RTX Spark Leak: 20-Core Model Closes In on M4 Max Performance

RTX Spark Leak: 20-Core Model Closes In on M4 Max Performance

New Geekbench 7 benchmark results have reportedly revealed a substantial performance improvement for NVIDIA’s upcoming RTX Spark, an ARM-based SoC developed in partnership with MediaTek.

The latest leak shows a 20-core CPU configuration reaching a multi-core score of 23,126, putting it just 11% behind Apple’s 14-core M4 Max reference score of 25,957.

The results represent a meaningful improvement over earlier engineering samples, which struggled to compete with Apple’s previous-generation M3 Max in Geekbench 6. The latest data also provides additional information about the CPU and GPU configurations expected across RTX Spark product tiers.

However, the benchmark results come from prototype systems running pre-release software, so they should be treated as an indication of potential rather than final retail performance.

⚡ 20-Core RTX Spark Closes the Multi-Core Gap

#

The latest results reportedly come from the Geekbench 7 database and involve two RTX Spark-powered laptop systems.

Both machines were configured with 64GB of unified memory and tested using a “Balanced” power profile.

The leaked configurations and scores are:

The 20-core RTX Spark therefore achieves approximately 89% of the M4 Max’s multi-core score, despite using six more CPU cores.

That is a considerably stronger result than earlier engineering samples suggested.

Single-Core Performance Remains a Weak Point

#

Single-threaded performance is where Apple’s M4 Max maintains its strongest advantage.

The 20-core RTX Spark recorded a Geekbench 7 single-core score of 2,570, compared with 3,404 for the 14-core M4 Max.

That represents a gap of approximately 24.5% in Apple’s favor.

Single-core performance matters for workloads that cannot efficiently distribute work across multiple CPU threads. Examples can include portions of application startup, lightly threaded software, certain scripting workloads, and legacy applications that have limited parallelism.

As a result, RTX Spark may not match Apple’s responsiveness in every CPU-bound scenario despite its much higher core count.

Multi-Core Performance Is Much More Competitive

#

The situation changes significantly in multi-threaded workloads.

With a score of 23,126, the 20-core RTX Spark trails the M4 Max by only around 11%.

That places the two platforms much closer in heavily parallel workloads such as:

• Video rendering.
• Batch processing.
• Parallel compilation.
• Multi-threaded content creation.
• Large-scale data processing.
• Certain AI and machine-learning workloads.

The result is particularly notable because the RTX Spark is being positioned as an ARM-based PC platform with integrated Blackwell-class GPU capabilities rather than simply as a conventional CPU competitor.

🔢 20-Core vs. 18-Core RTX Spark

#

The leaked data also provides a useful comparison between the two CPU configurations.

The 20-core model scores 2,570 in single-core testing versus 2,541 for the 18-core version, representing only about a 1.14% advantage.

In multi-core testing, the difference is more meaningful:

• 18-core: 21,776.
• 20-core: 23,126.
• 20-core advantage: approximately 6.2%.

The relatively small single-core difference is unsurprising because adding CPU cores does not inherently increase single-thread performance.

The larger multi-core advantage comes from the additional parallel execution resources available to the 20-core configuration.

Which Configuration Matters More?

#

For ordinary interactive workloads, the difference between the two CPUs is unlikely to be obvious.

For sustained multi-threaded workloads, however, the 20-core version has a measurable advantage.

The more important differentiator may ultimately be the GPU configuration rather than the CPU core count, particularly for users purchasing RTX Spark systems for AI, GPU-accelerated applications, content creation, or demanding graphics workloads.

🎮 Blackwell GPU Configurations Emerge

#

The CPU results are only part of the RTX Spark story.

Developer driver information reportedly reveals two different integrated Blackwell GPU configurations:

The flagship model’s 6,144 CUDA cores are particularly notable because that is the same nominal CUDA core count found in several desktop NVIDIA GPU configurations, including the RTX 5070 and RTX 3080.

However, identical CUDA core counts do not imply identical performance.

CUDA core counts should always be evaluated alongside GPU architecture, clock speed, memory bandwidth, power limits, thermal constraints, cache configuration, and workload characteristics.

Why the GPU May Matter More Than CPU Cores

#

The integrated Blackwell GPU is arguably the most important differentiator for RTX Spark.

Unlike conventional integrated graphics solutions that share relatively modest system resources, NVIDIA’s design is intended to bring a much more capable GPU architecture into an ARM-based PC SoC.

This could give RTX Spark systems a significant advantage in workloads such as:

• CUDA applications.
• AI inference.
• GPU-accelerated rendering.
• Video processing.
• Creative applications.
• Local generative AI workloads.
• Compute-intensive scientific applications.

For users whose workloads are heavily GPU accelerated, the jump from 5,120 to 6,144 CUDA cores could therefore be more meaningful than the relatively modest CPU difference between the two SKUs.

🧠 RTX Spark’s Unified Memory Architecture

#

Another important characteristic is the use of unified memory.

The leaked systems reportedly feature 64GB of unified memory, allowing the CPU and GPU to access a shared memory pool.

This architecture can be particularly useful for AI workloads because large models do not necessarily need to be divided strictly between separate CPU and GPU memory pools.

A conventional PC might have:

CPU → System RAM
GPU → Dedicated VRAM

A unified-memory architecture instead resembles:

CPU ─┐
     ├── Shared Unified Memory
GPU ─┘

This can simplify data movement between CPU and GPU workloads and potentially make it easier to run larger AI models locally.

