There have been rumors that NVIDIA will cooperate with MediaTek to build its own Arm architecture chips. The codenames mentioned at the time were N1 and N1X. At the earlier GTC Taipei conference, it was not surprising that this chip was actually brought, with the official name being “RTX Spark”, and it was announced that it would cooperate with Microsoft to create a new generation of Windows on Arm PC platform. NVIDIA claims that this is the most efficient PC chip in history, focusing on native AI agent tools, and will not only launch notebook products, but also mini computers.
NVIDIA RTX Spark officially unveiled: 1 PFLOP AI performance enters Windows laptops, mini computers will also be launched
RTX Spark is not a pure CPU, but an AI chip that integrates CPU, GPU, high-speed interconnect and unified memory. The CPU part is co-designed by NVIDIA and MediaTek, using Arm architecture, equipped with Blackwell RTX GPU, built-in 6,144 CUDA cores and fifth-generation Tensor Cores, supporting FP4 precision, and connected to the 20-core NVIDIA Grace CPU through NVLink-C2C.
NVIDIA claims that RTX Spark can provide up to 1 PFLOPS AI performance and is paired with up to 128GB of unified memory. In addition, the NVLink-C2C connection between RTX Spark’s GPU and CPU has a bandwidth of up to 600GB/s, which is 5 times that of PCIe Gen5:
Image source: NVIDIA via VideoCardz
NVIDIA emphasizes that RTX Spark is designed for “personal AI agents”, which means that AI agents can be executed on the computer instead of all through the cloud.
To provide users with peace of mind, NVIDIA and Microsoft will provide a new security foundation on Windows, including identity, isolation, policy and end-to-end security capabilities, so that native agents can execute under user control. NVIDIA provides an OpenShell runtime that allows users to define what the AI agent can and cannot do. It can also direct queries to local models in accordance with the privacy policy, or hide personal data before sending it to the cloud model.
Judging from the official application scenarios, RTX Spark can render ultra-large 3D scenes of more than 90GB, edit 12K 4:2:2 videos, generate 4K AI videos, and can also execute 120B parameter large language models locally, with a context length of up to 1 million tokens. Of course, it should be paired with 128GB of unified memory.
What’s more noteworthy is that Adobe will also optimize Photoshop and Premiere for RTX Spark, increasing AI and graphics performance by up to 2 times. Adobe Premiere will use RTX Spark’s unified memory, Blackwell GPU and TensorRT, and Photoshop will also be optimized for GPU-accelerated compositing, real-time filters, HDR and brush engines:
Of course, the gaming aspect isn’t too bad either. RTX Spark supports ray tracing technology, DLSS, Reflex and G-SYNC, and AAA games can provide over 100 FPS performance at 1440p.VideoCardz Supplement, NVIDIA has used “Indiana Jones and the Great Circle” as an example. The graphics performance can be close to the GeForce RTX 5070 Laptop GPU. However, NVIDIA has not announced the complete game list, image quality settings or power consumption limits:
Image source: NVIDIA via VideoCardz
RTX Spark notebooks will be available in 14-inch to 16-inch models, the thinnest can be 14mm, and the weight can be as low as about 1.36kg. Some models will use Tandem OLED screens and NVIDIA G-SYNC. In addition to laptops, there will also be mini computers:
The first wave of cooperative brands includes ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, and Acer and GIGABYTE will follow later. NVIDIA currently only says that RTX Spark notebooks and mini computers will be available this fall, and prices have not yet been announced.
Image source: NVIDIA via VideoCardz
Source: KOCPC Chinese