why this processor is so impressive

Apple’s M1 Ultra chip is a monster. This chip bypasses the physical limitations normally encountered when pairing two chips and promises incredible performance with low power consumption. Let’s take a closer look at what makes this new processor special.

Apple M1 Ultra

Apple M1 Ultra // Source: Apple

This week Apple hosted its keynote. Among all the announcements, they added “one last chip” to the M1 processor family. This chip called M1 Ultra is a new design that uses “UltraFusion” technology to link two M1 Max chips together, resulting in a massive processor that has 16 powerful CPU cores, four efficiency cores, an integrated 64 core GPU and supports up to 128 GB RAM. This chip will be present in the Mac Studio, a machine halfway between the Mac Mini and the Mac Pro.

YouTube LinkSubscribe to Frandroid

After listening carefully to Apple, which showed a lot of beautiful graphics, associated with a large number of superlatives, let’s take a closer look at what makes this new ARM chip so special.

M1 Ultra = M1 Max + M1 Max

First, what Apple did is almost nothing new. AMD already does this on quite a few of its Ryzen. It is a multi-silicon chip approach to fabricate larger chips for higher yields.

M1 M1 Pro M1 Max M1 Ultra
Engraving 5nm TSMC 5nm TSMC 5nm TSMC 5nm TSMC
Number of transistors 16 billion 33.7 billion 57 billion 2 × 57 billion
CPU 4 × Firestorm (S)
at 3.23 GHz 4 × Icestorm (E) at 2.064 GHz
6 × Firestorm (P)
at 3.23 GHz 2 × Icestorm (E) at 2.064 GHz
8 × Firestorm (S)
at 3.23 GHz 2 × Icestorm (E) at 2.064 GHz
16 × Firestorm (S)
at 3.23 GHz 4 × Icestorm (E) at 2.064 GHz
8 × Firestorm (S)
at 3.23 GHz 2 × Icestorm (E) at 2.064 GHz
GPUs 7 hearts om
1278MHz 896 EU
2.29 TFLOPS
14 cores on
1278MHz 1792 EU
4.58 TFLOPS
24 cores at
1278MHz 3072 EU
7.83 TFLOPS
48 cores at
1278MHz 6144 EU
15.66 TFLOPS
8 hearts for
1278MHz 1024 EU
2.61 TFLOPS
16 cores on
1278MHz 2048 EU
5.22 TFLOPS
32 hearts at
1278MHz 4096 EU
10.44 TFLOPS
64 cores on
1278MHz 8192 EU
20.88 TFLOPS
RAM 8GB LPDDR4X-4266
(dual channel, 64 bit)
16GB LPDDR5-6400
(dual channel, 128 bits)
32GB LPDDR5-6400
(four channels, 128 bit)
64GB LPDDR5-6400
(four channels, 128 bit)
16GB LPDDR4X-4266
(dual channel, 64 bit)
32GB LPDDR5-6400
(dual channel, 128 bits)
64GB LPDDR5-6400
(four channels, 128 bit)
128GB LPDDR5-6400
(four channels, 128 bit)
Memory bandwidth 68.2GB/s 204.8GB/s 409.6GB/s 819.2GB/s
neural motor 16 cores
11 TOPS
16 cores
11 TOPS
16 cores
11 TOPS
32 hearts
22 TOPS
transcoding 1x video decoding engine
1x video encoding engine 1x video encoding engine
ProRes Encryption/Decryption
1x decoding engine
2x video encoding engine 2x video encoding engine
ProRes Encryption/Decryption
2x decoding engine
4x video encoding engine 4x video encoding engine
ProRes Encryption/Decryption
Screen management Integrated screen

1 external screen
(1×6K60)
Integrated screen

2 external screens
(2×6K60)
Integrated screen

4 external screens
(3×6K60, 1×4K60)
5 external screens
(4×6K60, 1×4K60)

So we are not talking about 2nd generation Apple Silicon, but about a fourth expansion phase of the 1st generation. The aim is to deliver more power than the already impressive M1 Max. Instead of making an even bigger chip, the M1 Ultra uses the combination of two M1 Max chips.

Comparison between Apple M1 chips

Comparison between Apple M1 chips // Source: Apple

Not only did they combine two M1 Max chips on a single SoC, they made both dies look like one monolithic GPU. The result is a chip that is without a doubt one of the most interesting designs for a consumer SoC.

