Obviously Apple should kick nvidia to the curb as soon as they can use their own chips to achieve the same thing. But until then, yeah Apple should pay Nvidia (or AI partners that use Nvidia) whatever it takes to stay in the game. Within just a few years Apple will figure out how to use their own chips to do everything they (may) currently need Nvidia (or AI partners that use Nvidia) for. Generative AI is here to stay and the march towards general AI well never stop.
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 to Power AI Features With M2 Ultra Servers
- Thread starter MacRumors
- Start date
- Sort by reaction score
It will be interesting to see what AAPL does with the M4 Ultra Server chip…. They could end up following what NVIDIA has done with the Grace/Hopper superchip - GH200…. Grace (72c ARM w/ 480GB DDR5) CPU and a Hopper GPU w/ 144GB HBM3e memory…. Scale that up with a robust memory coherency fabric and they could have a winner.
Compare that to Tesla Dojo.. which still hasn’t seen the light of day (i.e. running production)…. Tesla still buying a _ton_ of H100
Compare that to Tesla Dojo.. which still hasn’t seen the light of day (i.e. running production)…. Tesla still buying a _ton_ of H100
That's one of those "l'll buy anything if its shiny and made by Apple" thought processes going on right there.
Ignore him he knows not what he speaks. As someone who's spent their entire career in data centers and recently doing Ai workloads working for a large hyperscaler the thought of using Apple hardware in that environment is laughable
People aren't looking at the bigger picture why Nvidia is so dominent in Ai. Look at why the DoJ blocked their purchase of ARM shortly after they acquired mellonox. Ai workloads are done in very large clusters consisting of GPUs in the thousands or tens of thousands using DLB and RDMA. All of which require infiniband NICs made by you guessed it; Nvidia.
It's the NICs that make Ai supercomputers possible in conjunction with Nvswith and link possible. Repairability aside of a socketed ram/cpu (2 of the most common falires) they would still need a raft of Nvidia equipment even if they used mi300x
NVIDIA is the leader, but time will tell if AMD (MI300X) and Intel (Gaudi3) for instance can chip away at their market share in any meaningful way.
Infiniband isn’t required per se, much to NVIDIA’s chagrin….other high-speed RDMA fabrics are plenty capable. Heck, three of the top five leadership class exascale systems use Cray Slingshot (200Gb) interconnect, which is based on Ethernet.
But Apple is not looking at using M2 Ultra servers for b2b training or b2b inferencing. They will apparently use it for b2c training and inferencing. Training meaning to get to know the user and keeping their data secure. For more general training, Tim Cook hinted in his earnings call last week that they are currently using hybrid solutions by tapping AWS, Google Cloud or other AI clouds. Once UXL Foundation releases the open source UXL stack for AI server GPU's by year end then Apple may be able to replace the need for third party clouds with Apple Silicon GPU's. At no time did Tim Cook hint that they will be adopting NVIDIA's proprietary Cuda platform. In fact, it appears they are laying the foundations to steer clear of it. Mr. Jobs taught him well to recognize "a bag of hurt" and drive around it.
Last edited:
You're correct that it isn't required but every bit counts and ethernet has too much overhead or so I'm told.
Mi300x is a niche product in that is has higher ram capacities which makes it good for specific tasks but AMD doesn't have volume yet. He'll Nvidia barely does. Hopper is effectively sold out and last I hear has a 50+ week lead time
They're using them because their product is not equipped for real non photography or capture work.
Power efficiency is great but again they can't scale and don't have a robust enough PCIE to do anything but be an expensive toy. The only reason why Intel is going well in servers is because they have more lanes than AMD and for certain rack shapes we need all of them.
Lastly the speed of which we roll through dimms and cpus would make using apple silicon at scale impossible from cost perspective
Yes, but the more important point is that they are using them in hybrid fashion on purpose because they do not want to purchase any Nvidia GPUs and commit iCloud farms to Cuda stack. They’re using them in hybrid fashion, so that the third party cloud, if it does not drop costs when UXL -based servers become available within a year or so, can be replaced easily with with other lower cost UXL cloud systems. They won't be just a little cheaper, but a lot cheaper than the 80% margins NVIDIA is asking.
What’s the point in committing to an expensive and proprietary stack that locks you into a single platform when they can use hybrid for now and have the freedom to switch up to any open source UXL based AI cloud that is equally powerful but much, much cheaper within a year or so?
When UXL stack for GPUs is released, it is also possible that Apple is not planning on switching up to the lowest priced AI clouds at all. They may be able to roll their own GPUs at that point and simply join the open source UXL Foundation to advance the platform faster as happened with Linux for server CPUs.
Apple Silicon and NVIDIA are both based on ARM. Apple Silicon team is entirely capable of catching up to NVIDIA with a little help from UXL Foundation. In fact, Apple's server GPUs may be even more efficient because they only have to deal with NPUs on Apple devices, not third party devices. Their balance of edge AI to cloud AI can be fine tuned dynamically to meet the needs of their own devices.
Overall, it makes no sense for Apple to commit to Cuda. Just a bag of hurt. It would be short term gain, for much longer term pain.
Last edited:
Were talking in circles so appologies; this isn't an argument for against Cuda this is specifically from a servicesbility stance. Ampere is a serviceable arm platform, so is amd and so is Intel. Apple silicon is not. What exactly makes AS more effective than ampere or anything oryn based? It's not if it's paired to AMD or Nvidia
Apple has over $100 billion cash on hand. They are able to make acquisitions in AI server sector if required. They appear to have their targets set on AI servers. Let's see how they do it.
