AMD Apple Laptop Tests Expose a Wide Battery Performance Gap

AMD Apple laptop performance moved in opposite directions during new unplugged testing, with one Ryzen system losing about half its benchmark speed on battery power.

Tom's Hardware compared recent laptops powered by AMD, Apple, Intel, and Qualcomm processors on AC and battery power. The publication tested each machine against itself, avoiding the misleading claim that one benchmark score identifies the fastest processor overall. That distinction matters because the results measure consistency, not absolute speed.

The widest gap appeared on an Acer Swift Go 16 AI with AMD's Ryzen AI 7 445. Its benchmark results fell by an average of 50.06 percent when unplugged. Apple's 14-inch MacBook Pro with an M5 Max averaged a 0.08 percent increase, effectively producing the same tested performance in both states.

Intel also stayed remarkably close to parity. Qualcomm's Snapdragon X2 Elite Extreme produced a more complicated result, gaining performance in one workload while losing a similar share in another. The outcome turns a familiar laptop specification debate into a system-design question: how much of the performance purchased at checkout remains available away from a wall outlet?

The Test Changed What Laptop Performance Means

A laptop's maximum benchmark score tells only half the story when unplugging it can change the machine's behavior.

The battery performance tests covered recent processors from the four major laptop chip vendors. Tom's Hardware used Cinebench 2026 single-threaded and multi-threaded workloads, plus a HandBrake video-transcoding task.

Cinebench renders a scene to measure processor performance. Its single-threaded test emphasizes the speed available to one processing thread. The multi-threaded run distributes work across more CPU resources, creating higher sustained demand.

HandBrake is an open-source transcoder, meaning it converts video from one format or resolution into another. The test converted a 4K video to 1080p, representing a sustained creative workload rather than a brief responsiveness check.

The Windows laptops used the balanced power profile, which was already the default on most tested systems. The MacBook Pro used Apple's automatic power-management profile. These settings matter because operating systems can change processor limits, fan behavior, and scheduling when a charger is removed.

Microsoft describes Balanced mode as a compromise between performance and battery life. Windows also offers modes that prioritize efficiency or maximum performance through its power mode controls.

Keeping the default or balanced configuration makes the exercise relevant to ordinary buyers. It does not reveal every laptop's highest possible unplugged score, but it shows what many owners encounter without changing advanced settings.

The tested machines included an Asus Zenbook Duo with Intel's Core Ultra X9 388H and an Asus Zenbook A16 with Qualcomm's Snapdragon X2 Elite Extreme. Apple was represented by a 14-inch MacBook Pro using the M5 Max.

Tom's Hardware tested two AMD configurations. The Acer Swift Go 16 AI used a Ryzen AI 7 445, while an HP OmniBook X Flip 14 used a Ryzen AI 9 465. An Intel Core Ultra 7 258V laptop was also included for generational context.

This was not a controlled processor laboratory comparison using identical chassis, cooling, batteries, and operating systems. Such a design would be impossible across macOS, Windows on Arm, and conventional Windows hardware.

Instead, the results capture an equally important reality. Buyers purchase complete laptops, not processors floating outside a chassis. Firmware, cooling, battery discharge limits, operating-system policy, and manufacturer tuning all shape the performance they receive.

That system-level perspective explains why the two AMD laptops did not behave alike. It also prevents the weakest Ryzen result from becoming a universal claim about every AMD-powered notebook.

The comparison therefore changed the useful question. Rather than asking which chip wins a benchmark, buyers should ask whether a particular laptop preserves its capabilities in the setting where they intend to use it.

For a desktop replacement that stays connected all day, a large battery penalty may be tolerable. For a photographer processing work on location, an engineer compiling code away from an outlet, or a student handling media projects between classes, it changes the product's practical value.

The conflict is no longer simply AMD Apple, x86 versus Arm, or Windows versus macOS. It is advertised performance versus accessible performance, tested at the moment a portable computer actually becomes portable.

Intel and Apple Made Unplugged Work Look Ordinary

Intel and Apple delivered the most consistent results, making removal of the charger almost invisible in these workloads.

