Computex 2026: NVIDIA Announcement Recap

June 9, 2026 0 By Lorena Mejia

NVIDIA RTX Spark

Announced as a groundbreaking “superchip,” the NVIDIA RTX Spark is designed to transform thin Windows laptops (down to 14mm and 3 lbs) and compact desktops into powerful, localized AI teammate ecosystems. By linking a 20-core NVIDIA Grace CPU to a Blackwell RTX GPU via a high-speed NVLink-C2C interconnect, the platform packs up to 1 Petaflop of AI compute and 128GB of unified memory to safely execute autonomous AI agents right on your device. The hardware integrates NVIDIA’s tech stack including CUDA, DLSS, and G-SYNC driving color-accurate tandem OLED displays while debuting a security partnership with Microsoft called NVIDIA OpenShell to safeguard user data. This premium, aluminum-clad hardware platform launches this fall from major manufacturers like ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and GIGABYTE systems to follow.

NVIDIA DLSS 4.5 Ray Reconstruction

Launching this August, the NVIDIA DLSS 4.5 Ray Reconstruction introduces a 2nd generation transformer model that unifies denoising and Super Resolution into a single AI network for all GeForce RTX GPUs. It replaces traditional hand-tuned denoisers with a supercomputer-trained model that boasts 35% more compute capability and processes 20% more parameters. The update dramatically enhances image quality, temporal stability, and lighting accuracy in ray-traced and path-traced games without sacrificing performance. Trained on an expanded dataset for superior spatial awareness closer to ground-truth visuals, this iteration also introduces finer temporal accumulation controls for developers, building directly upon DLSS 4.5’s core advancements like 6X Multi-Frame Generation to deliver sharper, clearer motion in next-gen gaming.

NVIDIA Vera CPU

The NVIDIA Vera CPU is a next-generation data center processor featuring 88 custom Olympus cores (176 threads) designed to eliminate infrastructure bottlenecks for agentic AI and reinforcement learning. By utilizing modular SOCAMM LPDDR5X memory, it achieves an unprecedented 1.2 TB/s of bandwidth and a 1.5 TB capacity, ensuring ultra-low-latency data access across its 164MB L3 cache. When deployed within the liquid-cooled Vera Rubin NVL72 rack-scale platform, this custom architecture completes local agent runtime loops up to 1.8x faster than traditional x86 server infrastructure while slashing token-processing costs tenfold.