MediaTek Joins NVIDIA to Build RTX Spark Superchip

According to NVIDIA's May 31, 2026 press release, NVIDIA unveiled RTX Spark, described as a new Arm-based superchip for Windows PCs built to run local personal agents, content-creation workflows, and high-end gaming. NVIDIA's release states the chip provides approximately 1 petaflop of AI performance, supports up to 128GB of unified memory, integrates a 20-core Grace CPU, and includes up to 6,144 Blackwell CUDA cores (NVIDIA press release). The release also specifies support for running 120B-parameter LLMs with contexts up to 1 million tokens and for handling ultralarge 90GB+ 3D scenes and 12K 4:2:2 video workloads (NVIDIA press release).

MediaTek is named as a collaborator on the CPU design in NVIDIA's announcement, and a MediaTek statement quoted in PR Newswire via The Manila Times attributes a direct quote to Vince Hu, Corporate Senior Vice President at MediaTek (PR Newswire / The Manila Times). PCMag's coverage and PC Watch's technical summary note the chip's similarity to NVIDIA's GB10/DGX Spark-class hardware and list early OEM commitments including ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI; additional outlets report Acer and GIGABYTE as follow-on partners (PCMag; PC Watch; Digitimes snippet).

NVIDIA's materials say RTX Spark combines technologies including NVIDIA CUDA, RTX ray-tracing, DLSS, FP4 precision, TensorRT, OptiX, Reflex and a custom NVLink chip-to-chip interconnect; the press release cites a high-performance interconnect and unified memory architecture as enabling the combined CPU+GPU+AI workloads (NVIDIA press release). PC Watch reports the chip's SKU-level specs-20-core Grace CPU, 6,144 Blackwell GPU cores, NVLink C2C interconnect and up to 128GB memory-as aligning closely with the previously announced GB10/DGX Spark family (PC Watch).

For practitioners: the integration of a multicore Arm CPU with a large Blackwell GPU and unified memory on a single SoC follows a broader industry trend toward tightly coupled CPU-GPU-AI fabrics for low-latency, high-throughput local model execution. Companies shipping comparable designs prioritize memory bandwidth, interconnect efficiency, and power-managed performance states to sustain extended agent and content-creation workloads without thermal throttling.

public reporting frames RTX Spark as NVIDIA's biggest push into consumer Windows PCs to date, aiming to make localized agentic AI and frontier-sized models feasible on laptops and small desktops. The inclusion of MediaTek as a CPU-design collaborator, reported in NVIDIA's announcement and in PR Newswire, signals cross-vendor cooperation between a major PC-software/AI vendor and a leading Arm SoC designer for client devices (NVIDIA press release; PR Newswire / The Manila Times). Early OEM listings across multiple outlets indicate strong partner interest ahead of fall product timelines (PCMag; Digitimes snippet).

Editorial analysis: observers will watch product power envelopes, sustained performance figures for long-running agent workloads, and real-world memory utilization when running 120B-class models or multistream 4K/12K media pipelines. Industry observers should also track software integrations-NVIDIA cites work with Microsoft on Windows agent primitives and with Adobe on rearchitecting Photoshop and Premiere for RTX Spark in NVIDIA's announcement-which will determine how effectively applications can leverage the new hardware (NVIDIA press release).

NVIDIA's announcement contains direct executive quotes, including Jensen Huang's statement that "The PC is being reinvented" (NVIDIA press release). MediaTek's Vince Hu is quoted as saying, "With RTX Spark, MediaTek and NVIDIA bring unprecedented local AI supercomputing directly to consumer PCs" in the PR Newswire text quoted by The Manila Times (PR Newswire / The Manila Times).