As early benchmarks show, this dual-die strategy benefits multi-threaded CPU and GPU workloads much more than single-threaded tasks — an area where Apple is already starting to fall behind. By letting the M1 Ultra’s two chips present themselves seamlessly as a single GPU, Apple has started a new technology race to put multi-chip GPUs in high-end consumer hardware.

By putting two M1 Max dies on one chip, Apple has doubled the number of components. This means twice the CPU cores, twice the GPU cores, twice the neural engine cores, twice the LPDDR5 memory channels, and twice the port processing (I/O) for the peripherals.

Mac Studio's spec sheet is impressive

Mac Studio’s datasheet is impressive // ​​Source: Apple

While M1 Max had 16 LPDDR5-6400 channels for a total of 408 Gb/s memory bandwidth, M1 Ultra doubles that to 32 LPDDR5 channels and 800 Gb/s memory bandwidth. An amount that Apple wouldn’t even fully utilize on this configuration. As with the M1 Max, Apple soldered the LPDDR5 chips directly to the SoC, for a total of 8 chips on the M1 Ultra (as you can see in the photo of one). So there is no way to add RAM afterwardsremember to choose the right amount of RAM for your needs while setting up the machine.

UltraFusion Technology

UltraFusion Technology // Source: Apple

What Apple presented as the technology UltraFusion can be described as a very short 10,000-lane highway. This secret ingredient is therefore a very fast interface that makes it possible to connect two M1 Max dies using a silicone interposer. In total, the two matrices have an available data rate of 2.5 Tb/s. According to AnandTech, this technology would resemble Intel’s EMIB technology or Elevated Fanout Bridge (EFB) technology. So Apple wouldn’t be the first to bring such technology to market, but they will undoubtedly be the first to democratize it.

More powerful than a GeForce RTX 3090 for 1/3 of its consumption

The GPU is overkill. According to Apple, the M1 Ultra’s graphics performance would be superior to Nvidia’s GeForce RTX 3090 graphics card, which is currently the fastest graphics card on the market. And what’s more, the M1 Ultra chip only consumes just over 100 Watts, which is 200 Watts less than the RTX 3090.

Apple M1 Ultra: why this processor is so impressive

The GA102 GPU used by NVIDIA has 28.3 billion transistors while the M1 Ultra has 114 billion. Obviously, Nvidia designed their graphics card differently: they didn’t take power consumption into account. Apple designed a larger GPU, with more silicon, to maintain lower clock frequencies (and thus optimize power consumption). Of course, Apple also has a huge advantage in the manufacturing process, using TSMC’s N5 etching process over Samsung’s 8nm process on the Nvidia RTX 30s.

The GPU specs of the Apple M1 Max

The GPU Features of the Apple M1 Max //Source: Apple

Interestingly, the version of the Mac Studio with M1 Ultra does not use the same cooling system as the M1 Max version. This results in a weight difference of almost 1 kilogram. The machines share the same 370W power supply, the extra weight is due to the M1 Ultra having a larger copper thermal module, while the M1 Max has an aluminum heat sink. FYI, copper, which is denser, with more thermal mass and conductivity than aluminum, can therefore hold more heat and move it faster.

Mac Studio Dual Fans

Mac Studio Dual Fans // Source: Apple

Given the performance of an M1 Max, therefore, Apple promises CPU and GPU performance levels for the M1 Ultra that eclipse all current configurations based on Intel CPUs and AMD graphics cards. However, further testing is needed to understand the scenarios that Apple illustrated in its presentation.

Nvidia, AMD and Intel are already in this GPU battle multi die† The theory is that if you take one powerful chip and glue it seamlessly onto another, you’ll get something twice as good. In theory it is simple. In reality, it’s not that simple, and while AMD has already managed to make a similar concept work for its MI200 supercomputer accelerator, no one else has done it for Apple for the general public.

We’ve entered a very exciting new era in GPU development, and Apple’s M1 Ultra is a preview of what’s to come from others who share the same goals as Apple. No doubt you understand much better what justifies the crazy price of the Mac Studio. With the Apple M1 Ultra SoC, the machine is sold from 4599 euros (Apple M1 Ultra, 64 GB and 1 TB).


To follow us, we invite you to download our Android and iOS application. You can read our articles, files and watch our latest YouTube videos.