A chip rated at 120 Platform TOPS sounds far more impressive than one that lists only 50 TOPS for its NPU. Yet when running Windows AI features, the two machines can perform about the same, because most AI tasks use just one processing block rather than all three together. The reason lies in how the number is calculated: Platform TOPS adds up the theoretical throughput of the CPU, GPU and NPU into a single figure, whether or not the three components actually work together on a given task. If you pick an AI laptop based on the biggest Platform TOPS number, you may well be overlooking a far more important spec: the NPU's own performance.
What Stacked TOPS Means on an AI Laptop
TOPS stands for Tera Operations Per Second, a unit that measures how many matrix operations a chip can perform in one second and is commonly used to gauge AI processing power. On a modern laptop, the figure shows up for three separate components: the CPU (general-purpose processing), the integrated GPU (graphics and parallel computing), and the NPU (Neural Processing Unit, a processor built specifically for AI).
Platform TOPS, or stacked TOPS, is the method of adding the theoretical throughput of all three components into one number. Intel pioneered the use of this approach in its product marketing, especially starting with the Lunar Lake generation (Core Ultra 200V). The result is headline figures like "120 Platform TOPS" on promotional material, even though the NPU, the component that actually handles AI in a power-efficient way, contributes only 48 of those 120 TOPS.
| Platform | NPU TOPS | GPU TOPS | CPU TOPS | Platform TOPS | Cinebench R23 multi-core |
|---|---|---|---|---|---|
| Intel Lunar Lake (Core Ultra 200V) | 48 | 67 (Arc 140V) | 5 | 120 | 10,301 |
| AMD Strix Point (Ryzen AI 300) | 50 | 31 (Radeon 890M) | 4 | 80-85 | 21,761 |
| Qualcomm Snapdragon X Elite | 45 | not disclosed | not disclosed | ≈45 (NPU only) | N/A (ARM) |
The core problem: most AI software today runs on only one component at a time. Stable Diffusion uses the GPU. LM Studio uses the NPU or CPU. The Recall feature in Windows Copilot+ uses the NPU. No task actually "stacks" all three at once the way the Platform TOPS figure implies. This is the gap between marketing and real engineering that Hung Phat wants you to understand before you decide.
Breaking Down the Real TOPS of Three Platforms
Each major AI chip platform today splits its TOPS differently across the three components. Intel Lunar Lake (Core Ultra 200V) reaches 120 Platform TOPS, but the internal breakdown is lopsided: the Arc 140V GPU contributes 67 TOPS, the NPU contributes 48 TOPS, and the CPU adds just 5 TOPS. In real benchmarks, the Core Ultra 7 258V scores 1,872 single-core and 10,301 multi-core points in Cinebench R23. Its multi-core score trails AMD's because it has only 8 cores and 8 threads at a 17W TDP, but in exchange it delivers battery life among the longest in its class.
AMD Strix Point (Ryzen AI 300) takes a more balanced approach: the XDNA 2 NPU contributes 50 TOPS, the Radeon 890M GPU adds 31 TOPS, and the CPU adds roughly 4 TOPS, for a total of 80-85 Platform TOPS. The total is lower than Intel's, but the NPU is slightly ahead. More importantly, the Ryzen AI 9 HX 370 scores 2,018 single-core and 21,761 multi-core points in Cinebench R23, nearly double Intel's multi-core performance thanks to its 12-core, 24-thread Zen 5/5c architecture. For compiling code, rendering video or exporting heavy files, that gap is real.
Qualcomm Snapdragon X Elite illustrates the very point of this article. Qualcomm does not publish a combined Platform TOPS figure the way Intel and AMD do; it cites only the 45 TOPS of its Hexagon NPU. In other words, whether the "total" looks big or small depends largely on whether the vendor chooses to add things up, not on any common measurement standard. It remains the only ARM platform in the group, with a 12-core Oryon architecture that scores 3,020 single-core and 15,969 multi-core points in Geekbench 6. The high single-core score keeps single-threaded tasks responsive, but its clearest advantage is still real-world battery efficiency.
One thing to keep in mind when reading the table above: GPU TOPS and NPU TOPS are not equivalent in energy efficiency. The Arc 140V GPU offers 67 TOPS, but on AI workloads it draws roughly 5-10 times more power than a dedicated NPU. An NPU is built to do exactly one job, sustained AI inference, so it consumes only about 1-2W, while a GPU can climb to 15-25W for the same task. That is why Lunar Lake laptops achieve 12-15 hours of real-world battery life even though their GPU has higher TOPS than AMD's.
Can You Really Trust the Platform TOPS Figure
The answer is both, depending on how you use your laptop. The NPU threshold of 40 TOPS and above is substantive: it is the minimum needed to enable offline Windows Copilot+ PC features such as Recall, Live Captions and Auto Super Resolution without a network connection. All three platforms, Intel Lunar Lake (48 TOPS NPU), AMD Strix Point (50 TOPS NPU) and Snapdragon X Elite (45 TOPS NPU), clear this bar. That is where the real value lies.
