Apple M3 Ultra vs M4 Max: Which Mac Studio Should You Buy?
Apple has released two Mac Studios with M3 Ultra and M4 Max chipsets. The Cupertino giant has called the M3 Ultra “the highest-performing chip” the company has ever created. This chip comes with a whopping 512GB of unified memory. But, how does it fare against the Apple M4 Max? To find out, we have compared the M3 Ultra and M4 Max chipsets in detail so you can easily decide which Mac Studio to get.
Apple M3 Ultra vs M4 Max: Specs Comparison
| Apple M3 Ultra | Apple M4 Max | |
|---|---|---|
| Fabrication Process | TSMC’s 3nm | TSMC’s 3nm |
| CPU | 28 cores (20P + 8E) 32 cores (24P + 8E) Up to 4.05GHz frequency |
14 cores (10P + 4E) 16 cores (12P + 4E) Up to 4.5GHz frequency |
| GPU | 60 and 80 cores HW Ray Tracing |
32 and 40 cores HW Ray Tracing |
| Neural Engine | 32 cores | 16 cores |
| Unified Memory (RAM) | 96GB to 512GB | 36GB to 128GB |
| Memory Bandwidth | Up to 819 GBps | Up to 546 GBps |
| Storage | 1 TB to 16 TB SSD | 512 GB to 8 TB SSD |
| Connectivity Port | Thunderbolt 5, 120 Gbps | Thunderbolt 5, 120 Gbps |
| Wireless Connectivity | Wi-Fi 6E, Bluetooth 5.3 | Wi-Fi 6E, Bluetooth 5.3 |
Apple M3 Ultra vs M4 Max: CPU
Let’s start by comparing the CPU performance of the Apple M3 Ultra and Apple M4 Max. In the Geekbench 6.4 test, the M3 Ultra with a whopping 32 CPU cores barely outperformed the M4 Max, which includes only 16 cores.
In single-core, the Apple M3 Ultra’s Geekbench score is 3,249 while the M4 Max achieves 4,082. In single-core CPU performance, M4 Max leads by around 25%.
The multi-core result is even more surprising. Despite packing 32 CPU cores, the Apple M3 Ultra managed to score 28,345 points while the M4 Max with just 16 cores, came at a striking distance, scoring 26,474 points. Here, the M3 Ultra leads by a small margin of 6.6%.
So what explains this surprising result? First, the Apple M3 was announced in late 2023 and it was fabricated on TSMC’s first-gen 3nm process node. It’s unclear whether the M3 Ultra is also built on the same node. In contrast, the Apple M4 series chips are built on TSMC’s second-gen 3nm node, offering higher transistor density and better efficiency.
On top of that, the M4-series chips are architecturally much better than the M3-series chips. Apple added SME units to the M4, and it’s designed on Armv9 architecture while M3 is developed on the older Armv8 design. This allows the M4 Max to operate at a higher clock speed, going up to 4.5GHz compared to M3 Ultra’s 4.05GHz.
That said, what is surprising is that those 32 CPU cores in M3 Ultra are struggling to scale beyond a certain point. This could be due to Apple’s UltraFusion packaging, which links two M3 Max chipsets to create the M3 Ultra chipset, instead of using a single monolithic die.
Or more likely, it could be Geekbench’s scaling limitation itself. In that case, we need to wait for other CPU benchmarks like Cinebench 2024 to ascertain the multi-core CPU performance of M3 Ultra.
Apple M3 Ultra vs M4 Max: GPU
The maxed-out M3 Ultra features a massive 80-core GPU, and the M4 Max comes with a 40-core GPU. In early Geekbench 6 GPU results, the M3 Ultra achieved 259,277 points while using Apple’s Metal graphics API. On the other hand, the Apple M4 Max scored 192,889 points with just half the cores.
Here, we can see that the M3 Ultra’s GPU outperforms the M4 Max GPU by about 35%, which is not substantial, given that it includes 80 GPU cores. That said, the bottleneck might be caused by Geekbench’s scaling problem. We should wait for 3DMark results to draw a definitive conclusion.
Apple M3 Ultra vs M4 Max: Neural Engine
Now, let’s talk about the performance of Neural Engine on AI workloads. In the Geekbench AI test, the M3 Ultra with a 32-core Neural Engine couldn’t beat the 16-core Neural Engine on the M4 Max, which is strange. Note that both devices are using the Core ML framework and the Neural Engine as the backend hardware.
| Geekbench AI | Apple M3 Ultra (32-core Neural Engine) |
Apple M4 Max (16-core Neural Engine) |
|---|---|---|
| Single Precision | 5,501 | 5,968 |
| Half Precision | 30,149 | 36,176 |
| Quantized | 33,362 | 48,872 |
In all three precision formats, the Apple M4 Max with half the Neural Engine cores, outperforms the M3 Ultra. While you get access to a large pool of memory on the Apple M3 Ultra, up to 512GB unified memory, which will allow you to load larger AI models, the Apple M4 Max is faster in AI workloads.
Apple M3 Ultra vs M4 Max: Memory Bandwidth
Finally, coming to memory, the Apple M3 Ultra can be configured from 96GB to a whopping 512GB of unified memory. With such a large memory, you can load AI models up to 600B parameters on a single Mac Studio. On top of that, the maxed-out variant offers memory bandwidth of up to 819 GBps which is great for running local AI models.
On the other hand, the Apple M4 Max allows you to configure unified memory from 36GB to 128GB. Its top-end variant offers memory bandwidth of up to 546 GBps, which is lower than the M3 Ultra, but it’s still quite fast.
Early Thoughts on Apple’s New Mac Studios
The initial benchmark results for the Apple M3 Ultra and M4 Max are quite surprising, to be frank. That’s why I am not drawing a final conclusion right now. However, early data suggests that the M3 Ultra, despite packing double the CPU, GPU, and Neural Engine cores, has struggled to deliver the vastly superior performance that we expected over the M4 Max.
It appears Apple has released the M3 Ultra to take on Nvidia’s Project DIGITS and Framework’s new AMD-powered Desktop for AI workloads in the consumer segment. Apple has mentioned LM Studio in its blog post which suggests that the company is keen on attracting users who want to run large language models locally. It also comes with half a terabyte of unified memory.
In my assessment, the Apple M4 Max-powered Mac Studio is a better buy for most users. It starts at $1,999, which is half the price of the M3 Ultra Mac Studio, and delivers absolutely phenomenal performance.
M4 Max is also architecturally a better chipset than the M3 Ultra with significant improvements in the CPU, GPU, and Neural Engine. Finally, I would suggest waiting for more benchmark results to ascertain the performance gap between M3 Ultra and M4 Max.
Source link
