CNET Lab Exclusive: Beyond More Cores, Higher Scores in the MacBook Pro

Yes, Apple graphics processing unit performance has obviously been getting better over time. This is what we’re used to seeing: a steady upward trend in GPU performance over the years, with the occasional generational discontinuity (such as the M3 Max in this chart).

What we don’t often think about is the diminishing returns to adding more cores. For example, while you may theoretically get better overall performance with a 40-GPU-core M-series Max chip than a 20-core Pro version, you might get less of a boost jumping from 20 GPU cores to 40 cores than you do bumping from 10 in a base chip to 20 in a Pro.

As prices continue to escalate, the incremental cost versus the performance of a better configuration may affect how you spend your money. I’ve been thinking about one aspect of the equation, the GPU, because that frequently governs which configurations you choose from when you’re looking for a certain level of gaming, pro graphics and AI performance.

Key takeaways:

While power constraints don’t seem to be the reason for the M5-generations behavior, low-power performance is disproportionately lower as a percentage of normal performance.
Based on my testing, while MacBook Pros — and Apple M-series processors in general — get faster every generation, they don’t necessarily improve equally across the processor line.
Results show that the M5 generation’s GPU performance doesn’t increases as consistently compared to the M4 series as you increase the number of GPU cores.

GPU evolution

There have been two major generational enhancements to the GPU since the original M1 (the move from Intel to Apple silicon doesn’t count, because Intel chips were awful).

The first overhaul occurred with the M3 generation in 2023, targeting traditional graphics performance areas like rendering and gaming. That’s when Apple added ray tracing accelerators and mesh shaders; ray tracing is a technique for producing more realistic lighting effects, and mesh shaders are a widely adopted method of more quickly and efficiently generating a rendered scene from basic geometric primitives than previous techniques.

That chip also added Dynamic Caching to more efficiently allocate memory between the CPU and GPU as needed, rather than reserving fixed amounts for each, preventing either from starving or overallocating. This was an important change because the Unified Memory Architecture essentially puts everything into a single pool.

Apple added AV1 video format decoding and hardware acceleration for AV1 video decompression to the GPU that year as well. And all the changes were bolstered by a general uplift in performance thanks to a switch to a 3nm process size.

The second shift came with the M5, launched in 2025, featuring enhancements targeting generative AI and ML performance, including a neural accelerator on each GPU core. Apple also refined the ray-tracing engine and shaders (responsible for traditional graphics rendering calculations) for gaming and traditional graphics.

In addition to architectural changes, generation-to-generation increases in overall processor speed and memory bandwidth can affect GPU performance. The latter is directly correlated with core count — chips with more cores have higher bandwidth by design. There was also a significant jump in clock speed from the M3 to the M4 generation, giving the M4 and later generations a noticeable performance boost.

Methodology meets madness

In practice, it’s hard to draw meaningful conclusions when pitting the performance of recent products against the older models using the original test results, because so many changes over time are beyond just hardware.

For computers, two major complicating factors are major generational updates to operating systems and changes within and across benchmarks, applications or games. I’ve seen significant differences in results for benchmark versions that should be comparable.

apple macbook pro 14 m5 sitting on a table in a living room

Complications become doubly problematic when it comes to fast-changing technology; in this case, the pressure to improve gaming graphics and the need to boost AI-related computation speeds using GPUs have led to rapidly evolving methods for measuring their performance. Apple’s Metal graphics API for MacOS continually evolves, with refinements and optimizations as well as support for new hardware capabilities.

So even for an apples-to-apples comparison, you’re comparing fresh-picked apples to ones languishing on the ground. The fresh apples will always taste better.

As a result, whenever new Macs come out, I inevitably end up retesting every model I want to compare in my reviews. (With the exception of battery life.)

For this story, I even had to run new tests on the two MacBook Pros I reviewed just five months ago. (I’d have needed to do it anyway for comparison with the new Mac Studio and Mac Mini.)

Another caveat to keep in mind is that this is based on a small sample. I don’t have a MacBook Pro with every chip version — as defined by the number of CPU and GPU cores — much less with every memory configuration, both of which can have a big impact on performance. There’s generally one version of the base M chip, and two each for the Pro and the Max. I try to account for factors like those when drawing conclusions.