The Apple M5 Max configuration produced the tightest range. Its Cinebench 2026 single-threaded score increased by 0.27 percent on battery. Its multi-threaded result rose by 0.85 percent, while HandBrake performance fell by 0.89 percent.

Those movements stayed below one percentage point in every workload. The average change was a positive 0.08 percent, which is small enough to treat AC and battery results as practical parity within this test.

The Intel Core Ultra X9 388H system was nearly as consistent on average. Its three results averaged a 0.01 percent increase when unplugged. That average, however, hides modest movement between individual tasks.

The Intel laptop produced the same single-threaded Cinebench result on AC and battery. Its multi-threaded score fell by 3.41 percent, while HandBrake performance increased by 3.43 percent.

Those opposing changes canceled each other in the average. They still suggest that workload-specific tuning can matter even when a summary number looks perfectly flat.

For most users, a movement near three percent will be less noticeable than the natural variation caused by background tasks, temperature, or application state. It is fundamentally different from losing about half of a system's measured output.

Apple's consistency is especially relevant because the M5 Max targets demanding users. Creative professionals often pay for higher CPU and graphics capacity specifically to complete intensive work quickly.

A fast laptop that preserves its speed on battery supports editing, encoding, development, and analysis outside a fixed office. It also makes performance easier to predict when a meeting room, aircraft seat, or worksite lacks accessible power.

Intel's result applies similar pressure within the Windows market. A buyer does not need to switch operating systems or processor architecture to get close AC and battery behavior in the tested workloads.

That weakens an old assumption that Windows laptops must substantially throttle when unplugged. It also raises expectations for every manufacturer shipping a premium machine around the same time.

The result does not establish that Intel and Apple always consume the same amount of energy while maintaining performance. Performance parity and battery endurance are separate measurements.

A processor can complete a task quickly while drawing more power, or it can reduce instantaneous performance and run longer. The test measured benchmark changes, not total work completed before the battery emptied.

That distinction becomes important for sustained jobs. A laptop preserving full speed for a video export still needs enough stored energy to finish the export. Users need both predictable performance and acceptable runtime.

Thermal conditions can also alter the picture. A short benchmark might finish before a chassis reaches its long-term temperature limit. Repeated rendering or encoding could produce different behavior as heat accumulates.

Even with those limits, the Intel and Apple results establish a useful baseline. Near-parity performance is not merely a future processor promise. It exists in shipping system designs under the tested conditions.

The AMD Apple comparison therefore pressures laptop makers more than chip vendors alone. If one platform can preserve performance, buyers can reasonably ask why another complete system cannot.

That question will increasingly shape reviews. Battery-life charts remain essential, but they do not reveal whether a laptop quietly trades away the speed shown in plugged-in benchmark results.

A machine that lasts longer because it becomes much slower offers a different bargain from one that combines long runtime with stable output. Those products should not receive the same portability judgment.

AMD Apple Results Show the Chassis Matters as Much as the Chip

The two Ryzen laptops produced dramatically different penalties, showing that AMD's result cannot be separated from each manufacturer's system tuning.

The Acer Swift Go 16 AI was the clear outlier. Its Ryzen AI 7 445 system lost 56.25 percent in Cinebench 2026 single-threaded performance after the charger was removed.

Its multi-threaded score dropped by 42.28 percent. HandBrake performance declined by 51.64 percent. Across the three workloads, its average loss reached 50.06 percent.

That is not a small efficiency adjustment. A buyer who selected the laptop using plugged-in results would receive a materially different performance class during battery operation.

The HP OmniBook X Flip 14, powered by a Ryzen AI 9 465, also slowed while unplugged. However, its average decline was 5.45 percent, far closer to the consistent behavior seen elsewhere.

The difference between those two AMD systems is central to interpreting the news. Both use recent processors from the same vendor, yet their average penalties are separated by more than 44 percentage points.

That spread points toward the platform around the processor. Laptop makers configure firmware, voltage behavior, temperature limits, fan curves, and battery discharge rules for each design.

A thin chassis may face different cooling constraints from a thicker one. A battery and power-delivery system may also limit how much energy the processor can draw during a sustained workload.

Manufacturers make those choices for several reasons. They must protect battery health, manage surface temperature, contain fan noise, and avoid unstable voltage under a sudden load.

