Apple reinvents the GPU detached from the CPU! Soon interchangeable RAM and SSDs?

Monday 23rd December 2024 07:50 PM

There was a time when CPU, RAM, and GPU were separate components, which manufacturers integrated onto a large motherboard, which connected all these little people together.

Ce concept is not dead, PCs still exist and for some retain this more modular architecturewhich allows you to change the GPU without changing the CPU and to add RAM as needed.

The SoC revolution

However, with the emergence of mobile devices, SoCs (System On a Chip) have revolutionized the sector by integrating all the components on the same chip. From the first iPhone to the latest iPad, this monolithic design has even invaded Macs, with Apple Silicon processors (M1, M2, M3, M4…).

Apple was rather right, because this architecture makes it possible to reduce costs (everything is engraved at the same time), and to discuss the components much more quickly while consuming less energy. The engineers have even released these SoCs in Pro, Max and Ultra versions, by making the chips bigger to offer more power. Initially adapted for mobile, this technology has invaded our laptops, and even stationary machines, such as Mac Studio or Mac Pro.

The SoC has its limits!

However, this concept also has its limits. It is for example expensive and complicated to have complex combinations -for example, having a lot of GPUs for AI and 3D or a lot of CPUs for more traditional tasks, such as audio. A small engraving error on one of the components renders the whole thing completely obsolete. Another concern, when CPUs and GPUs have to share the same space, they tend to overheat more quicklywe see this with MacBook Air for example where the demand on the GPU forces the CPU to lower its frequencies, in the absence of active ventilation.

According to the famous Ming-Chi Kuo, with the M5 chip (expected around 2026), Apple has decided to follow the trend established by the rest of the industry (AMD in particular), namely to separate the engraving of the different components and then bring them together on the same card. It is more expensive per unit, but it also offers more flexibility and fewer engraving problems, which can prove profitable over time, to be confirmed of course.

Pour Apple Intelligence ?

He is not there is no question here of really returning to separate CPUs and GPUs on interchangeable daughter cardsbut rather to use TSMC's SoIC-mH technique (the founder of Apple) which will allow engraving in 2.5D and more only flat as currently: we will obtain chips still in one piece, but with several layers of transistors etched separately, which should make it possible to better manage the rise in temperature in particular and to create more varied combinations or even to add other components.

This change would also be motivated by the needs of Apple Intelligence and the new servers in preparation (Private Cloud Compute (PCC). Indeed, for these machines, Apple undoubtedly needs astronomical quantities of Neural Engine compared to classic MX chips. Separating the components thus makes it possible to create more optimized machines according to the purpose of processing. SSD memory chips or RAM are already separate componentswhich Apple assembles at the end of the process.

Mac4Ever has long run on XServe, the servers made by Apple!

This would perhaps allow relaunch the Mac Pro with more modular chips and why not, dedicated GPUswhich could be added or modulated according to needs. The most teasing will retort that this was already the case under Intel, and they are not wrong: Apple today responds quite poorly to certain very demanding areas in 2D/3D calculations or even in terms of AI, but which are become almost marginal.