Targeting edge AI computing applications, NVIDIA announced its ultra-compact AI supercomputer, powered by a streamlined version of the GB10 Grace Blackwell superchip developed in collaboration with MediaTek.DGX SparkIt is expected to go on sale from October 15, including the FE version sold directly by NVIDIA, which will be priced at US$3999. Customized systems based on the GB10 chip built by partners including Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, MSI and other brands will also be gradually put on the market.
In addition to announcing the official launch of DGX Spark on the market, NVIDIA CEO Jensen Huang also personally handed this desktop AI supercomputer system to Tesla CEO Elon Musk.
在DGX Spark中採用的GB10晶片，分別包含由10組Arm Cortex-X925核心與10組Arm Cortex-A725核心構成的20核CPU，而GPU部分則採用6144組CUDA核心，並且配置256位元、128GB容量的LPDDR5x-9400統一記憶體，FE版本則額外搭載4TB儲存容量。
至於連接能力方面，則分別內建英偉達ConnectX-7 200Gb/s高速有線網卡、對應Wi-Fi 7與藍牙5.3的無線網卡，以及10Gbps一般有線網卡，外部I/O則包含4組USB4 Type-C、1組HDMI、1組10Gbps RJ45，以及2組200Gbps QSFP。
The operating system uses DGX OS and is pre-loaded with the NVIDIA artificial intelligence software stack. A single machine can handle 1 petaflop of FP4 sparse AI computing power, support device-side inference capabilities for models with a scale of 2000 billion parameters, and local fine-tuning of models with a scale of 700 billion parameters. If two machines are interconnected through ConnectX-7, it can handle inference operations on a scale of 4050 billion parameters.
NVIDIA's reduction of its DGX series supercomputers to desktop form factors and their affordable pricing suggests it plans to push AI computing power further to the edge, making it easier for more developers and research institutions to deploy AI technology.
Compared with the previous DGX series supercomputers that started at hundreds of thousands or even millions of dollars, DGX Spark is sold at a price of only $3999. It will obviously attract more developers and start-up teams with limited budgets, and even allow more educational institutions to use it for artificial intelligence technology teaching.
However, judging by the needs of the general consumer market, DGX Spark is clearly still targeted at specific professional applications. However, compared to the previous situation where massive AI computing power had to be accessed through cloud services, it can now operate AI models with larger parameter scales in a standalone form, which is expected to drive the development of more edge AI applications.