None of those goals automatically excuses a large performance drop. They do explain why attaching a processor brand to a result without naming the tested laptop would be inaccurate.

AMD has already shown that its silicon can support consistent portable performance in another product category. Valve's Steam Deck uses a semi-custom AMD system-on-chip and was designed around similar performance on and off its charger.

A system-on-chip, or SoC, combines major computing components into one package. The Steam Deck example demonstrates that an AMD-based device can prioritize predictable battery behavior when the complete system is designed for it.

Its existence sharpens the criticism rather than removing it. The weakest laptop result looks less like an unavoidable property of AMD architecture and more like an optimization problem that AMD and its manufacturing partners need to address.

Apple controls more of its stack. It designs the processor, operating system, power-management behavior, and the finished MacBook. That vertical control can reduce the number of parties responsible for balancing performance and battery constraints.

AMD supplies processors to many manufacturers across numerous chassis designs. This gives buyers more form factors and configurations, but it also creates wider variation.

Intel operates through a similarly diverse Windows ecosystem, yet its newest tested platform stayed near parity. That means ecosystem complexity alone cannot fully explain the Acer's result.

The AMD Apple split is therefore best read as a consistency problem. Apple's tested machine behaved predictably. AMD's two tested machines ranged from a moderate decline to a severe one.

For buyers, processor branding becomes an incomplete shortcut. A Ryzen AI sticker cannot answer how a specific model behaves on battery, just as an Intel or Snapdragon label cannot guarantee identical tuning across every laptop.

Procurement teams should pay attention to model-level reviews. The relevant evidence includes unplugged benchmarks, sustained-load tests, battery runtime, surface temperatures, and fan behavior.

Knowledge workers should also test their real applications during a return period. A benchmark may expose a broad limit, but a spreadsheet model, local code build, video export, or data-analysis workflow may respond differently.

The lesson is particularly important when teams document device evaluations. Keeping benchmark settings, firmware versions, and workload notes in a searchable technical knowledge base makes later purchasing comparisons more defensible.

A single percentage cannot explain the entire laptop. Yet a 50 percent average decline is too large to dismiss, especially when competing designs show that near-parity operation is attainable.

Qualcomm Won One Test and Lost Another

Qualcomm's Snapdragon system preserved average performance, but its large workload swings make that average less reassuring.

The Asus Zenbook A16 with Snapdragon X2 Elite Extreme averaged a 0.41 percent decline on battery. On that summary figure, it appears close to the Intel and Apple systems.

The individual results tell a more complicated story. Its Cinebench 2026 single-threaded score fell by 0.96 percent, which is effectively stable for practical interpretation.

Its multi-threaded Cinebench result increased by 13.06 percent on battery. HandBrake then moved in the opposite direction, falling by 13.33 percent.

Those changes nearly cancel in the average. They do not deliver the same predictability as Apple's sub-one-percent movement across all three tests.

The result illustrates why averages can conceal behavior that users will notice. A rendering task and a video transcode can exercise different code paths, processor units, and operating-system policies.

Qualcomm's position also carries an architecture question. Snapdragon X processors use Arm instruction architecture, while AMD and Intel laptop processors traditionally serve the x86 Windows ecosystem.

Windows 11 can run native Arm64 software and emulate many existing x86 and x64 applications. Emulation translates software built for another processor architecture so it can run without a complete rewrite.

Microsoft's Windows on Arm guidance says Arm-native applications provide the best path to performance, responsiveness, and battery life. Existing applications can still run through operating-system emulation.

That compatibility layer complicates comparisons. A workload compiled natively for Arm may behave differently from one passing through emulation, even on the same laptop and power profile.

The Tom's Hardware results do not support a broad claim that Arm is inherently more consistent or efficient. They show one Snapdragon laptop producing nearly equal average performance through offsetting changes.

That is meaningful progress from earlier Windows-on-Arm generations, which often faced larger application compatibility and performance concerns. It is not the same as eliminating workload variability.

For users, the key question is whether their specific software is Arm-native. Web browsers, office applications, media tools, development environments, and specialized plug-ins can follow different compatibility paths.