The less substantive part is the combined Platform TOPS total. When you use LM Studio to run a local language model, or AI noise-cancellation software in a meeting, the system uses only the NPU. Whether the stacked TOPS total is high or low has no effect on the speed or quality of that task. For reference, the RTX 4090 Laptop GPU, a discrete graphics card, reaches 686 TOPS INT8 with 9,728 shaders. That figure is not part of the Platform TOPS of any thin-and-light laptop, and when you need real AI power for rendering or model training, the dGPU is what decides it.
Hung Phat often hears customers ask, "Is a 120 TOPS machine better than an 85 TOPS machine?" and the answer is almost always: it depends on what you do. If you travel a lot, take online meetings that need real-time AI noise cancellation, or want to try power-efficient local LLMs, a sufficiently strong NPU (above 40 TOPS) is what you need, not a high Platform TOPS total. If you only do basic office work or play demanding games on a dGPU, the TOPS figure has almost no effect on your everyday experience. To pick the right threshold for your actual needs, you can also read 13 TOPS vs 45 TOPS NPU: Which AI Tasks Can Run Locally? or Laptop NPUs in 2026: 13, 40 or 50 TOPS - How Much Is Enough?
If you want a machine that already meets the Copilot+ PC standard and is in stock at Hung Phat, three representative options span the three platforms: the HP OmniBook X 14 inch AI (2024) runs a Snapdragon X Elite (45 TOPS NPU) and costs 25.7 million VND; the HP OmniBook X Flip Next Gen AI 16 (2025) offers a Ryzen AI 7 350 (50 TOPS NPU) or Core Ultra 5 226V option, at around 20-26 million VND; and the HP OmniBook X Flip Next Gen AI 14 (2025) comes with a Core Ultra 7 256V (Lunar Lake, NPU of about 47 TOPS) or Ryzen AI 5 340, at around 19.5-24.9 million VND. All three clear the 40 TOPS NPU threshold and so fully enable offline Copilot+ features; the main differences are the chip platform and price range. If you are still torn between upgrading the NPU and upgrading the RAM, the following article adds a useful perspective: Is a 45 TOPS NPU Really Worth the Money, or Should 32GB of RAM Be the Priority in 2026?
Frequently Asked Questions (FAQ)
How does stacked TOPS differ from the way TOPS was counted on earlier chip generations?
Before Lunar Lake and Strix Point, first-generation AI laptops such as Intel Raptor Lake Refresh or AMD Hawk Point had only a very weak NPU (under 16 TOPS) or no dedicated NPU at all, so every AI task fell to the CPU or GPU. Nobody stacked TOPS back then because the NPU was too weak to be worth mentioning. From 2026, with NPUs of 40-50 TOPS becoming the Copilot+ standard, vendors began adding up the three components to produce a more impressive total. The real-world difference is that new-generation NPUs can run mid-sized INT8 models (7-13 billion parameters) entirely locally, while older chips cannot, even with a more powerful GPU.
Will a laptop without a strong enough NPU become obsolete early?
Not necessarily, but it helps to be realistic. If you use a graphics workstation or gaming machine with a dedicated dGPU, where the RTX 4090 Laptop GPU reaches 686 TOPS INT8, the discrete card already handles AI more powerfully than any integrated NPU. A laptop without an NPU is limited only in offline Windows Copilot+ features: Recall, Live Captions and Auto SR will either not run or will run through the GPU with noticeable battery drain. For users who only do office work, watch movies and study online, lacking an NPU will not affect everyday experience over the next 2-3 years. Copilot+ features also depend on how deeply Microsoft integrates them into Windows, which is a software factor and less certain than hardware specs.
How much extra do I have to pay for an AI laptop with a qualifying NPU?
The 40 TOPS NPU threshold (the Copilot+ standard) currently appears in laptops from about 20 million VND and up. For example, the HP OmniBook X 14 AI with Snapdragon X Elite costs about 25.7 million VND, and the HP OmniBook X Flip 14 starts at around 19.5 million at Hung Phat. Compared with previous-generation Intel (Raptor Lake Refresh) or AMD Hawk Point laptops in the same segment, you pay roughly 5-10 million VND more for a sufficiently strong NPU. That difference also covers upgrades to the CPU architecture and integrated GPU, not just the NPU. If your budget is under 18 million VND, the market currently has few worthwhile options that meet the Copilot+ PC standard.
Which laptops at Hung Phat have an NPU that meets the Copilot+ PC standard?
The Copilot+ machines currently in stock at Hung Phat are mainly in the HP OmniBook X line: the HP OmniBook X 14 AI (Snapdragon X Elite, 45 TOPS NPU, about 25.7 million VND) suits frequent travelers thanks to its battery-efficient ARM platform; the HP OmniBook X Flip 16 (Ryzen AI 7 350 option with a 50 TOPS NPU, about 20-26 million) is strong in multi-core work; and the HP OmniBook X Flip 14 (Core Ultra 7 256V Lunar Lake, NPU of about 47 TOPS, about 19.5-24.9 million) balances battery life and performance. All three can enable Recall, Live Captions and Auto SR offline; choosing among them mostly comes down to your preferred chip platform, battery life and budget. If you need advice on a specific configuration, you can contact Hung Phat for help choosing the right machine for your needs.
