Every CPU With an NPU: TOPS Ranking Table

Four vendors currently ship a CPU with an on-die NPU — Intel (Core Ultra), AMD (Ryzen AI), Qualcomm (Snapdragon X) and Apple (M-series) — and the highest vendor-published rating today is Qualcomm's Snapdragon X2 Elite at up to 85 TOPS on the X2E-96-100, X2E-90-100 and X2E-84-100 parts. The full ranking is in the table below, one row per chip family, with the vendor that published each number and the precision it was quoted at. Read that precision column before you compare two rows: Intel labels its figure INT8, AMD quotes INT8, Qualcomm's own product pages state no precision at all, and Apple has published no TOPS figure whatsoever for the M5.

One thing to settle before the table, because it changes how you should read every row.

TOPS is a peak arithmetic rating for a block of silicon. It is the number of integer multiply-accumulate operations the NPU's array could retire per second if it were fed perfectly and never stalled. It is a ceiling derived from the design, not a measurement of anything running. It does not account for memory bandwidth, it does not account for whether any software on your machine can reach the NPU at all, and it is not comparable between vendors who quote it at different numeric precisions. A 50-TOPS NPU with no runtime targeting it produces exactly zero tokens per second.

TOPS became a shopping spec for one reason: Microsoft drew a line at 40. Everything else about the number is marketing convention. So use this page the way it is built — as a lookup table for "which chip is this and does it clear the bar", not as a performance ranking.

Which CPUs Have an NPU, and How Many TOPS?

Sorted by vendor-published NPU TOPS, highest first. The TOPS column is the NPU block alone — not the "platform total" that adds the GPU and CPU, which is usually the bigger number printed on the box (see the two-numbers problem below).

Where each number comes from: Qualcomm's per-SKU tables on its Snapdragon X2 Elite, Snapdragon X Elite, Snapdragon X Plus and Snapdragon X product pages. Intel's per-SKU "NPU Peak TOPS (INT8)" field on Intel ARK. AMD's Ryzen AI product pages, with the supported-silicon list cross-checked against the Ryzen AI Software release notes. Apple's own launch newsrooms for M1, M2, M1 Ultra, M2 Ultra, M4 and M5. Nothing in this table was measured by this site — these are the vendors' own peak ratings, reproduced with the labels they came with.

Three caveats that matter more than they look:

• Intel's Lunar Lake numbers vary by SKU inside the same family. 48, 47 and 40 are all Core Ultra 200V. Check the exact part number on ARK rather than assuming the flagship figure — a laptop advertised as "Core Ultra 200V" tells you nothing about which of the three you are getting.
• Panther Lake's "up to 50" is the top-tier figure. Lower Core Ultra 3 SKUs ship smaller GPU and AI configurations.
• Qualcomm's "Extreme" branding maps to specific part numbers. On Qualcomm's own X2 Elite table the 85-TOPS rating belongs to X2E-96-100, X2E-90-100 and X2E-84-100; the other four X2E parts are rated 80. A Snapdragon X2 Plus tier also exists — check its Qualcomm page for its own rating rather than assuming it inherits either number.

My Laptop Says 47 TOPS — Which Chip Is That?

Reverse lookup. Find the number you were shown, read across.

The two rows worth staring at are 100+ and 34. Both are platform totals that get quoted as if they were NPU ratings, and both are where most of the confusion in this topic comes from.

Why Does the Same Laptop Show Two Different TOPS Numbers?

Because vendors publish two different things and reviewers mix them freely.

The NPU figure is the neural block on its own. The platform figure adds the integrated GPU and the CPU's vector units to the same total. Intel makes this unusually explicit: Intel ARK exposes separate per-SKU fields for NPU Peak TOPS (INT8), GPU Peak TOPS (INT8), CPU Peak TOPS (INT8) and a Total Peak TOPS that is simply the three added up. For a Core Ultra 7 155H those come out as an 11-TOPS NPU inside a 34-TOPS platform — which is why "Meteor Lake, 34 TOPS" and "Meteor Lake, 11 TOPS" are both true statements about the same chip.