Kernel drivers present another constraint because Windows cannot emulate them in the same way as ordinary user applications. Specialized hardware, older peripherals, or security tools may still require native Arm64 support.

This issue does not appear in a short CPU benchmark, but it shapes the value of Qualcomm's laptop platform. Portable performance matters only when the applications and devices required for work function correctly.

Qualcomm's 13.06 percent Cinebench increase on battery is also a reminder that benchmark variation does not always mean throttling. Scheduler behavior, temperature, background activity, and run-to-run variance can produce counterintuitive gains.

The paired 13.33 percent HandBrake decline suggests a more workload-dependent policy. It deserves repeated testing across additional laptops before anyone treats it as a universal Snapdragon characteristic.

The skeptical reading is straightforward. Three workloads across one primary Snapdragon configuration cannot settle a platform-wide question.

The favorable reading is also legitimate. Qualcomm's average remained close to parity, and its worst observed decline was far smaller than the Acer Ryzen system's losses.

That places Qualcomm between two narratives. It can compete with Intel and Apple on headline unplugged consistency, but it has not yet made every demanding workload behave the same way.

Laptop manufacturers will need to tune Snapdragon systems beyond battery-life marketing. Reviewers will need to verify native application status and disclose it alongside their benchmark results.

Buyers should resist using a single average as the decision point. The most relevant number is the result from the workload that resembles what they actually do.

A developer compiling Arm-native code may see one pattern. A creator transcoding video may see another. A corporate user relying on an emulated legacy application may encounter a third.

The Qualcomm result is therefore neither a clean victory nor a failure. It is evidence that average parity can coexist with meaningful task-specific uncertainty.

What the Benchmarks Still Cannot Tell Buyers

Performance parity is valuable, but it does not reveal runtime, energy efficiency, thermal stability, or behavior across the full application ecosystem.

The first missing measurement is endurance under load. Tom's Hardware calculated the percentage difference between plugged and unplugged benchmark results, not how many demanding tasks each battery could complete.

A laptop can maintain full performance by consuming energy quickly. Another can slow down, conserve energy, and finish more total work before shutting down.

Neither behavior is universally better. A field photographer facing a deadline may value the fastest possible export, while a traveler may prefer lower speed over several additional working hours.

A stronger follow-up would measure both task completion time and energy consumed. Repeating the task until battery depletion would show how much useful work each system completes per charge.

The second uncertainty is chassis representation. The comparison included specific laptops, not every device built around each processor family.

The Acer and HP results already show why this matters. Two AMD configurations produced sharply different penalties, preventing one machine from defining an entire vendor.

The same caution applies to Intel, Qualcomm, and Apple. Apple's tighter hardware range may reduce variation, but different MacBook sizes and cooling designs can still behave differently.

The third limitation involves benchmark coverage. Cinebench and HandBrake are relevant CPU workloads, yet they do not represent gaming, local AI inference, software compilation, photo processing, browser workloads, or GPU rendering.

Modern laptop processors contain more than CPU cores. Many include integrated graphics and neural processing units, or NPUs, designed to accelerate selected machine-learning operations.

Those components can follow separate power limits. A laptop might preserve CPU performance while reducing GPU speed, or maintain an NPU workload while constraining another subsystem.

Display brightness, wireless activity, storage behavior, memory configuration, and connected devices also affect practical battery use. None is captured by comparing a few processor benchmarks alone.

The fourth question is repeatability. Small positive or negative movements can fall within ordinary benchmark variation unless tests run enough times under controlled conditions.

Apple's movements below one percent should therefore be described as practical parity, not as proof that battery operation makes the processor faster. Intel's average of plus 0.01 percent deserves the same caution.

Qualcomm's opposing 13 percent swings and Acer's roughly 50 percent declines are large enough to demand attention. Even so, repeated runs would help separate consistent policy from transient conditions.

Firmware updates can change these outcomes after launch. Manufacturers routinely adjust fan control, voltage tables, scheduler behavior, and power profiles through BIOS or operating-system updates.

That creates a moving target for reviews. A laptop tested at release can behave differently months later, while two retail units may ship with different firmware versions.

Battery age adds another variable. A fresh battery can sustain voltage and current differently from a worn one, particularly under a demanding load.

