Nvidia Unveils DGX Spark and DGX Station Supercomputers at GTC 2025
At the GTC 2025 event in San Jose, Nvidia introduced two supercomputers – DGX Spark and DGX Station. These systems use the NVIDIA Blackwell Ultra platform, designed for developers, researchers, data scientists, and students.
Nvidia also introduced Next-gen GPUs: Blackwell Ultra (2025), Vera Rubin (2026-27), and Feynman (2028).
DGX Spark:
DGX Spark previously called Project Digit, it measures 150 x 150 x 50.5mm (5.9″ x 5.9″ x 2″)and available at $3999.
It features the GB10 Grace Blackwell Superchip with a Blackwell GPU capable of 1,000 trillion operations per second (TOPS).
DGX Spark features a Blackwell GPU with fifth-generation Tensor Cores, FP4 support, and NVLink-C2C, which enables high-bandwidth memory sharing between the GPU and Grace CPU.
It includes 128GB of LPDDR5x unified memory and up to 4TB of NVMe SSD storage.
DGX Spark Equipped with an M.2 slot for extra storage and supports WiFi 7, Bluetooth 5.3, and various ports including:
4 x USB4 (40 Gbps)
1 x 10 GbE LAN (ConnectX-7 Smart NIC)
1 x HDMI 2.1.
DGX station:
DGX Station is a larger desktop supercomputer that will launch later this year through partners like Asus, BOXX, Dell, HP, Lambda, and SuperMicro, with the price yet to be announced.
DGX Station resembles a PC tower instead of a typical data center unit.
It contains an NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip with 72 Neoverse V2 CPU cores and a Blackwell Ultra GPU.
DGX station offers up to 784GB of coherent memory (288GB GPU + 496GB LPDDR5x CPU).
It features ConnectX-8 SuperNIC for up to 800Gb/s speeds, ideal for large-scale AI workloads and compatible with NVIDIA CUDA-X AI, NVIDIA NIM microservices, and NVIDIA AI Enterprise.
Next Gen GPUs:
Blackwell Ultra: This GPU is designed to deliver 20 petaflops of performance, enhancing AI computation capabilities.
Blackwell Ultra introduces larger memory capacities compared to its predecessors, aiming to meet the increasing demands of complex AI models.
Vera Rubin: Vera Rubin anticipated to perform 3.3 times faster than systems running Blackwell Ultra, significantly boosting AI processing speeds.
Rubin Ultra: It expected to deliver 100 petaflops of performance, representing a substantial leap in AI processing power.
It designed as a rack-scale architecture capable of connecting 576 next-generation GPUs, facilitating large-scale AI data center operations.
Feynman: It expected to incorporate next-generation HBM5 memory, enhancing data throughput and efficiency for AI workloads.
Transforming Industries with AI Supercomputers
AI supercomputers are indispensable for advancing artificial intelligence research and applications.
They offer immense computational power, facilitating faster training, complex simulations, and large-scale data.
Industries such as healthcare, finance, and autonomous driving depend on AI supercomputers for breakthroughs in drug discovery, fraud detection, and self-driving technology, thereby accelerating innovation and addressing real-world challenges.
News Gist
At GTC 2025, Nvidia introduced the DGX Spark and DGX Station supercomputers powered by the Blackwell Ultra platform.
DGX Spark, priced at $3,999, offers 1,000 trillion operations per second.
DGX Station, launching later this year, features 784GB memory and supports high-speed AI workloads, collaborating with partners like Dell and HP.
Next-gen GPUs Blackwell Ultra, Vera Rubin, and Feynman are announced with enhanced AI computational capabilities.