They look similar on a desk, but they are built for different jobs. An ASUS NUC 16 Pro is an x86 Windows PC with an Intel NPU and Arc graphics, best at everyday office work plus light, local AI. An NVIDIA RTX Spark machine pairs an Arm CPU with an RTX GPU and up to 128GB of shared memory under Windows on Arm, so it can run much larger AI models while still behaving like a Windows PC. An NVIDIA GB10 system such as the NVIDIA DGX Spark or ASUS Ascent GX10 is a Linux AI workstation with 128GB of unified memory and NVIDIA's full data-centre software stack, aimed at developers running, fine-tuning and serving models. At SourceIT the NUC 16 Pro X9 388H with 32GB and 1TB is S$2,775.81, RTX Spark is sold in laptops from S$4,974.08, and GB10 systems with 128GB start at S$7,955 (listed 3 Nov 2026).
How are the three built differently?
| ASUS NUC 16 Pro (Core Ultra X9 388H) | NVIDIA RTX Spark (N1X) | NVIDIA GB10 (DGX Spark, GX10) | |
|---|---|---|---|
| CPU | Intel x86, Panther Lake | Arm, 20 cores | Arm (NVIDIA Grace) |
| AI hardware | NPU 50 TOPS + Arc B390 GPU 122 TOPS (INT8) | Blackwell RTX GPU, 6,144 CUDA cores | Blackwell GPU; NVIDIA quotes 1 petaFLOP FP4 with sparsity |
| Memory | 32GB or 64GB soldered LPDDR5x at SourceIT (up to 96GB per ASUS) | Up to 128GB shared | 128GB unified |
| Memory bandwidth | About 136.5 GB/s theoretical (LPDDR5x-8533, our calculation) | About 300 GB/s (reviewers' calculation; NVIDIA has not published it) | 273 GB/s (NVIDIA) |
| Operating system | Windows 11 (x86) | Windows 11 on Arm | NVIDIA DGX OS (Ubuntu Linux) |
| Main AI software | OpenVINO, Windows ML, PyTorch XPU, Ollama via Vulkan | CUDA on Windows on Arm (Ollama, llama.cpp); some Linux-only tools need WSL | CUDA, TensorRT-LLM, vLLM, NVIDIA playbooks |
| Form at SourceIT | Mini PC | Laptops | Mini desktop |
| SourceIT price | S$2,775.81 (32GB/1TB); S$4,281.65 (64GB/1TB) | From S$4,974.08 (32GB); S$10,773.20 (128GB, ASUS ProArt P16) | From S$7,955 (128GB/4TB); NVIDIA DGX Spark S$9,300 |
Intel and ASUS specifications for the NUC 16 Pro; NVIDIA for GB10; ServeTheHome and StorageReview (Oct 2026) for RTX Spark bandwidth; MSI's EdgeMesa N AI+ page for the N1X core counts. The NUC bandwidth is our arithmetic from the memory speed and is not an Intel figure. SourceIT prices listed 3 Nov 2026. We have not tested these systems ourselves.
What happens when you give them real work?
The fairest comparison is by task rather than by headline TOPS, because the published numbers come from different tools, settings and quantisation. Treat these as indications, not a head-to-head test.
- Office apps, line-of-business software and Windows tools. The NUC runs everything a normal x86 Windows PC runs. RTX Spark runs most Windows apps, with x64 programs translated by Microsoft's Prism emulation, and some games such as F1 25 and Battlefield 6 were reported incompatible. GB10 runs Linux only, so it is not an office PC.
- A small chatbot, around 8 billion parameters. StorageReview measured 26.18 tokens per second for Llama 3 on the NUC 16 Pro's GPU using UL Procyon on OpenVINO. On an RTX Spark Surface Laptop Ultra, StorageReview measured 56.4 tokens per second for Llama 3.1 8B in MLPerf. Ollama's own test of a GB10 gave 38.0 tokens per second for Llama 3.1 8B at q4_K_M. All three are usable for one person.
