Exploring Apple Silicon’s local AI performance with the Mac Studio and M4 Max — M4 Max beats GB10 and Strix Halo in decode throughput, but memory bandwidth isn’t everything

Our local AI explorations in 2026 have so far focused on two main platforms: Nvidia’s GB10, as seen in the DGX Spark and Dell Pro Max with GB10, and AMD’s Ryzen AI Max+ 395, aka Strix Halo, as seen in the Corsair AI Workstation 300 and the Ryzen AI Halo. We’ve generally favored GB10 systems for these kinds of local AI development sandboxes. Their solid all-around performance and broad AI software compatibility make them easy to love, even if raw LLM inference throughput isn’t that high.

But Apple’s Mac Studio is another compelling option for local AI trailblazers who want systems with large unified memory pools, powerful GPUs, and high memory bandwidth to enable faster tokens-per-second throughput than either GB10 or Strix Halo can provide. And until someone builds another unified memory SoC with a memory bus as wide as what Apple uses for its Max and Ultra chips, Apple Silicon is, in fact, the only game in town for more memory bandwidth from a chip of this design.

Why focus so much on memory bandwidth for local LLM performance to begin with? To keep it very short, the decode phase of LLM inference is virtually always sequential. Every output token (starting with the first one generated as part of the prefill process) has to be run through every layer in a model in order to generate the next token, and so all the active model weights in a layer have to be streamed to the GPU from its memory as the token makes its way through each one.

Because the amount of computation required for each individual token at each layer is tiny, the speed of the entire decode process basically becomes dependent on how fast those model weights can be streamed in from GPU memory. That makes systems with large memory pools (for holding large models) and high memory bandwidth (for streaming those model weights to the GPU) appealing, and Apple Silicon is the only platform that lets you have both characteristics at a (relatively) reasonable cost.

Apple’s last Mac Studio refresh was in March 2025 and marked the introduction of two versions of the system. The higher-end version uses the M3 Ultra SoC with 819GB/s of memory bandwidth across both the 28-CPU-core, 60-GPU-core mode and the 32-CPU-core, 80-GPU-core model.

The M4 Max version of the Mac Studio encompasses two versions with greater divergence in specs than just core counts. The base M4 Max includes a 14-core CPU (10 performance cores and four efficiency cores) with 410GB/s of memory bandwidth, and it can be upgraded to a 16-core CPU (12 performance cores and four efficiency cores) with 546GB/s of memory bandwidth.

We reviewed both of these systems at launch, but Apple generously lent us another M4 Max with the 16-core CPU, 40-core GPU, and 128GB of memory for LLM-specific tests. At the time, the price of that system was $3,699 or so, close to the DGX Spark’s $3,999 list price.

But as the RAMpocalypse has rolled on, the Mac Studio has become more limited in its configuration options and harder to find in stock. The M4 Max version we tested now tops out at 64GB of RAM, and lead times for that system are over two months out. Regardless, it’s the most comparable system that Apple has for the DGX Spark and GB10 platform more generally, at least back when dollar-for-dollar meant something in this market.

Before we get into performance results, we wanted to dig into the nitty-gritty of the M4 Max’s 40-core GPU, but Apple doesn’t share much (if any) detailed architectural information on its GPUs, including the number of shader ALUs that make up each GPU core, so it’s hard to directly compare the M4 Max to GB10 or the Radeon 8060S in the Ryzen AI Max+ 395.

The Radeon 8060S on board Strix Halo has 40 RDNA 3.5 compute units, each with 64 shader ALUs for a total of 2,560. AMD clocks the 8060S high, at 2900MHz, to achieve its roughly 14.8 TFLOPS of peak FP32 vector math.

Meanwhile, GB10 has 48 Blackwell SMs with 128 shader ALUs each for 6,144 total CUDA cores, and it too clocks relatively high, at 2450MHz in our testing. It offers a whopping theoretical 30 TFLOPS of FP32 vector compute.