Also, over time, I’ve noticed that chip performance across the desktops and laptops has been consistent enough that I’m comfortable using Mac results as a proxy for MacBook Pro performance, which is why I’ve used results from Mac Studios and the M6 Mac Mini to supplement or for informational purposes where it makes sense.

Best Apple MacBook of 2026

Falling multipliers

While more GPU cores will invariably deliver better performance for applications that take advantage of them, how much of a boost you’ll get from additional cores varies a lot across chip generations, the applications you use and how many cores you’re adding.

To analyze GPU core scaling, as defined by how much better the performance gets when you increase the number of cores, I’ve used the ratio of scores or frame rates on a variety of benchmarks as a proxy. Note that linear scaling — doubling the cores delivering double the score — is rare in general, not just on benchmarks. At the very least, there’s usually some overhead that lowers the multiplier.

For example, on Cinebench 2026’s GPU test, the M4 Max’s 40-core GPU performed almost exactly 2x faster than the M4 Pro’s 20-core GPU, while the M5 Max performed only about 1.8x faster. This equal-or-lesser scaling appears across the board, to varying degrees, for the M5 vs. M4 generations.

I thought one possible interpretation was that the M5 Max (at least the 40-GPU-core version) is memory-bottlenecked sooner than the M4 Max; in other words, that the M5 Max’s GPU performance might scale better if you could configure it with 256GB rather than the current maximum of 128GB. But unless the M5 Ultra’s 80 cores were bottlenecked by 256GB, which is unlikely, that’s not the case.

In general, one potential culprit is the pattern of chip updates, in which an architectural change in one generation is followed by optimizations in the next. If true, an eventual M6 Max’s GPU should scale better than the M5 Max, especially given the switch to the M6’s smaller process size.

On the other hand, Apple added two more GPU cores to the M6, which may presage an increase in core count for the higher-end versions of the chips as well. Or extra cores might be necessary to deliver a material improvement from generation to generation.

Core contributions

AI exercises the GPU much differently than traditional graphics, so it’s no surprise that performance scales differently. But it’s also a harder metric to pin down. Unlike gaming and graphics, which use the GPU exclusively for most operations, AI processing uses both the Neural Engine/NPU and neural accelerators depending upon system resources and computation type, and while benchmarks like Geekbench AI can programmatically isolate which device to use, in reality, workloads tend to be distributed based on whatever makes sense at the time.

At an absolute level, gaming performance has been rising over time. For instance, on a Cyberpunk 2077 benchmark with ray tracing, there’s a generational increase in what I call “performance per core,” a proxy calculated by the score divided by the number of GPU cores, to compare across different chip variants. The M3 generation delivers roughly 1 frame per second per GPU core; the M4 rises to about 1.5fps, and the M5 to about 2.3fps — except for the M5 Max, which drops to around 1.8fps.

That translates to about a 56% boost from the Pro to the M5 Max, compared to a 63% increase for the same in the M4 line. The gap widens if you look at the Solar Bay Extreme benchmark, which measures similar types of rendering: 97% for the M4 Pro to the M4 Max, but only 78% from the M5 Pro to the M5 Max.

But you can also see that while the M4’s scaling was pretty consistent across the board — it stayed roughly the same from 10 to 40 cores — the M5’s dropped as the core count doubled. And the M5 Ultra, which isn’t, and likely will never be, available in a MacBook Pro, continues that trend.

For AI, M4 is less consistent, in that the GPU scales a bit differently from the M4 through the M4 Max on some tests, but the M5 still obviously drops as you increase core count.

What about power?

MacBook Pros have essentially two power settings, which the system can switch between manually or automatically: high performance and low power. Unplugged, the default is high performance, which, as far as I’ve been able to tell, runs the same on battery as it does when plugged in. It occurred to me that maybe the M5 line is more power-hungry than previous generations, and that, therefore, more cores would require proportionally more power, and that the higher-core-count chips were getting throttled.

But while that could be true — there’s a similar drop in performance at low power as cores are added — that wouldn’t explain the M5 Ultra unless the Mac Studio throttles as well.

Though the M6 does seem to deliver better per-core performance than previous generations, that appears to be true for every base chip; we won’t know whether it looks more like the M4 or the M5 until we’ve got a Pro and a Max.