Why Project Aquarius Vanished: What Happened to Apple’s First CPU?

本站添加: 2026-05-23

[00:00:00]In 1986, inside a separate building in Cupertino called Valley Green 3, Apple installed a $15 million Cray X-MP 48 supercomputer.

[00:00:11]The site was prepared in only 6 weeks, one of the fastest Cray installations on record up to that point.

[00:00:18]Apple CEO John Sculley had authorized the purchase, and according to Cray's own in-house magazine, he made the first call to the sales team himself.

[00:00:27]The machine was there to help over 50 engineers design Apple's first internally developed central processing unit.

[00:00:34]The project was called Aquarius. It would never produce a working chip.

[00:00:39]You were watching the beginning of the RISC era. The Berkeley RISC papers had landed. Stanford had produced the work that would become MIPS. The MIPS R2000 had appeared in 1985.

[00:00:51]Acorn in Cambridge had begun shipping the original ARM inside the Archimedes.

[00:00:56]Sun was finalizing SPARC. AT&T was deep in the CRISP design.

[00:01:02]Hewlett-Packard was preparing PA-RISC.

[00:01:05]The pattern was already visible to anyone working close to the silicon.

[00:01:09]The winning move was to strip the instruction set down, lean on the compiler, and build the chip around a specific host machine.

[00:01:17]You knew this because you had seen the comparative benchmarks. You had read the Patterson and Hennessy papers. The small teams kept shipping working silicon while large teams at established semiconductor firms struggled to match their results.

[00:01:32]Acorn had taken the original ARM from architectural concept to working silicon in roughly 18 months with a team you could fit in a single conference room.

[00:01:41]Their simulation rig was a network of BBC micros. The first prototype chip worked the day it came back from the foundry. Apple was about to attempt the opposite of all of that with more money and more engineers and less to show for it.

[00:01:56]The Aquarius effort had three internal names. Aquarius referred to the overall project. The architecture specification was called Scorpius, and Harys was the planned initial implementation.

[00:02:07]Jean-Louis Gassée, head of the Macintosh division, championed the effort within senior management. Sam Holland, an Apple engineer whose background was in supercomputer design, was placed in charge.

[00:02:20]The architecture targeted four cores on a single die. It included SIMD execution for parallel data processing across the cores.

[00:02:28]It also included an interprocessor bus designed for tight communication between processing units.

[00:02:34]The published Scorpius specification, dated January 1989 and now archived on the Internet Archive, describes a detailed multi-core design with 16 general-purpose registers per processing unit, seven local registers per unit, and eight global registers shared across the chip.

[00:02:54]The team at Project Deep reached over 50 engineers working out of a building isolated from the rest of Apple's campus. The Cray X-MP 48 alone carried a price tag of 15 million 1986 dollars.

[00:03:08]Its top-end configuration provided four CPUs with a theoretical peak performance of over 800 megaflops and eight million words of memory.

[00:03:17]Site preparation took six weeks. The machine became, by Apple's own claim, the first Cray system in the world to run Unix, specifically Unix System 5, integrated with engineering help from AT&T. The Cray itself was connected through a hyperchannel link to a VAX 11/785, a VAX 11/780, and a Sun 2 workstation, with the VAX systems and Sun bridged, in turn, to a network of Macintoshes over AppleTalk and Ethernet through a Kinetics converter.

[00:03:48]The internal infrastructure built around the Aquarius effort was the largest internal engineering network Apple had assembled to that date. The premise behind the spending was straightforward in the boardroom. Steve Jobs had departed Apple in 1985 after his power struggle with Scully. Scully was running the company alone. Gassée and Holland had convinced him that Apple needed to control its own silicon roadmap rather than depend on Motorola's 68000 line as it slowed.

[00:04:18]Holland's specific argument, as Gassée later recounted, was that a quad-core Apple design processor would dramatically outpace the evolution of Intel and Motorola chips for years to come and lock in the Macintosh position at the top of the personal computer industry.

[00:04:36]Scully accepted the argument. Funding followed quickly. The Cray purchase was justified to outside investors as a modeling tool for future Apple hardware, which, in the most generous reading, was technically true.

[00:04:48]Access to the Aquarius building was restricted to project staff only. The internal name traveled inside the company only as far as it needed to.

[00:04:56]Cost overrun risk was absorbed entirely by Apple's then strong Macintosh and Apple II revenue.

[00:05:04]The obvious explanation for the failure is that Apple was not a semiconductor company and did not have the manufacturing depth of Intel or Motorola. The obvious explanation is incomplete.

[00:05:15]The deeper force was a cultural import that sat inside the architecture from the first specification.

[00:05:21]Sam Holland came from supercomputers.

[00:05:23]The features he proposed, quad-cores, SIMD execution, multiprocessor parallel modes, an interprocessor bus, integrated graphics support, were supercomputer features rendered in microprocessor form.

[00:05:38]The Cray purchase was not just an expense. It was diagnostic.

[00:05:42]It signaled that Holland was solving a supercomputer design problem at the scale of a single chip.

[00:05:48]Every other risk project of the period was winning by doing the opposite.

[00:05:52]Acorn designed the original arm with about a dozen engineers and used a network of BBC micros for simulation.

[00:06:00]The Berkeley risk team produced working silicon with graduate student labor and academic budgets. Stanford did the same.

[00:06:08]The defining property of the risk era was that small teams could win by removing instructions, simplifying instruction decode, building deep pipelines, and pushing complexity into the compiler. Apple kept adding a features.

