Apple is designing more powerful chips rather than just end it with the M4 series?
I am shocked, SHOCKED I say!
I am shocked, SHOCKED I say!
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Working on Chips for New Macs and AI Servers
- Thread starter MacRumors
- Start date
- Sort by reaction score
They will need to make a chip for at least Mac Pro in order to make server grade Apple Silicon chip but currently, SoC design is the worst for that. Nobody use SoC for server and super computer like Nvidia and Intel. They were just thinking to connect like MCM or Ultra but it turns out it was a failure and they ditched Extreme chip by connecting 4x Mac chips.
Since Apple is struggling with AI development as they only have 50,000 of 5 years old GPU, they really need a powerful chip.
Since Apple is struggling with AI development as they only have 50,000 of 5 years old GPU, they really need a powerful chip.
Apple's chip design to the rescue, we'll never have to sleep again.
Errr, there is very little to indicate that Apple is building general purpose 'server chips'. The rumors have focused on very AI specific workload chips. That is not what Nuvia was doing. Nuvia was making a "Ampere Computing" and Amazon Graviton competitor.
In the current Qualcomm chips, all the "AI" tech is derived from Qualcomms already had. ( Just like the graphics subsystem). Nothing "Nuvia team" is making a big dent there.
For general purpose serving the generic (Linux) centric parts of Apple web services where an "Apple server" chip would make much difference from what Ampere eventually rolled out ( and Graviton is readily available for the outsourced Apple web services. )
The notion that Apple was going to hold all of its talent long term is deeply flawed. Other members have run off to RISC-V projects and left Apple also. Arm itself is picking up the pace of more performant cores (e.g., Neoverse server cores that several are using ; Ampere (for a while), Microsoft, Nvidia , etc. ). Nvidia also could have thrown money at folks (which is really part of the "do server chips" play. Qualcomm paid a giant bucket of money for Nuvia. Getting 'bought out' likey was always part of the plan.)
As long as this "AI server chip" is a 'behind the curtain' hardware solely for Apple Intelligence, it really isn't going to make much of a difference until Apple Intelligence as an overall system is competitive. Software and Hardware.
There is zero indications that Apple is trying to go into the general purpose web services business ( Microsoft Azure, Amazon AWS , Google Cloud , etc. ) business. Pretty good chance this "AI Server" chip never shows up in a product that is for sale at retail.
Not really 'decade behind' AWS, Azure , or G-Cloud if simply just pulling workload in house and/or creating new AI inference workloads that didn't exist before.
This story dovetails with Apple spending $500 billion over the next four years manufacturing AI servers that will be distributed across the US.
The Mac Pro is primarily a single user workstation. It pragmatically needs to be able to drive multiple diplays so that the single user can see the GUI. Very good chance that Apple's "AI server chip" has no display function units. First, Apple display processing units are abnormally big. Likely space die trade off will lean more toward the AI compute need to do rather than the display task absolutely not going to do in a data center. Basically same thing for Thunderbolt subsystem. More likely traded for the rumored Broadcomm internode communication system.
A Mac Pro with no display process and no Thunderbolt won't meet the modern criteria for being a "Mac" anymore.
Errr.. Nvida?
A Quick Introduction to the NVIDIA GH200 aka Grace Hopper
The NVIDIA GH200 or "Grace Hopper" is far from a single product. We have a quick guide so when someone says "GH200" you know what to look for
Apple's solution is much more rigidly "Homogenous Unified Memory" but others are on the MCM and/or multiple die path also. There is no DIMMs slots for mega-modular memory there.
Nvidia SuperChips are a Failure? really???? The longer term trendline toward far more power efficient AI compute is heading directly at chiplets and/or MCM.
Apple's possible resources are not limited to that. 'AI' is not equated with solely just large as possible language models. Apple has some self imposed constraints that go beyond avaialble GPUs. Privacy , not streamrolling copyrights. not hovering up as much customer prompt data a possible. etc. Apple has been more focused on inference on modest resources than on powering the most monstrous MW consuming machine possible.
Back to Mac Pro ... it has been capped at normal USA wall socket power levels. Not 'brown out' the neighborhood power levels.
Apple's AI server chip is far more likely to be the most power efficiencent (Perf/W ) of its immediate rivals far more so than the "Most Powerful" option.
A lot of yapping saying nothing. Facts are facts.Qualcomm bought Nuvia explicititly to challenge Apple ARM chips. Server-wise, you make it seem like server plans were set and would remain static and never changing, where they would clearly know industry wide plans for server chips. .
No need, as Luca Maestri has found a stack of pentium chips in a garage sale for a very reasonable price, so Apple can use these instead.
I hope they allow to configure top of the line GPU cores available. Macbook pros, only lack in GPU, else everything is awesome!. I know how much they can pack in a chassis so thin, but a little thicker is fine if i'm able to configure, 96 gpu cores and above or even 128 GPU
Another misstep by current Apple leadership. Hopefully, it doesn't come back to bite them.
Yaaaawwn ... People still care about AI ? Can't wait to ask it about how my NFT blockchain is doing in the metaverse.
These are not server CPU’s in the normal sense ie like Xeon. These are specialist chips that will go into AI Servers dedicated for handling AI services in the Cloud.
Don’t expect OSX Server and xServe to be making a return with this.
In this case they are adapting previous designs, so this is more like 3 years, instead of working from scratch
But would they have as good mobile chips today if their efforts were split onto the servers chips? Can you make this assumption on this single fact?
I mean a big chunk of the original Apple Silicon team left Apple to develop it on their own. They lost the same amount of people either way. And the Nuvia team was posting about their server chips leapfrogging Apple in performance and power efficiency before being acquired by Qualcomm.
It was a huge three pronged short sighted management failure. 1. Lost a big chunk of top Apple Silicon engineers, 2. Gave competition, Windows and Qualcomm a nice post Intel and X86 gift and sadly and worst of all 3. They needed a server and AI chip development team anyway. Their short sightedness left them behind in AI and the server space that unlike Apple, everyone else could clearly see.
Last edited: