Nvidia announces Blackwell ASIC architecture with quantum integration and optimized AI server infrastructure

At Computex 2025, NVIDIA unveiled the DGX Spark, a compact personal AI supercomputer designed for developers, researchers, and data scientists. The system is powered by the NVIDIA GB10 Grace Blackwell Superchip, delivering 1 Petaflop of performance and featuring 128GB of unified system memory (LPDDR5X). The DGX Spark supports AI models with up to 200 billion parameters for local development and inference, and comes preloaded with the NVIDIA AI software stack including PyTorch and Jupyter.

NVIDIA announced significant advances in ASIC chiplet integration through NVLink Fusion, a technology that enables companies to integrate custom ASICs or CPUs with NVIDIA's GPUs using NVIDIA's high-speed NVLink interconnect. This approach provides high-speed connectivity, flexibility to pair NVIDIA processors with custom silicon, and scalability for demanding AI workloads such as large-scale model training and agentic AI inference. NVIDIA has partnered with MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys, and Cadence to bring this technology to market, enabling semi-custom designs for data centers, automotive applications, and industrial workloads.

The company also launched the NVIDIA Accelerated Quantum Research Centre (NVAQC) in Boston, Massachusetts, to address challenges in quantum computing including qubit noise and experimental quantum chip design. For AI factories, NVIDIA introduced the Blackwell Ultra GPUs, including the B300, which offers a 1.5x performance boost over standard Blackwell GPUs. The B300 provides 15 PFLOPS of dense FP4 compute and 30 PFLOPS with FP4 Tensor Cores, with 288GB of HBM3e memory, powering platforms like the GB300 NVL72 (72 GPUs) and HGX B300 NVL16 server-class unit.

Additionally, NVIDIA unveiled the DGX Station, a more powerful desktop supercomputer featuring the GB300 Grace Blackwell Ultra Desktop Superchip with 20 petaflops of AI performance and 784GB of unified memory, suitable for enterprise use as a standalone or centralized compute node. The company also announced new networking platforms—Spectrum-X and Quantum-X—leveraging silicon photonics to reduce energy consumption and support larger GPU clusters, alongside RTX Pro Blackwell GPUs for PCs and servers combining AI and graphics capabilities for professional users, and NVIDIA Dynamo, an open-source inference framework to enhance reasoning model performance on NVIDIA GPUs.