This matters for exactly one practical decision: Copilot+ eligibility is judged on the NPU alone. A machine with a 34-TOPS platform total and an 11-TOPS NPU is not a Copilot+ PC, no matter how the sticker reads.

Generational platform totals climb fast for a reason that has nothing to do with the NPU — the integrated GPU is doing the heavy lifting in that sum. That is also the honest explanation for why platform TOPS roughly doubled between Intel's Lunar Lake and Panther Lake generations while the NPU itself moved from 48 to about 50.

Which Chips Clear Microsoft's 40-TOPS Copilot+ Bar?

Microsoft's own developer documentation is unambiguous about the threshold. The Copilot+ PCs developer guide describes a Copilot+ PC as Windows 11 hardware "powered by a high-performance Neural Processing Unit (NPU) ... that can perform more than 40 trillion operations per second (TOPS)", and states that "Many of the new Windows AI features require an NPU with the ability to run at 40+ TOPS". Microsoft's consumer-facing requirements add 16GB of RAM and 256GB of storage.

Clears the bar: every Snapdragon X and X2 part, every Core Ultra 200V part, Panther Lake, and every AMD Ryzen AI 300, AI PRO 300, AI Max and AI 400 part. Note the one wording wrinkle: Microsoft says "more than 40", while Intel's Core Ultra 5 200V SKUs are rated at exactly 40 and still ship in machines sold as Copilot+ PCs — so in practice the bar reads as 40 and above.

Does not clear it: Intel Meteor Lake (11), Intel Arrow Lake (13), AMD Phoenix (10) and AMD Hawk Point / Ryzen 8000G (16). These are all real NPUs and all of them run Windows Studio Effects-class workloads; they simply predate the line Microsoft drew.

Not applicable: every Apple M-series chip. The Neural Engine is a genuine NPU, but Copilot+ is a Windows program and Apple's silicon is not evaluated against it. If you are weighing a Copilot+ machine against a discrete GPU for actual local inference, our Copilot+ PC vs RTX comparison for local AI is the honest version of that trade-off.

The practical consequence of every current Windows part clearing 40: TOPS has stopped being a differentiator. When the whole shelf qualifies, the badge tells you nothing, and screen, battery, memory capacity and memory bandwidth are what is left to choose on.

Are These TOPS Numbers Comparable Across Vendors?

No, and the precision column is where that breaks. An "operation" in a TOPS figure is whatever the vendor decided to count.

• Intel states its precision in the spec field itself: ARK's label is literally "NPU Peak TOPS (INT8)". You always know what you are getting.
• AMD quotes XDNA 2 at INT8. The same block also supports Block FP16, so AMD marketing occasionally quotes throughput figures at two precisions for one chip.
• Qualcomm publishes "45 TOPS" and "80 TOPS" on its product pages with no precision qualification anywhere on them. Comparing a Qualcomm number to an Intel INT8 number is an assumption, not an equivalence.
• Apple says "trillion operations per second" and has never attached a precision to it in a launch newsroom.

The scale of the problem is easiest to see on GPUs, where the conventions are documented. NVIDIA's Ampere architecture whitepaper quotes the GeForce RTX 3090's tensor throughput at 284 INT8 TOPS and 568 INT4 TOPS, and doubles both again for structured sparsity. Same silicon, four legitimate numbers, a 4x spread between the smallest and the largest. Nothing stops an NPU vendor from quoting the INT4 or sparse figure and letting you assume it was INT8.

The rule for reading any TOPS claim: an unlabelled number is not a specification, it is a headline. If the precision is not stated, treat the figure as usable only for comparing chips from the same vendor.

Why Has Apple Stopped Publishing a TOPS Figure?

