Nvidia Details RTX Spark Platform Combining Grace CPU And Blackwell GPU

Nvidia has detailed the technical architecture of its RTX Spark processor platform, engineered to power high-performance Windows laptops and compact workstations. Fabricated as an integrated system-on-chip, the silicon combines an energy-efficient central processing unit with desktop-class graphics hardware on a unified package.

NVIDIA and Microsoft readied developer tools for RTX Spark, including the first PyTorch-CUDA build for Windows on Arm and support for 97% of the top 15,000 Python packages.

RTX Spark laptops are available globally starting at $2,599.

As a point of comparison, Apple M4 Max delivers approximately 540 GB/s bandwidth, enabling the M4 Max to generate tokens faster for models up to roughly 70B parameters.

Developed in engineering partnership with MediaTek, the platform couples custom Grace CPU silicon alongside Blackwell graphics compute units. The design marks an aggressive expansion into personal computing, targeting creative professionals and software developers running local artificial intelligence models.

Silicon Microarchitecture And Interconnect Bandwidth

Manufactured using TSMC advanced three-nanometer semiconductor lithography, the processor functions across a configurable thermal profile ranging from 45W to 80W. This power envelope enables notebook manufacturers to construct high-throughput portable machines weighing approximately three pounds while sustaining up to 99 percent of peak mains performance on battery reserves.

Internal silicon communication relies on an NVLink-C2C interconnect that provides 600 gigabytes per second of bidirectional bandwidth between processing elements. This interface allows the Grace processing cores and Blackwell graphics units to access a shared unified memory pool of up to 128GB LPDDR5X operating at 9,400 Mbps.

Compute Configurations And Localized Intelligence

Under premium N1X hardware configurations, the processor provides up to 20 CPU execution cores split evenly between performance and efficiency clusters, paired with 6,144 graphics compute cores. A secondary configuration delivers 18 central processing cores alongside 5,120 graphics cores to accommodate thinner form factors.

By pairing fifth-generation Tensor hardware with unified memory pools, the architecture processes massive generative models locally on-device. This localized execution allows developers and enterprise users to deploy agentic software tools securely without sending sensitive data through external cloud networks.