A test configuration combining an 18-core CPU and a 5,120-core GPU has appeared for NVIDIA's Windows SoC "RTX Spark." In an entry dated August 7, 2026, Geekbench 7 listed this previously undisclosed reduced configuration alongside a 20-core/6,144-core configuration close to the official specifications. Both entries are no longer viewable, but the posted result pages recorded the core configuration, clock speed, and even 64GB of memory. Before any ranking in the performance race could be determined, what emerged was the possibility that NVIDIA is testing an RTX Spark configuration with both CPU and GPU scaled down.
What the 18-core, 5,120 GPU-core configuration suggests about a two-tier lineup
The 20-core entry was identified as "NVIDIA RTX Spark N1X (6144-core GPU, 20-core CPU)," with the CPU split into two clusters of 10 cores each, and a base frequency of 4.00GHz. A test unit running Windows 11 Pro executed Geekbench 7.0.0, recording a single-core score of 2,570 and a multi-core score of 23,126. The power plan was set to Balanced, with 64GB of memory.
The other was identified as "NVIDIA RTX Spark N1X (5120-core GPU, 18-core CPU)." Its CPU clusters were split into 10 and 8 cores, with a base frequency of 3.90GHz. Running on Windows 11 Enterprise, it scored 2,541 single-core and 21,776 multi-core. This unit also used the Balanced setting and 64GB of memory.
The official specifications for RTX Spark, announced by NVIDIA on May 31, list a 20-core Grace CPU, a 6,144 CUDA-core Blackwell RTX GPU, and up to 128GB of unified memory. The product page describes the CPU as having "up to 20 cores" and memory as "up to 128GB." The 18-core version has not been announced, but if the identifier in the benchmark is accurate, it represents a configuration with 10% fewer CPU cores and 16.7% fewer GPU cores. In other words, NVIDIA may be testing a lower-tier configuration with both CPU and GPU scaled back.
However, the identifier is also information passed to Geekbench from firmware or software. NVIDIA has not disclosed whether it will bring the 18-core version to market, nor under what name or price it might sell. Power consumption is also unknown. At this stage, all that can be confirmed is that at least one Windows test unit reported this configuration name.
Reducing the 20-core by 10% still results in only a 5.8% drop in multi-core score
In a comparison between the two entries, the 18-core version's single-core performance was 1.1% lower than the 20-core version. Given that the base frequency difference was 2.5%, it is not unusual for scores to be close when the core configuration handling single-threaded work does not change significantly. On the other hand, the multi-core score was 5.8% lower. This gap is smaller than the 10% reduction in CPU core count, indicating that—at least in this particular run—the 18-core version retained parallel performance reasonably well.
There is also a notable error to flag in Tom's Hardware's article. The body text described the 18-core version's score of 21,776 as "about 6% faster" than the 20-core version, but the table in the same article listed it as -5.84%. Calculating the difference against 23,126 confirms that the 18-core version was in fact 5.84% slower. Before debating which configuration is superior in the shipping product, one must first read the direction of the comparison correctly.
That said, the two units were not tested under identical conditions. The provisional OEM names differ, as do the Windows editions. The cooling capacity of the chassis, the CPU's power limits, and background processes are all unknown. The impact of the reduced GPU core count cannot be judged from CPU test scores alone, and AI inference, content creation, and gaming performance will require separate measurements.
Why a single score cannot determine market ranking
Geekbench 7 was released just recently, on July 23, with updated workloads for video/audio processing and game physics. The multi-core test has also changed, now running only the processes that are actually parallelized in real applications as multi-threaded tasks. Therefore, any score gap against RTX Spark test units measured with the older Geekbench 6 cannot simply be attributed to driver improvements.
The Geekbench Browser's processor chart aggregates user-submitted results and only lists CPUs with at least five unique results. In this case, each RTX Spark configuration has only a single entry, and the result pages currently return 404 errors. Any comparison suggesting that the 20-core version outperformed a specific Intel or AMD mobile CPU may hold true for that particular run, but it does not guarantee average performance or reproducibility.
What can be read from a single test-unit benchmark is a configuration difference, not a ranking of shipping performance. While it serves as strong evidence for considering the existence of an 18-core version, connecting it to purchasing decisions for mass-produced units requires the same Geekbench 7, the same power conditions, and multiple measurements.
Three conditions to verify with fall products
NVIDIA has stated that laptops and small desktops equipped with RTX Spark will launch in fall 2026. ASUS, Dell, and HP will offer initial products. Lenovo, Microsoft Surface, and MSI are expected to join around the same time, followed by Acer and GIGABYTE. While the official announcement emphasizes the 20-core CPU and up to 128GB of memory, both units in this case had only 64GB. Memory capacity, too, may become part of the product tiering.
For mass-produced units, the first thing to confirm is whether the 18-core/5,120 GPU-core configuration will actually remain as a real SKU. Next, if the price difference and power envelope relative to the 20-core version become clear, the meaning of the 5.8% multi-core gap can be properly assessed. Finally, repeated measurements of CPU, GPU, and AI processing on identical chassis would help isolate exactly where performance differences emerge. Only once these three conditions are met can we judge whether the 18-core version genuinely expands the range of performance and price options.