Because Apple moved the AI compute somewhere the Neural Engine number would not describe.

Apple published a Neural Engine figure for every generation from M1 through M4: 11 trillion operations per second on M1, 15.8 on M2 ("over 40 percent more than M1", in Apple's wording), 31.6 on M2 Ultra, and 38 on M4 from a 16-core design. Then it stopped. The M5 announcement describes "a faster 16-core Neural Engine" with no number at all, and puts the emphasis elsewhere: "The 10-core GPU features a dedicated Neural Accelerator in each core, delivering over 4x peak GPU compute compared to M4."

So on M5 the AI silicon is distributed across the GPU cores, and a single Neural Engine TOPS rating would understate the chip. Any "Apple M5 = X TOPS" figure you find is somebody's estimate, not an Apple specification.

The M3 generation is a second, quieter gap: Apple's M3 launch said only that "The Neural Engine is up to 60 percent faster than in the M1 family of chips" and never published a TOPS number. The widely-circulated "M3 = 18 TOPS" is arithmetic on that claim — 11 x 1.6 = 17.6 — not something Apple ever printed. We have marked both rows "Not published" rather than quietly filling them in.

If you are sizing an Apple machine for local models, the numbers that actually decide it are unified memory capacity and bandwidth, not the Neural Engine — our Apple M5 guide for local AI works through those.

Which Processors Have No NPU At All?

Worth knowing, because "has an NPU" is not a property of a brand, it is a property of a specific die.

• Intel Core 14th generation and earlier — no NPU on any desktop or laptop part. The NPU arrives with Core Ultra.
• Intel Core Ultra 200S desktop — has one, rated 13 TOPS. So desktop x86 is not NPU-free, it is just far below the Copilot+ bar.
• AMD Ryzen 9000 desktop (Granite Ridge) — no NPU. The AM5 chips with an NPU are the 8000G APUs.
• AMD Ryzen 8500G and 8300G — no NPU, while the 8700G and 8600G in the same product line have one. Same series, same socket, different silicon.
• AMD Ryzen 7045 "Dragon Range" laptop parts — desktop dies in a laptop package, no XDNA block.
• Apple — every M-series chip since M1 has a Neural Engine, and so does every A-series chip. There is no Apple-silicon Mac without one.

AMD's own Ryzen AI Software release notes are the most reliable list of which AMD silicon the NPU toolchain actually supports: as of the 1.8 release it names "AMD processors codenamed Phoenix, Hawk Point, Strix, Strix Halo, and Krackan Point", found in "Ryzen 200 Series, Ryzen 7000 Series, Ryzen PRO 7000 Series, Ryzen 8000 Series, Ryzen PRO 8000 Series, Ryzen AI 300 Series, Ryzen AI PRO Series, Ryzen AI Max 300 Series". If your chip is not on that list, no amount of TOPS on the box will get AMD's own tooling onto it.

Does a Higher TOPS Number Mean Faster Local AI?

For LLMs, almost never — because token generation is bound by memory bandwidth, not by arithmetic.

Here is the arithmetic, shown rather than asserted. A 7B model at Q4_K_M is roughly 0.6 GB per billion parameters, so about 4.2 GB of weights. Generating each token requires reading every weight once. On a Snapdragon X Plus, Qualcomm publishes the memory bandwidth as 135 GB/s. That gives a hard ceiling of 135 / 4.2 = ~32 tokens per second before you have considered the NPU at all. Doubling the NPU from 45 TOPS to 85 does not move that ceiling by a single token, because the bottleneck is the memory bus, not the multiply-accumulate array.

That is the whole reason a 128GB Apple or Strix Halo machine outruns a higher-TOPS thin-and-light on real local models: it can hold a larger model at all, and it can read it faster. The mechanism is worked through in our guide to memory bandwidth and local LLM speed, and the model-size-to-memory mapping is in the VRAM and unified-memory requirements table.

