★ The Information Suggests, in an Aside, That Apple Scrapped Work on a Quad-Max/Double-Ultra M-Series Chip

• Facebook
• Tweet
• Pinterest
• Reddit
• Email

Reporting from The Information suggests that an M-series tier above Ultra might remain years away.

Wayne Ma and Qianer Liu, in a piece today for The Information (paywalled up the wazoo, sadly), “Apple Is Working on AI Chip With Broadcom”:

Apple is developing its first server chip specially designed for
artificial intelligence, according to three people with direct
knowledge of the project, as the iPhone maker prepares to deal
with the intense computing demands of its new AI features. Apple
is working with Broadcom on the chip’s networking technology,
which is crucial for AI processing, according to one of the
people. If Apple succeeds with the AI chip — internally
code-named Baltra and expected to be ready for mass production by
2026 — it would mark a significant milestone for the company’s
silicon team. […]

Broadcom typically doesn’t license its intellectual property,
choosing instead to sell chips directly to customers. In its
arrangement with Google, for instance, Broadcom translates
Google’s AI chip blueprints into designs that can be manufactured,
oversees its production with TSMC and sells the finished chips to
Google at a markup.

But Broadcom appears to be taking a different tack with Apple.
Broadcom is providing a more limited scope of design services to
Apple while still providing the iPhone maker with its networking
technology, one of the people said. Apple is still managing the
chip’s production, which TSMC will handle, another person said.
Additional details of the business arrangement couldn’t be learned
[sic]1

I’ll go out on a limb and say that it’s Apple choosing to take a different tack with Broadcom than Google did, rather than a choice in any way driven by Broadcom. The Information’s own “arrangement with Google” link above points to this year-ago report that opens: “Google executives have extensively discussed dropping Broadcom as a supplier of artificial intelligence chips as early as 2027, according to a person with direct knowledge of the effort. In that scenario, Google would fully design the chips, known as tensor processing units, in-house, the person said. The move could help Google save billions of dollars in costs annually as it invests heavily in AI development, which is especially pricey compared to other types of computing.” Why would Apple ever agree to an arrangement like that?

The hint of obsequiousness to Broadcom suggests to me, pretty clearly, that it’s sources from Broadcom who provided the leaks for this story.

Anyway, what really caught my eye in this report wasn’t the AI server chips, but rather the following (emphasis to key paragraph added), included seemingly only as an aside even though I thought it was the most interesting nugget in the report (vague shades of Fermat’s Last Theorem):

Apple’s silicon design team in Israel is leading development of
the AI chip, according to two of the people. That team was
instrumental in designing the processors Apple introduced in 2020
to replace Intel chips in Macs.

Apple this past summer canceled the development of a
high-performance chip for Macs — consisting of four smaller chips
stitched together — to free up some of its engineers in Israel to
work on the AI chip, one of the people said, highlighting the
company’s shifting priorities.

To make the chip, Apple is planning to use one of TSMC’s most
advanced manufacturing processes, known as N3P, said three
people with direct knowledge. That would be an improvement over
the manufacturing process used for Apple’s latest computer
processor, the M4.

What they’re talking about regarding a cancelled high-end Mac chip would be a hypothetical M-series chip with (effectively) double the specs of an Ultra, which I presume would only be available in a future Mac Pro, and, just pulling adjectives from Apple’s marketing dictionary, I’d bet would be called the “M# Extreme” (where “#” is the M-series generation number). The M1 and M2 Ultra chips are, effectively, two M1/M2 Max chips fused together with something called a silicon interposer that offers extremely high-speed I/O between the fused chips. Performance doesn’t exactly double, but it comes close. A hypothetical quad-Max “Extreme” would effectively double the performance of the same-generation Ultra chips. Such a chip, available exclusively in the Mac Pro, would give the Mac Pro a much more obvious reason to exist alongside the Mac Studio (which, to date, has offered Max and Ultra chip configurations).

But if Apple’s work on that quad-interposed M-series chip was cancelled only “this past summer”, and was for a generation of chips using TSMC’s next-generation N3P process, that would mean it was slated for the M5 or M6 generation, not the M4.2 The M4 generation is fabbed using TSMC’s N3E process, and any additional variants beyond the M4 Max, slated for updated Mac Studios and Mac Pros next year, were designed long before this past summer.

I feel like it’s a lock that there will be an M4 Ultra chip next year, with the performance of two M4 Max chips fused together. Or, perhaps the M4 Ultra will be a standalone design, not two Max chips fused. The M-series Max chips have always been their own designs — not two Pro chips fused together. The same could be true for Ultra chips, starting next year, or some generation further into the future.

But I’ve had my fingers crossed that we’ll also see an “M4 Extreme” — or whatever Apple would decide to call a tier above “Ultra” — sooner rather than later. If The Information’s reporting is correct, however, either we’ll see a quad-Max M4 chip next year, and then it will skip a generation because the engineering team was redirected to work on these AI server chips, or, those engineers were working on the first quad-Max M-series chips, and now the first such M-series chips have been punted even further into the future, if ever. Today’s report has me thinking, sadly, that could be a few years off, at the soonest.

Read More