Nvidia RTX Spark Analysis - Finally an answer to Apple's M5 Pro & M5 Max?

Nvidia uses a custom-built MediaTek CPU with up to 20 cores in four clusters. You get up to ten performance cores (ARM Cortex X925 up to 4.1 GHz) and ten efficiency cores (ARM Cortex A725 up to 2.8 GHz). It is manufactured at TMSC in a 3 nm process. The CPU tile also includes the memory interface for up to 128 GB LPDDR5x-9400 RAM, display connectors (HDMI 2.1a & DisplayPort 2.0) as well as high-speed I/O connectivity (4x USB 4.0, PCIe 5.0, Wi-Fi 7).

The GPU part is a familiar Blackwell GPU (GB20B) with either 5120 or 6144 CUDA cores running at up to 2535 MHz. Since the chip generation is identical to current consumer GPUs like the GeForce RTX 5070, the N1X GPU can also take advantage of all current software optimizations like DLSS5 or Frame Generation. 

The maximum memory bandwidth is 300 GB/s, which is comparable to Apple's M5 Pro SoC. Both the CPU and GPU tiles are connected via NVLink with a bandwidth of 600 GB/s. In addition to the N1X with 20 CPU cores and 6144 GPU cores, there is also a smaller N1X with 18 CPU cores (9x performance + 9x efficiency) as well as 5120 GPU cores. Since the name is identical, you have to pay attention to the exact specs when you purchase a laptop with RTX Spark. Please see the full technical specifications of the two N1X versions in the table below.

In order to make a meaningful comparison between the different processors and graphics cards, we look at the power consumption in addition to the pure performance in synthetic benchmarks from which we then determine the efficiency.

The consumption measurements are always carried out on an external display so as to eliminate the different internal displays as influencing factors. Nevertheless, we measure the overall consumption of the system here and not just compare the pure TDP values.

We start with the single-core performance and the N1X performs pretty well. It is faster than AMD's Zen 5 chips and on par with Intel's latest Panther Lake chips. Qualcomm's new X2 chips are a bit faster and Apple is obviously on a completely different level in terms of single-core performance.

During single-core workloads, the N1X CPU core consumes roughly 9 Watts with a total system power draw of ~30 Watts. The efficiency is solid only slightly behind a Panther Lake Core Ultra X chip in the Asus Zenbook Duo, while it is better compared to all the AMD rivals. Snapdragon X2 chips and Apple M5 chips are superior here, though.

Multi-core performance is very good and the RTX Spark N1X can keep up with the fastest AMD Strix Halo systems. Intel's Panther Lake is easily beaten and most Snapdragon X2 Elite chips as well, only the rare Snapdragon X2 Elite Extreme is slightly faster. Compared to Apple's processors, the N1X can just beat the latest M6 chip, but falls behind the M4/M5 Pro & Max models.

The N1X consumes up to 85 Watts and then settles at 70 Watts in the Performance mode of the Asus ProArt P16. The multi-core efficiency is roughly on par with the Strix Halo systems. Intel's Panther Lake systems are better (but are also much slower) and the second generation of Snapdragon processors is also superior, just like Apple's M4/M5 chips. We will add efficiency data for the slower power modes of the ProArt P16 soon.

As we mentioned before, the Blackwell GPU is basically what you get in consumer GeForce cards. Keep in mind that some benchmarks like Time Spy have to be emulated, which will affect the performance. We therefore limited our benchmark selection to native ARM tests like 3DMark Steel Nomad. Here, the GPU performance is comparable to powerful versions of the Nvidia GeForce RTX 5070 and around 50 % faster than the Radeon 8060S. Intel and Qualcomm cannot compete in this regard. Apple's M5 Pro GPU is also slower (-28 %), while the M5 Max GPU is faster (+27 %).

Gaming is obviously another important aspect with a Blackwell GPU and this is also very interesting considering the ARM version of Windows. From our experiences with Snapdragon laptops over the last two years, we know that some titles are not compatible. This includes titles like F1 25 or Battlefield 6, for example, even though we already saw a working version of F1 25 at Gamescom, so this might change soon. Then there is also the performance aspect, since the vast majority of titles have to be emulated right now. We had the chance to test the Marvel Rivals in a native version and the performance advantage was roughly 20 % in 1080p and with much higher 1% low results (9 vs. 29 fps), so this is a significant difference. We also faced issues with stutters in titles like Assassin's Creed BlackF Flag or Shadows. We are currently still testing and update this article soon.

For this comparison, we only included Cyberpunk 2077 as well as Baldur's Gate 3 and the emulation definitely takes a toll. Compared to the synthetic benchmarks, the N1X loses ground to the RTX 5070 and is closer to the RTX 5060 in Cyberpunk, for example, while being competitive in Baldur's Gate 3. The larger memory is obviously an advantage in demanding games and we expect there will be native ARM version for the majority of upcoming titles. As of right now, gaming is definitely possible, especially with features like DLSS and Frame Generation, but laptops with powerful GeForce GPUs are (still) superior for gaming.

We also checked the efficiency while gaming, but once again keep in mind that Cyberpunk is the x64 version (there is no native ARM version for Windows yet, only for macOS), which affects the performance and therefore also the efficiency. The numbers are still solid and the N1X is just slightly worse compared to systems with the GeForce RTX 5060/5070. It beats the AMD Strix Halo systems and also the MacBook Pro 16 with the M5 Pro. 

Local AI is the major talking point of all chip manufacturers. But what does local AI mean and how can you use it? In simple terms, if you have lots of memory, you can download models and run the locally on your machine without using cloud services and obviously without paying for them. There are obviously different models and what you do will be completely unique from user to user. Once you have a model running, you can pretty much use it as a chat bot like you would with ChatGPT or Gemini, but agents take AI workloads one step further. We used Hermes and we used it to create slides for our YouTube video review of RTX Spark, for example, which can also be seen in the next section (Creator Benchmarks). Instead of simply answering your questions, the agent can get information from different sources and create documents. Agents can do much more and this is just a simple example, but even this approach already improved our workflow. 

The big problem in evaluating the performance is that the evolution of AI models is rapidly progressing and creating a database is problematic. We are also still trying to find the best benchmarks and right now, we use Mlperf 2.0 as the main comparison. Geekbench AI on the other hand is not really suitable right now as you can only run the CPU test on the RTX Spark N1X. Our Mlperf results show a mixed picture, and the N1X trades places with the M5 Pro depending on the individual test. We will continue testing the AI performance and update this article once we have additional results for you. 

We also performed some creator benchmarks. In Adobe Photoshop, the new N1X performs well and is on par with AMD's Strix Halo SoC, but the M5 Pro is clearly ahead thanks to the powerful CPU. There are rumors about an update for Photoshop, which is supposed to take more advantage of GPU acceleration, so we are curious to see the performance as soon as it will be released. The Resolve benchmarks are a bit disappointing, but considering the benchmark does not need much VRAM and the software loves fast GPUs and is heavily optimized for macOS, the results make more sense. However, once you work with a lot of RAW footage and fill up the VRAM quickly, the N1X definitely has the edge over GPUs like the RTX 5070.