Way back in the day when Apple first revealed the M1 GPU, however, the company disclosed that each core encompassed 128 execution units. Assuming that the same organization has carried through to current Apple GPUs, the 40-core M4 Max GPU would have 5,120 execution units or shader ALUs. That provides an idea of the scale of the GPU, at least, though without knowing clock speeds, we’re hesitant to speculate further about peak FLOPS or other theoretical performance measures.

Apple also only provides raw memory bandwidth figures for the M4 Max, but we can do some back-of-the-napkin math to surmise that the chip likely uses LPDDR5X-8533 on a 512-bit bus to achieve its 546 GB/s of rated memory bandwidth, twice as wide as both GB10 and Strix Halo. The lower-end M4 Max likely uses LPDDR5X-6400. In turn, the M3 Ultra likely uses a 768-bit-wide bus with LPDDR5X-8533 across both of its variants.

We covered the design of the Mac Studio in detail in our initial review, but it’s worth a quick tour to re-familiarize yourself with it. The top and sides of the system are largely featureless, save for an inlaid black Apple logo up top. The front panel has two 10Gbps USB Type-C ports and a UHS-II SD card reader.

The Studio’s thermal design uses a clever intake system that draws in air from vents around the system’s circular foot before exhausting it through a vent at the back of the chassis.

The M4 Max Studio we tested has a whopping four Thunderbolt 5 ports on its back panel with support for data rates up to 120Gbps. Although we’re not testing it today, you can enable RDMA over Thunderbolt 5 for low-latency clustering of Mac Studios using utilities like Exo. These ports also support Thunderbolt displays or DisplayPort 2.1 Alt Mode.

This system also has two USB 3 ports running at 5Gbps, 10Gb Ethernet, an HDMI 2.1 port, and a headphone jack, so it’s got all the connectivity you could practically want out of a compact desktop in 2026.

With those basics out of the way, let’s get to testing.

To test LLM inference performance, we once again turn to llama.cpp, which includes native Metal/MLX support on Apple Silicon when built from source. We used the llama-benchy project as our benchmarking harness.

We tested three models of various memory footprints and architectures: Qwen 3.6-35B-A3B, a recent mixture-of-experts model; Gemma 4 12B, a small dense model, and OpenAI’s gpt-oss-120b, a large mixture-of-experts model. We used Unsloth’s four-bit quantizations of these models for our testing.

Out of the gate with Qwen 3.6-35B-A3B, the Mac Studio delivers prompt processing speeds that beat Strix Halo but are still behind GB10, especially with longer contexts.

Flip over to the throughput results, and the M4 Max pulls ahead of both GB10 and Strix Halo, turning in higher tokens-per-second throughput than both those systems at all context depths.

But despite having double the memory bandwidth of GB10 and Strix Halo on paper, the real-world improvement in token throughput on the M4 Max is just 25% on Qwen 3.6-35B-A3B. That’s certainly welcome, but it’s not nearly the speedup you might expect with twice the memory bandwidth on tap.

The dense Gemma 4 12B, on the other hand, really seems to let the Mac Studio’s bandwidth advantage shine. Tokens-per-second throughput scales more linearly with memory bandwidth with this model, and so the Mac Studio delivers a 1.8X increase in throughput over GB10 and a 2.26X increase over Strix Halo.

But prompt processing speeds are only in line with (or even slightly worse than) Strix Halo here, which isn’t a fantastic result for latency if that’s important to your application.

Finally, we’ll look at gpt-oss-120b. Prompt processing again lands between Strix Halo and GB10, and tokens-per-second throughput is far higher than with either of those systems. But the Mac Studio’s tokens-per-second throughput is, on average, 1.5x higher than GB10 and 1.6x higher than Strix Halo, despite its much higher memory bandwidth. Those are still impressive leads, but they show that memory bandwidth is just one component in judging the performance of an inference pipeline.

On the whole, our LLM inference testing shows that the M4 Max GPU can generate many more tokens per second than competing systems with large unified memory pools, but a memory bandwidth specification alone isn’t a good proxy for delivered performance. The actual boost you get from the M4 Max in LLM inference is highly dependent on the architecture of the model you want to run, so real-world benchmarking is still important.