What NPUs are genuinely good at is the workload they were designed for: small, always-on, low-power models — background blur, live captions, transcription, on-device search indexing, Windows Studio Effects. Those run at a fraction of the power a GPU would need, and that is the real product. For the buying decision across all four vendors, our NPU comparison covers what each one is actually good for; if you are choosing hardware for local models specifically, start with the honest guide to running local AI on a laptop instead.

How Do I Check My Own Chip's Rating?

On Windows, Task Manager will confirm you have an NPU but will not tell you its TOPS. Microsoft's documentation notes that "For devices with NPUs, the Task Manager can now be used to view NPU resource usage" — utilisation, memory, driver version, physical location. No rating. So:

1. Get the exact part number: Settings → System → About, or in PowerShell, Get-CimInstance Win32_Processor | Select-Object Name.
2. Look that exact SKU up on Intel ARK or the matching AMD or Qualcomm product page. Not the family — the SKU. Lunar Lake alone spans 40, 47 and 48 TOPS.

On a Mac, the chip name under → About This Mac is all you need; Apple publishes the Neural Engine figure only in the launch newsroom for that generation, and for M3 and M5 it never published one.

FAQ

Is a 10 or 16 TOPS NPU useless?

No — it is just not a Copilot+ NPU. The XDNA 1 and Intel NPU 3 blocks were designed for exactly the workloads they still do well: camera effects, noise suppression, transcription, small vision models running at a fraction of the power the GPU would draw. What they cannot do is satisfy Microsoft's 40-TOPS gate for the Windows AI feature set, and they were never going to be the thing that runs a 7B language model.

Will my NPU speed up Ollama or LM Studio?

Not today. Ollama's own hardware documentation lists NVIDIA GPUs, AMD Radeon via ROCm, Apple Metal and Vulkan — there is no AMD XDNA backend, no Intel Core Ultra NPU backend and no Qualcomm Hexagon backend. Reaching an NPU takes software compiled specifically for it, which is a different toolchain per vendor. If Task Manager shows your NPU flat at 0% while the CPU cooks, nothing is broken.

Two laptops have the same chip but advertise different TOPS. Why?

Usually one of two things: they are different SKUs inside the same family (Lunar Lake spans 40 to 48 TOPS under one "Core Ultra 200V" umbrella), or one vendor is quoting the NPU figure and the other the platform total. Compare part numbers, not family names.

Is a 45-TOPS Snapdragon slower than a 48-TOPS Intel?

That question cannot be answered from those two numbers. They are peak ratings from vendors using different counting conventions — Intel's is explicitly INT8, Qualcomm's carries no stated precision — and neither reflects which runtimes can reach the silicon on your machine. Three points of rated TOPS is well inside the noise of that comparison.

What is the fastest NPU you can buy right now?

On vendor-published ratings, Qualcomm's Snapdragon X2 Elite in its top three SKUs, at up to 85 TOPS. AMD's Ryzen AI 400 at 60 is the highest x86 rating, and Intel's Panther Lake tops out around 50. Whether the fastest-rated NPU is the fastest machine for what you want to run is a different question, and for local LLMs the answer is usually no.

Sources

• Microsoft — Copilot+ PCs developer guide (the 40-TOPS threshold, Task Manager NPU reporting)
• Qualcomm platform pages: Snapdragon X2 Elite, Snapdragon X Elite, Snapdragon X Plus, Snapdragon X
• Intel ARK — per-SKU NPU Peak TOPS (INT8), GPU Peak TOPS (INT8), CPU Peak TOPS (INT8) and Total Peak TOPS fields
• AMD Ryzen AI Software release notes — supported processor codenames and series
• Apple Newsroom: M1, M1 Ultra, M2, M2 Ultra, M3, M4, M5

Every figure on this page is a vendor's own published peak rating, reproduced with the precision label it was published with. This site does not own this hardware and has measured none of it.