Nvidia在GTC会议上展示DGX Spark和DGX Station两款个人AI超级计算机，它们由Grace Blackwell平台驱动，适用于处理大型AI模型。Spark已开启预购，Station暂未定价。两款产品各有特点，且有多家OEM合作伙伴将推出相应版本。

🎈DGX Spark是Mac Mini大小的AI超级计算机，曾名Digits，采用GB10 Blackwell Superchip，具备强大性能

🚀DGX Station配备更强大的GB300 Blackwell Ultra桌面超级芯片，性能强劲

🤝多家OEM伙伴将制造DGX计算机的不同版本，包括华硕、戴尔、惠普等

Nvidia has revealed its new DGX Spark and DGX Station “personal AI supercomputers” at today’s GTC conference, which are both powered by the company’s Grace Blackwell platform and designed for users to work on large AI models with or without a connection to a datacenter. The Spark is going up for preorder today.

The Spark is the new name for Nvidia’s $3,000, Mac Mini-sized “world’s smallest AI supercomputer” that was announced with the name “Digits” at CES earlier this year. Its larger Station counterpart, currently with no price tag, is aimed at “AI developers, researchers, data scientists and students to prototype, fine-tune and inference large models on desktops.”

The Spark is powered by Nvidia’s GB10 Blackwell Superchip, featuring a GPU with fifth-generation Tensor Cores and FP4 support. The GB10 is optimized for the Spark’s smaller desktop form factor. However, it can still deliver “up to 1,000 trillion operations per second (TOPS) of AI compute for fine-tuning and inference with the latest AI reasoning models, including the NVIDIA Cosmos Reason world foundation model and NVIDIA GR00T N1 robot foundation model.” Spark features 128GB of unified memory, and up to 4TB of NVMe SSD storage.

@verge

DIGITS is a personal AI supercomputer that can handle huge AI models, up to 200 billion parameters, all while running from a standard power outlet. It uses NVIDIA’s new GB10 Grace Blackwell Superchip, which will allow users to develop and test AI models locally. These functions historically used to need larger processing power systems to work, but now is sized perfectly be atop a desk. #ai #supercomputer #tech #techtok #ces

♬ original sound – The Verge

Meanwhile, the Station, with its larger size, accommodates Nvidia’s recently announced more powerful GB300 Blackwell Ultra desktop superchip that “provides 20 petaflops of AI performance and 784GB of unified system memory.”

Nvidia also said OEM partners will make versions of the DGX computers. Asus, Dell, HP, Boxx, Lambda, and Supermicro will build their own Stations, which will be available later this year. Meanwhile, the Spark will have versions made by Asus, Dell, HP, and Lenovo. You can reserve one today on Nvidia’s website, with deliveries expected this summer.

Nvidia isn’t the only company building GPUs with lots of unified memory that can be used for local LLMs. Rival AMD has the Ryzen AI Max+ “Strix Halo,” and HP is putting the 128GB version into a laptop, while Framework has put it into a $2,000 desktop. Both let the GPU access up to 96GB of VRAM.