Image generation performance

LLMs aren’t the only AI workload that one might want to run locally. Image and video generation also matter, and the performance of these diffusion models is generally bound by the compute throughput of the underlying GPU.

Running our Flux.2 Klein image generation test on the Mac Studio initially failed due to an apparent lack of GPU support for the FP8 data type that the ComfyUI workflow uses by default.

We were able to modify the ComfyUI workflow to use the base FP16 version of Flux.2 Klein on the Mac Studio, but that initial compatibility issue shows that you might not enjoy as smooth an image or video generation experience on Apple Silicon as you might on GB10 or even Strix Halo.

Compatibility issues aside, the M4 Max GPU requires even more time than the already pokey Radeon 8060S in this image generation test. Nvidia has repeatedly demonstrated the DGX Spark running as an image and video generation companion for Macs in its demos this year, and this result demonstrates why that pairing might make sense.

Image generation performance with the latest M5 GPU and its dedicated Neural Accelerators for matrix math might be better here, but we don’t have an M5-powered system at hand to compare against the M4 Max. Apple does claim that the M5 Max is “up to 3.8x faster” at image generation than the M4 Max, so take heed if this workload is important to you.

General CPU performance

We already gauged the CPU performance of the M4 Max in our initial review, but we wanted to stack it up against Strix Halo and GB10 directly to see how it performs against these popular platforms in the AI space. To establish a CPU performance ballpark for the M4 Max, we fired up the just-released Geekbench 7 and got to it.

In this benchmark, the M4 Max turns in a whopping 27% advantage in single-core performance and a 21% boost in multi-threaded performance versus the Ryzen Max+ 395’s 16-core, 32-thread Zen 5 x86 CPU.

And comparing Arm to Arm, the M4 Max delivers a 39% single-threaded advantage over the Cortex-X925 “big” core on the Nvidia GB10, along with a 22% advantage in multi-threaded work.

Those results are seriously impressive on their own, but Apple has even more advanced CPU cores in its arsenal now on board the M5 series. We’d expect even larger advantages in general use from those chips.

Geekbench is fine for a quick synthetic test, but I wanted to make sure that its results held in a real-world task, so I compiled the CPU-only version of llama.cpp from source across all three of our test systems using the Clang 21 compiler running across all available physical cores.

With all 16 of its CPU cores in the game, the M4 Max finishes this compilation job in about 37 seconds, anywhere from 33% to 37% faster than the Nvidia and AMD competition here. That’s even faster than the M4 Max’s Geekbench multi-threaded performance delta would suggest.

At least going by these high-level tests, Apple’s CPU performance among SoCs with large unified memory pools is best-in-class, and by a wide margin. And again, that’s with the last-gen M4, not the latest M5.

When judging the efficiency of Apple Silicon or any other platform, one might be tempted to just glance at a power meter like a Kill-a-Watt with the system under load and leave it at that, but instantaneous power draw isn’t the right metric when judging the efficiency of a system for a given task.

Task energy (the amount of power consumed over time) is much more informative. A system that needs a lot of power but finishes a task quickly may consume less task energy and is therefore more efficient than one that consumes a moderate amount of power but takes much longer to finish a job.

To judge inference efficiency, we set up a fixed-length workload with gpt-oss-120b using llama.cpp and llama-benchy: first processing a 2048-token prompt against a 32768-token context depth, then generating 1024 output tokens. We chose that context depth to ensure that our test ran long enough to collect useful data. We recorded this data in watt-hours so that it reflects both energy consumed and time to completion.

At least in this synthetic workload, the M4 Max is more efficient than Strix Halo, but the sheer prompt-processing speed of GB10 lets it finish much faster and use less total power than the Mac Studio does, even though it’s outputting fewer tokens per second in the decode stage.

This test once again suggests that fixating on tokens-per-second throughput alone is perhaps not ideal for judging a system’s inference performance. An ideal LLM inference system would have both high prompt processing speeds and high tokens-per-second throughput for the shortest possible inference turn.

