the NVIDIA H100 even reached 25,000 US dollars (around Rp. 414.8 million) per unit. This means that the NVIDIA DGX Spark offers access to high-level AI computing at a much lower cost, although its performance is not as strong as the data center class chip.
Memory and performance advantage
According to the Register report, the GPU performance on the Grace Blackwell GB10 chip used by the DGX Spark is equivalent to the RTX 5070, but has a big advantage on the memory side. If the RTX 5070 only has 12 GB of video memory, then the DGX Spark carries 128 GB unified memory. This makes DGX Spark capable of running much larger AI models without having to rely on external cloud or servers.
With a price of Rp. 66 million, Spark is said to be a solution for universities, startups, and individuals who want to innovate in the field of AI without big capital.
