RTX Spark: Everything you need to know

The arrival of RTX Spark has ignited a tech conversation like few others: a mix of curiosity and the potential to change the world of AI, which is currently dominated by the cloud, leaving local execution for very limited cases. It’s not just “another GPU,” but a concept that promises to redefine how we understand AI acceleration in personal and laptop computers.

RTX Spark Equipment Specifications Table

RTX Spark internal architecture

As you know, the RTX Spark is based on a low-power ARM chip for laptops, similar to those found in Apple laptops or those equipped with Qualcomm Snapdragon processors. It’s not an NVIDIA Grace, as some believe. In this case, it’s a design created by NVIDIA and MediaTek called N1X. The computing architecture is quite interesting, featuring:

• Heterogeneous CPU (big.LITTLE): ARM-based and co-designed by MediaTek and NVIDIA, featuring high-performance and efficiency cores, totaling 20 cores (configured in 10+10 clusters, likely Cortex-X925 + Cortex-A725). It uses the ARMv9 ISA and is designed to deliver good performance in general tasks and good efficiency to reduce battery consumption, as it is primarily intended for laptops. It can only process 20 threads in parallel, as it lacks SMT. Furthermore, these cores can operate at frequencies up to 2.9 GHz.
• The NVIDIA Blackwell GPU features a design with 48 Streaming Multiprocessors (SMs), resulting in a total of 6144 CUDA cores, 192 Tensor Cores, and up to 48 Ray Tracing cores. This makes it a superchip with remarkably high graphics performance, similar to that of an RTX 5070. It boasts a base clock speed of 1665 MHz and a Boost clock speed of up to 2418 MHz, with a raw FP4 performance of 1 PTFLOPS. While capable of handling all types of graphics workloads, such as video games, with high performance, it is primarily focused on delivering local AI workloads.
• Memory and I/O: It’s based on up to 128GB of unified LPDDR5X memory, with a 256-bit bus and 16 channels to achieve high bandwidth and prevent bottlenecks. The SoC also integrates 12 PCIe 5.0 lanes and PCIe 4.0 lanes for connecting expansion devices such as NVMe SSDs and other high-speed components, depending on the OEM. Furthermore, it’s important to note that both chips are in the same package, surrounded by SDRAM memory chips for low access latency, and the CPU and GPU are interconnected by a high-performance NVLink C2C bus with a capacity of up to 900 GB/s.
• Thermal profiles and power : The N1X has a TDP of between 45-80W depending on the performance demanded at any given time. Thanks to its design, it can maintain high frequencies sustainably with minimal thermal throttling, even in thin chassis such as those found in laptops.

Performance in LLM models and vision

The combination of Blackwell’s architecture with a massive unified memory system positions RTX Spark laptops as true local AI workstations, eliminating absolute cloud dependency for small to medium-sized models, although it’s not yet possible to work with very large models, as is the case with products like DGX Spark. However, it can locally run LLM models of up to 120B, with context windows of up to 1M tokens .

If you are also interested in training and fine-tuning for AI developers or data scientists, the platform also allows the use of large models up to 200B using quantization techniques, taking advantage of the native FP4 accuracy of the 5th Generation Tensor cores.

Overall performance

The high-end NVIDIA N1X SoC (System on Chip) breaks down the traditional gap between energy efficiency and raw power, unifying the CUDA ecosystem with the Windows environment on ARM. This portable platform boasts 1 PFLOPS of raw AI performance , along with solid performance for running general-purpose software.

With a fairly low TDP range, and in line with what we’ve seen from the competition, it allows for a high-performance, quiet, efficient laptop with powerful graphics, offering an experience running AAA titles at 1440p resolutions that can reach averages of 100 FPS in many cases, thanks to DLSS technology.

For content creation and video editing, the RTX Spark can handle 4K video in real time and smooth editing , even with demanding professional formats like 12K video in 4:2:2 profile . And, 3D creators who need to render complex scenes benefit from over 90GB of dedicated graphics memory, thanks to the absence of VRAM and the use of unified memory.

Of course, unlike the Grace CPU in the DGX Spark, this CPU has been optimized for general software workloads. Likewise, Adobe has completely redesigned key tools like Photoshop and Premiere Pro for this architecture, achieving a 2x performance increase in AI and graphics acceleration compared to x86 software.