We’re curious how Apple’s latest M5 Max GPU would perform on this workload, as the company claims big gains on prefill compared to the M4 Max and has given that chip even higher memory bandwidth than the M4 Max. If that’s true, the combination of fast prefill and high tokens-per-second on decode would likely make the M5 Max competitive with GB10, if not an outright leader here.

We also gauged idle and load noise levels of these systems to see what it’s like to share space with them. For reference, the noise floor in our testing environment is about 32.6 dBA. We measure sound pressure levels at a distance of 18” from these systems to approximate what it’s like to work with them on top of a desk.

The Mac Studio is impressively quiet both at idle and under load. It barely raises the noise level in our testing environment at idle, and under load, its fans are barely perceptible. Practically any other noise you make or hear in your space will be louder. This is superlative performance.

Our time with a 128GB M4 Max Mac Studio demonstrates why these systems have been so popular among local AI trailblazers. For starters, the spec we tested has an incredibly potent 16-core CPU that’s leagues ahead of both Nvidia’s GB10 and AMD’s Ryzen AI Max+ 395 in both single-threaded and multi-threaded performance.

The massive memory bandwidth advantage of the 40-core GPU configuration of the M4 Max versus Strix Halo and GB10—546 GB/s vs 273 GB/s—also translates directly to higher tokens-per-second throughput in local LLM workloads. The M4 Max’s LLM inference performance is by far the highest of any unified memory platform we’ve tested.

But our tests demonstrate that there’s much more to delivered performance than that raw bandwidth figure. Depending on the model you choose, you might see the M4 Max generate as few as 27% more tokens per second or as many as 82% more TPS than an Nvidia GB10 box, at least with the llama.cpp inference engine.

This result emphasizes that you can’t just eyeball memory bandwidth and conclude that twice the bandwidth means twice as many tokens per second for LLM inference. Testing actual workloads remains essential.

And the Mac Studio doesn’t totally wipe the floor with the competition in our tests. For work like image generation, the Nvidia GB10 GPU still holds a large advantage over the M4 Max. Apple claims this workload will perform significantly better on the most recent M5 Max with its latest GPU architecture.

Beyond its impressive raw performance with LLMs and in general processing tasks, the Mac Studio is a master class in industrial design. It looks distinctly classy on a desk, and its cooling design keeps it incredibly quiet. Under load, it’s barely louder than the Nvidia GB10-powered Dell Pro Max and Ryzen AI Halo are at idle. And when the Studio itself is idling, it’s practically silent. This level of performance in such a compact, quiet, and refined chassis is incredibly impressive.

There’s one major asterisk, though: the ongoing RAMpocalypse means that even if you do want to use a Mac Studio as a foundation for your local LLMs, these systems are no longer available in the configuration that we tested.

A Studio with the high-end M4 Max as seen in this review currently tops out at 64GB of RAM for $5299, and the M3 Ultra version tops out at 96GB (with an even higher 819GB/s of memory bandwidth) for the same price. But neither of these configurations are available from stock, and lead times for both are about two months out, meaning that you practically can’t buy a Mac Studio new right now.

If you really want to try out local AI on Apple Silicon with a large memory pool to match, the best current option may be a MacBook Pro with an M5 Max inside, which is still readily available from stock, still allows you to get up to 128GB of RAM, and benefits from Apple’s latest GPU architecture with dedicated Neural Accelerators for matrix math operations.

But depending on your SSD choice, a 128GB M5 Max MacBook Pro with the highest-end 16-core CPU and 40-core GPU is still a $7K to $10K system today. Even in today’s world of rising prices for access to frontier models, that’s a lot of tokens from the cloud – and, almost assuredly, more intelligent ones – than anything you can get from a single 128GB system from any vendor.

But if you’re dead-set on a single Apple Silicon system with the most memory possible, and you want it now in order to ride the local AI wave, a maxed-out M5 Max MacBook Pro is your only ticket at the moment. Given the refinement and flexibility of the Mac Studio as a desktop or server, we can only hope Apple brings that chip to this system, too.