- A large model such as gpt-oss-120b. Its GGUF file is about 59 GiB, so it does not fit the NUC's 32GB or 64GB. With 128GB, Techtesters measured 63.3 tokens per second on an MSI Prestige N16 Flip AI+ (tool and settings not published), and NVIDIA measured 55.4 tokens per second on GB10 with llama.cpp. Memory size, not TOPS, decides what you can run.
- Serving many users. GB10 runs vLLM and SGLang natively. StorageReview measured gpt-oss-120b on a Dell Pro Max GB10 scaling to about 671 tokens per second in total at batch 64. vLLM does not support Windows natively, so RTX Spark machines need WSL for it.
Choose by the model you need to run and the operating system you need to live in. The TOPS number on the box will not tell you either.
Why are the TOPS figures so far apart?
Intel's figures are INT8. NVIDIA's GB10 petaFLOP is FP4 with sparsity, a lower-precision format counted with a 2:4 sparsity feature that doubles the figure. A community measurement on NVIDIA's developer forum put GB10 at about 511 TFLOPS dense FP4 against about 1,014 sparse. Neither can be divided by the other to say how much faster one machine is. We explain this in Intel NPU vs NVIDIA GPU: Are AI TOPS Becoming the Most Misleading PC Specification?
Which one fits which buyer?
- NUC 16 Pro: office desktops, reception and meeting-room PCs, digital signage and staff who want Copilot+ features and small local models. ASUS markets the NUC 16 Pro as a Copilot+ PC.
- RTX Spark: one person who wants a portable Windows machine that can still load 70B to 120B-class models locally, with CUDA for creative and AI apps.
- GB10: developers and small teams prototyping, fine-tuning or serving models, and anyone who wants NVIDIA's Linux AI stack on a desk.
For a closer look at the two NVIDIA machines, see NVIDIA DGX Spark vs RTX Spark: Same AI Ambitions, Two Very Different Computers and NVIDIA RTX Spark vs GB10: Is Windows Finally Ready to Challenge Linux for Local AI? What the 120B and 200B model claims mean is covered in RTX Spark vs DGX Spark: What NVIDIA's 120B and 200B AI Model Claims Really Mean. To see what the NUC can do with local AI in practice, read Can a S$2,000 ASUS NUC Run the Same AI Apps as a S$7,000 GB10?
Buying in Singapore
SourceIT supplies the ASUS NUC 16 Pro X9 388H, 32GB, 1TB, Windows 11 Pro (S$2,775.81) and 64GB version (S$4,281.65); RTX Spark laptops including the MSI Prestige N16 Flip AI+ 32GB (S$4,974.08), Lenovo Yoga Pro 9n 64GB (S$5,897.20) and ASUS ProArt P16 128GB (S$10,773.20); and GB10 systems including the NVIDIA DGX Spark 4TB (S$9,300) and ASUS Ascent GX10 128GB/1TB (S$10,350). Prices listed 3 Nov 2026. Browse the ASUS NUC, RTX Spark and DGX Spark collections. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is the main difference between an ASUS NUC, RTX Spark and GB10? The NUC is an x86 Windows PC with an Intel NPU and Arc graphics; RTX Spark is an Arm Windows machine with an RTX GPU and up to 128GB shared memory; GB10 is a Linux AI workstation with 128GB unified memory and NVIDIA's data-centre software stack.
Can an ASUS NUC run large AI models like gpt-oss-120b? Not the configurations SourceIT lists. The model's file is about 59 GiB, more than the NUC 16 Pro's 32GB or 64GB. RTX Spark and GB10 machines with 128GB can run it.
Which runs Windows apps best? The NUC, because it is a standard x86 Windows PC. RTX Spark runs Windows on Arm with emulation for x64 apps, and GB10 runs Linux only.
Are Intel and NVIDIA TOPS figures comparable? No. Intel quotes INT8, while NVIDIA's GB10 figure is FP4 with sparsity. Compare by memory size, bandwidth and software support instead.
How much do they cost at SourceIT? On 3 Nov 2026, the NUC 16 Pro X9 388H 32GB/1TB was S$2,775.81, RTX Spark laptops started at S$4,974.08, and GB10 systems with 128GB started at S$7,955.
