TL;DR

For playersOffer from Amazon

Play games on Amazon Luna for your next adventure, included with Prime

• A rotating selection of games, no download needed
• Play on TV, laptop or phone
• Fast, free delivery for your gear, too

Start playing with Prime →Free trial · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

An XDA Developers investigation found evidence that Nvidia’s RTX Spark contains a physical neural processing unit, not merely a software interface that sends Copilot+ tasks to the GPU. The writer ran ResNet-50 on the 55-TOPS NPU in FP16 and INT8, though INT8 was not faster, and Nvidia has not publicly explained much about the accelerator or whether related hardware is present in GB10 systems.

An XDA Developers investigation found that Nvidia’s RTX Spark includes a separate neural processing unit capable of running neural networks, rather than merely presenting the GPU as an NPU for Windows Copilot+ features. The device is rated at 55 TOPS, but Nvidia has offered little public detail about its design or intended uses, and testing found no speed advantage for INT8 over FP16 on ResNet-50.

XDA’s Lead Technical Editor Adam Conway said he spent about a week examining firmware, drivers and Nvidia development tools to identify the accelerator. His investigation found that the NPU has its own registers, clock domain and power reporting, separate from the RTX Spark’s GPU. Firmware on the HP system also includes a “DLA Support” setting and ACPI logic that can hide the device from Windows.

Conway reported successfully running neural networks on the NPU and checking that the GPU was not secretly carrying out the computation. He ran ResNet-50 in both FP16 and INT8 and said both produced correct results. However, the INT8 run was not faster than FP16, despite INT8 often being used for lower-precision inference. Nvidia’s platform documentation, as described in the report, says the platform does not support INT8, making the working test another unresolved feature of the device.

The Windows driver identifies the accelerator as “NVIDIA NPU,” while internal driver references call it “NVIDIA_DEV.DLA.” Tools examined by Conway identify its architecture as NPU3, and its device UUID begins with “DLA-.” He also found it has a separate clock, reported as idling at 430 MHz and reaching 1,562 MHz. These findings support the case that it is a distinct accelerator, although they do not establish its full capabilities or performance across workloads.

A Separate Path for AI Workloads

The findings matter because the NPU appears to be physical compute hardware with its own device identity, not simply a label Windows uses for part of the GPU. That distinction could affect how applications are assigned AI workloads, how the device is measured, and whether developers can make use of it outside built-in Windows features.

For now, practical access appears limited. Conway said standard third-party software could not access the NPU through normal interfaces, and running test networks required Nvidia’s beta development tools. The 55-TOPS figure gives a headline measure of capability, but the investigation does not establish real-world throughput, power efficiency, or how the NPU compares with other accelerators. The reported lack of an INT8 speed benefit also cautions against assuming that its advertised capacity translates into faster performance for common AI tasks.

From DGX Spark to RTX Spark

Conway said he had previously tested Cyberpunk 2077 on Lenovo’s ThinkStation PGX and initially described RTX Spark as the GB10 hardware Nvidia already sells in DGX Spark, running Windows. The NPU investigation complicates that understanding. Nvidia’s block diagram places the DLA/NPU on the SoC die alongside display, camera and audio components, while the GPU sits on a separate die connected through Nvidia’s C2C link.

The report also describes firmware references to a DLA in GB10, even though Linux did not expose such a device in Conway’s examination. That leaves open whether the relevant hardware is absent from GB10, disabled, or simply unavailable through the software he tested. Separately, German publication ComputerBase reportedly obtained Nvidia’s description of the RTX Spark NPU at a San Francisco event as a 55-TOPS accelerator intended for Copilot+ features.

“The NPU runs a full ResNet-50 correctly in INT8, but it isn’t any faster than FP16.”

— Adam Conway, XDA Developers Lead Technical Editor

Performance and GB10 Remain Open

Nvidia has not publicly provided extensive information about the NPU’s architecture, workload support or performance, according to the report. The available tests do not show how it performs across a broad set of models, whether applications beyond the tested networks can use it, or how much power it consumes under sustained workloads. A 55-TOPS rating alone does not answer those questions.

Access is another limitation: the retail driver includes cuDLA, but Conway said it detected one DLA device with hardware version 2 and that only hybrid mode worked. In that mode, tasks must be submitted through a CUDA stream on the GPU; standalone mode returned an “unsupported operation” error. The report’s excerpt does not detail the complete toolchain or establish that ordinary developers can use the accelerator through released, documented software.

It is also unclear whether the DLA referenced in GB10 firmware corresponds to a physically present but disabled block, or whether the firmware contains references for other configurations. The investigation does not resolve whether RTX Spark’s NPU and the GB10-referenced DLA are the same implementation or how they differ.

Documentation Is the Next Test

The next useful step would be public documentation from Nvidia describing the NPU’s supported models, precision modes, software interfaces and measured performance. Updated drivers or development tools could show whether developers can run workloads on it without relying on beta components and hybrid GPU submission.

Further examination of GB10 firmware and system configurations could also clarify whether that platform contains a disabled DLA or only references to one. Until Nvidia addresses those points, the investigation establishes that RTX Spark’s NPU can perform neural-network computation, but its broader role and practical value remain unsettled.

Key Questions

Does RTX Spark have a physical NPU?

XDA’s investigation found separate registers, clocks and power reporting, and the author ran neural networks on the device while checking that the GPU was not doing the work. Those findings support the conclusion that RTX Spark has a distinct hardware accelerator.

How powerful is the RTX Spark NPU?

Nvidia reportedly described it as a 55-TOPS accelerator for Copilot+ features. The report does not provide enough broader benchmark data to establish its real-world performance across applications.

Can developers use the NPU now?

Conway accessed it using Nvidia’s tools, including beta development components, but said normal third-party software could not reach it through standard interfaces. The extent of supported developer access remains unclear.

Does the NPU make INT8 workloads faster?

Not in the reported ResNet-50 test: Conway said the network ran correctly in INT8, but was not faster than in FP16. That result is specific to the test described and does not establish performance for other models.

Is the same NPU present in GB10 systems?

The investigation found DLA references in GB10 firmware, although Linux did not expose a DLA in the system Conway examined. It remains unknown whether the hardware is missing, disabled or inaccessible through the tested software.

Source: XDA Developers