However, unified memory is not automatically equivalent to dedicated high-bandwidth GPU memory. Actual performance depends heavily on memory bandwidth, latency, cache architecture, GPU workload characteristics, and software optimization.

⚠️ CUDA Core Counts Do Not Equal Desktop GPU Performance

#

The 6,144-CUDA-core figure may sound impressive when compared with desktop GPUs, but direct performance comparisons would be misleading.

Laptop SoCs operate within significantly different thermal and power envelopes from desktop graphics cards.

A desktop GPU can sustain considerably higher power consumption and clock speeds because it has a dedicated cooling system and substantially larger thermal budget.

An integrated SoC must balance:

• CPU power consumption.
• GPU power consumption.
• Memory subsystem power.
• Thermal limitations.
• Laptop battery requirements.
• Sustained cooling capacity.

Consequently, an RTX Spark with 6,144 CUDA cores should not be expected to perform like a desktop GPU with the same nominal core count.

Architecture Matters More Than the Raw Count

#

The useful comparison is therefore not:

6,144 CUDA cores = desktop RTX 5070 performance.

Instead, the correct interpretation is:

6,144 CUDA cores indicate that NVIDIA is allocating substantial GPU compute resources to the highest-end RTX Spark configuration.

Actual performance will depend on how those resources are clocked, fed with data, and sustained under real workloads.

💻 RTX Spark Could Become a Major ARM PC Experiment

#

The broader significance of RTX Spark extends beyond benchmark scores.

NVIDIA and MediaTek are attempting to combine:

• ARM CPU architecture.
• NVIDIA Blackwell GPU architecture.
• CUDA software support.
• Unified memory.
• AI acceleration.
• Windows on ARM.

That combination could create a different type of Windows PC platform.

The biggest challenge is not necessarily raw silicon performance. Software compatibility and optimization will be equally important.

Windows on ARM has improved considerably, but the ecosystem still contains applications, drivers, utilities, and games that were originally designed around x86 architectures.

NVIDIA’s CUDA ecosystem could provide RTX Spark with a major advantage in professional and AI workloads, but broader consumer adoption will depend on application support and efficient software translation where native ARM versions are unavailable.

🔬 Prototype Results Require Caution

#

The leaked Geekbench scores should not be treated as final specifications.

The systems appear to be engineering or pre-production hardware, and benchmark performance can change significantly as manufacturers refine:

• Firmware.
• CPU power management.
• GPU drivers.
• Scheduler behavior.
• Memory configuration.
• Thermal policies.
• Application-level optimization.

The reported “Balanced” power profile is another important variable. A production laptop operating under a higher-performance profile could potentially deliver different results, while a thinner or battery-focused system could perform below the leaked figures.

Earlier RTX Spark engineering samples reportedly produced weaker results, demonstrating how much performance can change during the development process.

Final Retail Performance Remains Unknown

#

The most meaningful benchmarks will arrive once commercially available RTX Spark systems can be tested under consistent conditions.

At that point, useful comparisons should include more than Geekbench:

• Sustained CPU performance.
• GPU rasterization.
• Ray tracing.
• CUDA compute.
• AI inference.
• Video encoding and decoding.
• Battery efficiency.
• Thermal throttling.
• Application compatibility.
• x86 emulation performance.
• Memory bandwidth.

These measurements will provide a much clearer picture of whether RTX Spark can compete with Apple Silicon and traditional x86 Windows PCs across different workload categories.

📈 The Performance Picture So Far

#

The current leak suggests a fairly clear hierarchy.

CPU single-core: Apple M4 Max remains substantially ahead.

CPU multi-core: The 20-core RTX Spark is considerably more competitive, trailing by roughly 11% in the leaked Geekbench 7 comparison.

GPU: The flagship RTX Spark reportedly offers 6,144 CUDA cores, giving it potentially significant advantages in CUDA and AI workloads, although raw core count cannot predict final performance.

Memory: 64GB unified memory configurations could make the platform attractive for local AI workloads that require more memory than typical consumer laptops provide.

Software: Still an unknown factor until retail systems and production drivers become available.

🔍 The Bottom Line

#

The latest RTX Spark benchmark leak presents a much more competitive picture than earlier engineering samples.

The 20-core model reportedly reaches 23,126 points in Geekbench 7 multi-core testing, just 11% behind Apple’s 14-core M4 Max reference score. Single-core performance remains weaker, with the RTX Spark trailing by roughly 24.5%.

The GPU configuration may ultimately be more important. The top-end model reportedly combines 20 CPU cores with 6,144 Blackwell CUDA cores, while the lower configuration pairs 18 CPU cores with 5,120 CUDA cores.

That combination could make RTX Spark particularly interesting for AI, CUDA, content creation, and GPU-accelerated workloads.

Still, the most important caveat remains the same: these are leaked prototype results.

Final performance will depend on production silicon, firmware, drivers, thermal limits, power profiles, memory bandwidth, and software optimization. The identical CUDA core count shared with some desktop GPUs should also not be interpreted as equivalent real-world GPU performance.

If these results hold after commercial launch, however, RTX Spark could represent a significant step for Windows on ARM—particularly by bringing NVIDIA’s CUDA and Blackwell ecosystems into a high-performance ARM PC platform.