These limitations do not invalidate the comparison. They identify exactly what the initial results should trigger next: deeper model-level testing, clearer disclosure, and closer scrutiny of manufacturer defaults.

The benchmark also exposes a communication problem. Laptop specifications usually promote processor model, memory capacity, storage, display quality, and maximum battery-life claims.

They rarely state how much application performance remains available when unplugged. Buyers must search reviews for that behavior, assuming reviewers measured it at all.

A useful industry response would add AC-to-battery performance retention to standard laptop testing. Reviewers could report the percentage retained across CPU, GPU, and representative application workloads.

Manufacturers could publish their own performance curves at different power modes. Independent testing would still be necessary, but buyers would at least gain a claim that can be checked.

Enterprise buyers have another reason to care. A fleet may perform consistently at office desks yet slow sharply during travel, field service, classroom use, or flexible work.

That variability affects project estimates and employee expectations. It can also undermine the rationale for buying higher-end processors if workers frequently operate away from chargers.

The proper conclusion is not that every buyer should choose Apple or Intel. Operating-system requirements, application support, repairability, graphics performance, and form factor remain major considerations.

The defensible conclusion is narrower. Unplugged performance deserves to become a standard purchasing metric, and the AMD Apple gap shows why processor specifications cannot replace complete-system evidence.

The Next Tests Will Decide Whether AMD Can Close the Gap

Three signals will show whether the weakest result was an isolated laptop problem or a broader challenge for AMD's mobile platform.

The first signal is repeat testing across more Ryzen AI 400-series laptops. Reviewers need machines from several manufacturers, with different chassis sizes, batteries, cooling systems, and firmware configurations.

If most models stay close to the HP OmniBook's 5.45 percent average decline, the Acer result will look like an implementation outlier. That would weaken claims of a broad AMD architecture problem.

If several systems approach the Acer's roughly 50 percent loss, pressure will shift toward AMD. The company would need to explain how its reference guidance and platform controls allow such wide unplugged penalties.

The second signal is firmware behavior on the Acer Swift Go 16 AI. An update that materially improves battery performance retention would point toward system tuning rather than a fixed silicon limitation.

Reviewers should repeat the same Cinebench and HandBrake workloads after every relevant BIOS, driver, and Windows power-management update. They should retain identical profiles and environmental conditions.

A large improvement would strengthen the argument that laptop vendors control much of this outcome. No improvement would leave questions about battery delivery, cooling design, or platform-level limits.

The third signal is broader workload testing across Intel, AMD, Qualcomm, and Apple. CPU rendering and video conversion should be joined by code compilation, photo exports, GPU tasks, and local AI inference.

Those tests need two outputs: performance retention and total battery energy used. Together, they would reveal whether a machine maintains speed efficiently or simply spends its battery faster.

Qualcomm needs this expanded test almost as much as AMD. Its opposing Cinebench and HandBrake movements show that average parity does not guarantee workload parity.

Intel must also prove that the Core Ultra X9 result persists across more manufacturers. Its near-zero average is encouraging, but the Windows ecosystem contains too many designs for one laptop to settle the issue.

Apple faces a different test. Its M5 Max MacBook Pro maintained benchmark performance, but buyers still need to know how sustained full-speed work affects useful runtime and thermal behavior.

Over the next product cycle, review language should become more precise. "Fast on battery" must be separated from "long battery life," because they describe different product qualities.

A laptop can excel at one, both, or neither. Treating those outcomes as a single efficiency score hides the tradeoff buyers actually face.

The AMD Apple comparison offers a practical starting point. Apple delivered the steadiest individual results, Intel produced the steadiest average, Qualcomm was variable by task, and AMD ranged from moderate loss to a severe outlier.

No single brand won every dimension because this test did not measure every dimension. It did establish that removing a charger can expose choices hidden by conventional specification sheets.

Before choosing a laptop for demanding mobile work, look for model-specific unplugged tests using your type of application. Compare the percentage of performance retained, then compare how long the machine sustains it.

If reviewers begin publishing both measurements, manufacturers will have a clearer incentive to optimize complete systems. If they do not, buyers will keep discovering the real limits only after leaving the outlet behind.