Compatibility with frameworks and libraries

The system, which we can enjoy on laptops and mini PCs, currently offers compatibility with Windows on ARM. This means you can use Windows 11 with all that entails, including a vast amount of natively compatible software (ARM builds), but also the ability to run x86 software as if it were native thanks to the emulation layer. Generally, with the Prism emulator, you can use any x86 program or video game you use on a PC, although it’s true that in some cases there may be minor compatibility issues or reduced performance due to the emulation layer.

Performance when running 32-bit and 64-bit x86 software depends heavily on the hardware resources consumed by the emulator and the optimization of the drivers in each case.

On the other hand, the RTX Spark is fully compatible with CUDA 12.1, cuDNN, and OpenCL 3.0, TensorRT, OptiX, DLSS , and a multitude of inference runtimes. Furthermore, it’s important to highlight that, being a system with unified memory for the CPU and GPU, sharing RAM, it allows frameworks that rely on shared access to tensors and buffers to fully utilize its capabilities without having to create copies in separate memories. This is important, for example, for working with PyTorch and TensorFlow.

We also have ONNX Runtime , very useful as a portability layer, and together with TensorRT, it allows converting models trained in different frameworks for optimized execution in the NVIDIA runtime.

As for graphics and multimedia libraries , for those who don’t just want to focus on the AI ​​aspect, you also have compatibility with the technologies you can see in a desktop GeForce RTX, and full compatibility with a multitude of video games, graphics engines like those of Blender, OTOY, VFX pipelines, etc., as well as the usual video codecs for high-resolution editing.

Prism Emulation

Another important detail to keep in mind before you jump into a Windows 11 ARM system is the limitations of the Prism emulator , which allows you to run many x86/x64 apps when native ARM64 binaries don’t exist. For productivity, it’s recommended not to use Prism and to look for native ARM software whenever possible.

However, in most cases it works quite well when emulating only in user mode, and you can configure different profiles for each application from the Windows settings. But, as with Rosetta 2 on macOS, there are also some limitations :

• Virtualization software such as VirtualBox, VMWare, etc., fails due to a lack of compatible kernel modules for ARM and the absence of certain virtualization acceleration instructions or technologies that x86 has but ARM does not.
• For similar reasons, some antivirus programs also fail because they require certain kernel drivers or network filters to function.
• Development tools that use real-time JIT/LLM may also experience slowdowns or errors during execution.

Regarding the performance loss due to the system having to run not only the software itself but also the emulation layer, there is no single universal figure. Practical reports and documentation indicate that the penalty can range from small losses (10–20%) to significant degradation (30–50%) depending on the app, its reliance on AVX/AVX2, its CPU or I/O intensiveness, and whether the app benefits from Prism’s translation cache.

Conclusion

At that time, Apple held a lucrative market niche as the “only” company with high-performance ARM-based chips in its M-Series . Although some ARM-based laptops and mini PCs began to appear, all those chips from Qualcomm, MediaTek, Rockchip, etc., were no match for the M-Series chips in the MacBook and Mac Mini.

As has happened in the past, when Apple launches something, it generates a wave that many others try to ride. In this case, with the arrival of ARM in these types of devices (especially Chromebooks), many also jumped on the bandwagon, but this time, that momentum doesn’t seem to have been enough, despite the improvements in the latest Qualcomm Snapdragon X Elite SoCs. The enormous dependence on the x86 ecosystem that still exists in the PC world is one of the reasons for this slowdown in the widespread adoption of ARM.

Now, many have their eyes on NVIDIA’s RTX Spark as a truly compelling alternative to Apple’s SoCs, viewing these new products as the key to reducing this reliance and finally allowing ARM to succeed. It’s true that it has two strengths that Apple lacks: firstly, it’s compatible with Windows, giving you access to the vast amount of compatible software available, and thousands of games that aren’t available on Mac. This could attract gamers who want an ARM-based system but are hesitant to switch to the Mac platform precisely because of this.

Another advantage of the RTX Spark is its significantly superior performance for running local AI compared to the M-Series, making it practically the only current laptop option for those looking to work with modern models. Even better, it supports CUDA, which is currently unrivaled.

That said, and without diminishing the merits or expectations surrounding the RTX Spark, in my opinion, the number of units sold will be low compared to x86. I think many things still need to change for the hurricane that many predicted would wipe out x86 to materialize…