RTX Spark can run some Windows apps through Prism, but that does not ensure a required device driver will work. XDA’s Adam Conway reported that Cyberpunk 2077 and Forza Horizon ran through Prism on an HP OmniBook Ultra 16. In a separate test, Windows detected an RTX 4080 eGPU, but NVIDIA’s Arm64 driver did not provide acceleration. A project author also reports running vLLM with CUDA on Windows ARM64, while Derivative staff described an emulation issue with TouchDesigner.
Prism emulation does not cover every driver
Prism translates x86 application code running in user mode, outside the core operating system. Kernel-mode drivers run within that core and need native Arm64 support. That is why an app can run through emulation while a driver for a connected device cannot.
| Windows on Arm path | Reported example | Practical implication |
| Emulated user-mode app | XDA reported Cyberpunk 2077 and Forza Horizon running through Prism on an HP OmniBook Ultra 16. | Those results apply to the named games in that test setup. |
| Kernel-mode GPU driver | In XDA’s test, Windows detected an RTX 4080 in a Plugable Thunderbolt 5 enclosure over a PCIe Gen4 x4 tunnel, but the NVIDIA Arm64 driver lacked the card’s device ID. | Windows used Microsoft Basic Display Adapter, so the connection did not deliver NVIDIA GPU acceleration. |
| Windows ARM64 AI software | A project author reported running vLLM with CUDA on RTX Spark. | One reported software configuration does not establish compatibility for other models or apps. |
The eGPU result separates a working connection from a working driver. The USB4 drivers established the PCIe tunnel and Windows detected the card; the missing device ID in the Arm64 NVIDIA driver prevented it from using NVIDIA acceleration. It is a result from that tested pairing, not a blanket verdict on every eGPU.
A separate account covers a Surface Laptop Ultra with RTX Spark, including apps running through Prism and controls for shared memory allocation.
A project author reports vLLM running on Windows ARM64
A project author reports running vLLM with CUDA on RTX Spark Windows ARM64, using Qwen3.8-Flash-Next NVFP4, MTP speculative decoding and CUDA graphs. The author’s project post reports 70.44 tokens per second for single-request decoding, averaged across 80 SPEED-Bench coding conversations. The author compared that result with 35.37 tokens per second for NVIDIA’s llama.cpp RTX Spark recipe.
That same post reports a separate 16-request coding workload at 285.85 tokens per second aggregate, or about 17.9 tokens per second per request on average. For AA-AgentPerf, the author reports 734 seconds for the project setup versus 1,118 seconds for NVIDIA’s recipe. These figures belong to different workloads and metrics, not a single general-purpose speed rating.
Windows compatibility is one part of the broader move toward running AI locally, covered in NeoTeo’s earlier overview of NVIDIA’s push to bring local AI into the home.
TouchDesigner has a separate emulation issue
When a user reported that the installer failed, Derivative staff suggested its /Extract option and /DIR to choose an extraction location; TouchDesigner.exe is in the extracted bin folder, according to the staff reply.
In a separate reply dated October 7, Derivative staff reported that mimalloc.dll did not work correctly in emulation and said the team was considering a fix. Staff also said native Arm builds were under consideration if Windows Arm hardware gained market share, as described in the follow-up.
NVIDIA’s RTX Spark specs describe hardware, not app support
NVIDIA’s Windows on Arm Porting Guide describes the highest-end RTX Spark SKU with a 20-core Arm v9.2 CPU, an integrated Blackwell GPU and up to 128 GB of coherently shared LPDDR5 memory. It lists a 256-bit memory interface and 300 GB/s of bandwidth. Those specifications describe the platform’s hardware; they do not determine whether a particular Windows app or driver supports Arm64.