[00:06:23]There was a second related problem that compounded the first.

[00:06:27]Aquarius had no target machine. Acorn designed arm specifically for the Archimedes. Sun designed Spark for the Sun 4 workstation. MIPS designed the R2000 for a specific workstation product line. Motorola designed the 88000 for known mini computer customers. Aquarius was designed for future Apple personal computers that did not yet exist and had not been specified anywhere in writing.

[00:06:53]With no host machine to constrain the design, scope expanded. As scope grew, the architecture acquired more features rather than fewer.

[00:07:02]The team of 50 engineers could not converge on a final implementation.

[00:07:06]Steve Sakoman, an Apple engineer who would later move on to lead the Newton effort, urged Gassée to cancel Aquarius early.

[00:07:15]Sakoman was ignored. By 1988, after two years of work in a Cray running expensive simulation cycles around the clock, the project had produced no test silicon and no agreed implementation plan.

[00:07:29]The early stack architecture had been abandoned in favor of a risk design, but the new design carried so many additional features that it had lost the simplicity advantage that defined the rest of the risk field.

[00:07:41]The Cray itself revealed an awkward fact in late testing. The Sun workstations attached to it could run the team's simulation workloads at roughly the same speed as the Cray itself.

[00:07:52]Sculley ran out of patience in 1988.

[00:07:55]Holland was removed and replaced with Al Alcorn, designer of Pong at Atari, the engineer who had hired the young Steve Jobs at Atari as a technician in 1974, and a by 1985, an Apple fellow holding the company's most senior engineering title.

[00:08:13]Alcorn brought in microprocessor expert Hugh Martin from Ridge Computers, a workstation maker that had built one of the early commercial 32-bit RISC processors.

[00:08:24]Martin reviewed the Scorpio specification and described it to both Alcorn and Sculley with one word, ridiculous.

[00:08:32]His recommendation was that Apple stop trying to compete with Intel and Motorola on raw silicon design, and instead use an existing CPU for the next Macintosh generation, allocating its engineering effort to operating system software and industrial design.

[00:08:50]Areas where Apple actually held competitive advantage. Sculley accepted the verdict in 1989.

[00:08:56]Aquarius was canceled. Hugh Martin was assigned to lead a new project called Jaguar, building a multi-processor workstation around four Motorola 88100 chips.

[00:09:08]Jaguar produced no shipping product. A parallel project led by veteran engineer Jack McHenry called Cognac, took the same Motorola 88100 and built a heavily modified Macintosh LC around it, running classic Mac OS.

[00:09:24]The Cognac team produced working hardware. The internal name was the RISC LC. The board ran classic Mac OS applications through emulation of the 68000 instruction set, demonstrating that an existing risk CPU could carry the Macintosh forward without an Apple design processor underneath it. That machine became the direct ancestor of the first Power Macintosh, which shipped in 1994 on the PowerPC 601.

[00:09:53]A chip designed not by Apple alone, but by the AIM Alliance of Apple, IBM, and Motorola.

[00:10:01]The decision Hugh Martin had recommended in 1989, let someone else design the silicon, became Apple's actual strategy for the next two decades.

[00:10:12]The Cray XMP 48 outlived the project that justified its purchase.

[00:10:18]With Aquarius canceled, the supercomputer at Valley Green 3 ran at far below capacity.

[00:10:24]Apple's industrial design group eventually used the machine for plastic injection mold simulations on Macintosh enclosures.

[00:10:32]A small graphics team used it to model video compression algorithms, work that fed directly into the vector quantization decoder shipped with QuickTime 1.0 in 1991, and into the tree-structured encoder used in the virtual museum project. The most visible product of the Cray during this period was the rendered credit roll on a short computer animated film called Pencil Test, produced using the machine's then novel frame buffer hardware. Seymour Cray, founder of Cray Research, reportedly commented at the time that Apple was using a Cray to design its next computer, while he was using an Apple to design his next Cray.

[00:11:11]The remark traveled inside both companies. It captured the inversion at the center of the Aquarius story.

[00:11:18]Supercomputer hardware solving a microcomputer design problem, while the supercomputer architect reached for a Macintosh to do work that did not require a Cray at all.

[00:11:28]The fact that the Aquarius engineers eventually discovered their attached Sun workstations could run the same simulation jobs at roughly the same speed as the Cray only sharpened the lesson in retrospect. Apple's first attempt to design a central processing unit ended in 1989 without producing test silicon. The Scorpio specification survives as an archive document dated January 1989 available on the Internet Archive.

[00:11:56]After the cancellation, the Cray was reassigned to graphics industrial design and animation work and was eventually retired.

[00:12:04]The 50 engineers were dispersed across other projects inside Apple including the Newton tablet effort that Sakoman had moved to Bubb Road to lead.

[00:12:12]Apple did not return to designing its own CPU as a primary effort for over 20 years.

[00:12:19]The company joined the AIM alliance for PowerPC in 1991, transitioned the Mac to Intel x86 in 2006, and only began shipping its own central processor designs in volume after acquiring PA Semi in 2008, and shipping the A4 in the original iPad in 2010.

[00:12:41]The first Apple designed Mac CPU, the M1, arrived in November 2020, 34 years after Sam Holland walked into John Scully's office in Cupertino with a quad-core RISC proposal and a $15 million Cray XMP 48 was crated into a building called Valley Green 3.