Nvidia RTX Spark SoC Impresses in Hands-On Test Across AI, Creative and Gaming Tasks

The publication said its reviewer attended Nvidia's RTX Spark event in San Francisco earlier this month and then spent almost every waking moment testing the chip. The RTX Spark is Nvidia's first system-on-a-chip, combining an 18-core CPU and a 5,120-CUDA-core Blackwell RTX GPU in one design, or a 20-core CPU and 6,144-core GPU in a higher-end version. TechRadar described it as an integrated chipset in the vein of Apple's M-series silicon, with the Blackwell RTX GPU also bringing considerable gaming performance.

In Geekbench 7.1 tests of the RTX Spark N1X chip, which has 20 CPU cores, TechRadar recorded a single-core score of 2,623, below the 4,288 it got for Apple's M5 chip. Its multi-core score of 21,899 far exceeded the M5's 17,926, which TechRadar said should translate into more power for tasks such as content creation, software development and AI.

TechRadar said Nvidia focused heavily on hybrid agentic AI at launch. It reported that the RTX Spark lets users choose local models and ring-fence them for parts of an agentic workflow while sending specific demanding tasks to cloud frontier models. That approach gives users more granular control over how much they rely on public versus private models, the publication said, rather than forcing a choice between loose security and a full lockdown. TechRadar said the hybrid workflow could be valuable for people who vibe-code complex products or generate demanding 3D scenes.

The reviewer said cloud-based AI models are powerful but raised security concerns about uploading private data and blindly trusting the companies behind some models. Local open-source models that can be vetted and never transmit sensitive data outside the network feel safer, TechRadar said, but not everyone wants to cut themselves off from frontier models.

In hands-on agentic tests, TechRadar ran the MLPerf Client benchmark to simulate data analyst and software engineering agent workflows using the Qwen 3 8B local model. It said each run averaged 63.7 and 50.8 seconds and processed up to 14,005 and 10,238 tokens, respectively. In ComfyUI generative image and video workflows, the Flux2.Dev NVFP4 and Flux2.Klein 4B Base NVFP4 models generated images in 46.5 and 16.2 seconds, while the LTX 2.5 (Distilled) NVFP4 model took 88 seconds. TechRadar said the results were broadly in line with Nvidia's promises and showed how quickly the RTX Spark can generate media using local models.

The publication also spent days training its own AI agent with Hermes and the Qwen3.8 27B open-source local model. On its Extra High setting, the agent was given CSV analytics data and asked to tidy the spreadsheet, analyze the data and present key findings; after 12 minutes it produced a pivot table and observations. TechRadar said it was not surprised by the conclusions but was impressed that the agent identified the most essential points without secure information leaving the laptop.

A separate email-digest experiment was less smooth. TechRadar said the first attempt to connect its Hermes agent to an online email client spent 75 minutes trying methods to exchange tokens with Outlook's online client before crashing unexpectedly. On a second attempt, with the reviewer providing feedback when the agent drifted or hit hurdles, it set up the brief in 70 minutes. TechRadar said some experiments required patience and active prompting to get the best results.