Tuesday, April 2nd 2024
Apple M3 Ultra Chip Could be a Monolithic Design Without UltraFusion Interconnect
As we witness Apple's generational updates of the M series of chips, the highly anticipated SKU of the 3rd generation of Apple M series yet-to-be-announced top-of-the-line M3 Ultra chip is growing speculations from industry insiders. The latest round of reports suggests that the M3 Ultra might step away from its predecessor's design, potentially adopting a monolithic architecture without the UltraFusion interconnect technology. In the past, Apple has relied on a dual-chip design for its Ultra variants, using the UltraFusion interconnect to combine two M series Max chips. For example, the second generation M Ultra chip, M2 Ultra, boasts 134 billion transistors across two 510 mm² chips. However, die-shots of the M3 Max have sparked discussions about the absence of dedicated chip space for the UltraFusion interconnect.
While the absence of visible interconnect space on early die-shots is not conclusive evidence, as seen with the M1 Max not having visible UltraFusion interconnect and still being a part of M1 Ultra with UltraFusion, industry has led the speculation that the M3 Ultra may indeed feature a monolithic design. Considering that the M3 Max has 92 billion transistors and is estimated to have a die size between 600 and 700 mm², going Ultra with these chips may be pushing the manufacturing limit. Considering the maximum die size limit of 848 mm² for the TSMC N3B process used by Apple, there may not be sufficient space for a dual-chip M3 Ultra design. The potential shift to a monolithic design for the M3 Ultra raises questions about how Apple will scale the chip's performance without the UltraFusion interconnect. Competing solutions, such as NVIDIA's Blackwell GPU, use a high-bandwidth C2C interface to connect two 104 billion transistor chips, achieving a bandwidth of 10 TB/s. In comparison, the M2 Ultra's UltraFusion interconnect provided a bandwidth of 2.5 TB/s.
Source:
HardwareLuxx
While the absence of visible interconnect space on early die-shots is not conclusive evidence, as seen with the M1 Max not having visible UltraFusion interconnect and still being a part of M1 Ultra with UltraFusion, industry has led the speculation that the M3 Ultra may indeed feature a monolithic design. Considering that the M3 Max has 92 billion transistors and is estimated to have a die size between 600 and 700 mm², going Ultra with these chips may be pushing the manufacturing limit. Considering the maximum die size limit of 848 mm² for the TSMC N3B process used by Apple, there may not be sufficient space for a dual-chip M3 Ultra design. The potential shift to a monolithic design for the M3 Ultra raises questions about how Apple will scale the chip's performance without the UltraFusion interconnect. Competing solutions, such as NVIDIA's Blackwell GPU, use a high-bandwidth C2C interface to connect two 104 billion transistor chips, achieving a bandwidth of 10 TB/s. In comparison, the M2 Ultra's UltraFusion interconnect provided a bandwidth of 2.5 TB/s.
26 Comments on Apple M3 Ultra Chip Could be a Monolithic Design Without UltraFusion Interconnect
The flaw, called GoFetch, is based on the microarchitecture design of the Apple Silicon, which means that it cannot be directly patched and poses a significant risk to users' data security. The vulnerability affects all Apple devices powered by M-series chips, including the popular M1 and M2 generations. The M3 generation can turn a special bit off to disable DMP, potentially hindering performance.
www.techpowerup.com/320693/apple-m-series-cpus-affected-by-gofetch-unpatchable-cryptographic-vulnerability
I'd presume that the memory usage and management is entirely different than w/ a Windows PC. (Kinda like how the iPhones got/get away with so little RAM vs. Android devices)
My issue w/ the latest MacBook Air is, the thing is a lil hotbox. Many board-level components have efficiency and MTBF 'curves' based on their thermal operating conditions.
What I'm getting at: The design is implicitly 'eating into' the thermal tolerances of everything inside the Air.
Like everything Apple makes, it's built to fail in a highly 'calculated' manner, and 'technically excusable' once the issues become known.
Unfortunately when you need to work with other ecosystems or use their ecosystem in a "non-approved" way that you start to run into issues. And from a business/enterprise perspective, trying to get support for macOS from IT is a pain. "We don't know how to use or support Macs".
And I say this as the only Mac-savvy person in my office (which is a curse, believe me).
One of my computers is the M1 MacBook Pro 13” 8gb with 256GB storage.
I am using so many Firefox windows with tons of tabs it is ridiculous.
Machine learning flies and I rarely need to transfer tasks to my M2 Max 32GB studio or Amazon EC2 instances.
The only time I have hit issues with the ram is when I try to do a stupid Python data processing that tries to load all the input AND output data in RAM. Using more intelligent methods like chunking etc. solves this while Firefox, safari, vlc, vscode, jupyter etc are running.
I am always astonished how I have practically no RAM issues considering the 8GB.
Have I build my Windows and Linux PCs, FreeBSD ZFS NAS boxes and OpenBSD routers/firewalls? Yup. But having a TOOL that works without me having to do a thing and allows me to concentrate 100% on the actual task I am doing is fantastic.Use the right tool for the right job.
Would you use OpenBSD for desktop publishing or serious audio/video work?
Would you use a Windows PC for a router/firewall/VPN Concentrator?
Would you use a FreeBSD box for gaming?
We might want to have a tool that does it all 100% and yesterday, but the reality is using the best tool and OS for the job is a great way to get things done.
Overall fuck Apple. I hope the DOJ rips them apart.
I hope the DOJ fails, the arguments are pitiful. If you don't like it, don't buy it. If you don't care about it, don't swarm Apple posts. Simple.
Macs make my development work easier and smoother. Quantifying the “creative” process is difficult.
When the problem only needs more performance/brute force as in more cores or terabyte+ RAM i can just spin up EC2 instances and pass the dumb processing there.
Creatively solving problems works much better for me on macOS. And I use various other operating systems: Windows, various Linuxes, OpenBSD and FreeBSD.
Also, why are people so upset that others make different purchasing and equipment choices?! Apple, Nvidia, Microsoft, Amazon etc offer various products and everyone is free to buy whatever one seems good for the specific use case.
For some a maxed out 96 cores Threadripper workstation with a terabyte of RAM is the best purchase, for another a Mac is the best purchase. Both can be 100% right depending on the specific use cases.
If someone needs the absolute best single thread performance for a specific use case an i9-14900KS with 1-2 performance cores enabled only and OC applied might be the best choice bar none.
I didn't want to pay for extra storage - so I went with the OWC 1M2 NVMe M.2 SSD enclosure and slapped a 4TB WD SN850X in it for an additional $325 total.