Wednesday, July 5, 2017

Steve Jobs and the Missing “Intel Inside” Sticker

Ken Segall:

The Intel Inside marketing strategy will be studied in business schools around the world for decades. It represented bold thinking and bold spending.

[…]

Apple’s internal testing showed that the newest PowerPC processor was faster than Intel’s fastest chip. With a real competitive advantage to work with, we did what any feisty agency would do: we declared war on Intel.

Suddenly, it was to our advantage that Intel had become the unifying, driving force in PCs. We didn’t have to attack any PC maker by name—we could take on the entire PC industry simply by attacking Intel.

Apple did a lot of Photoshop demos. My recollection is that aside of graphics tasks that emphasized floating point, PowerPC-based Macs were mostly slower than Intel-based PCs. This was especially noticeable with compilation. The stated reason for the Intel transition was performance per watt, but by that time the high-powered PowerPCs were behind schedule, too. When I got the original Core Duo iMac in 2006, its performance blew away the dual-G5 tower that I had been using, even though the Core Duo was derived from the mobile Pentium architecture. So of course it was great in notebooks, which were still stuck using G4s.

11 Comments RSS · Twitter

"The stated reason for the Intel transition was performance per watt, but by that time the high-powered PowerPCs were behind schedule, too."

IIRC, what was going on was that no one was willing to invest the big bucks in PowerPC development necessary to keep up with Intel's capital expenditures. This is why PowerPC was behind schedule, but the implications of the lack of capital expenditure went far beyond mere delays in schedule...

@Chucky That’s definitely a big part of what was going on, though I’m not sure Apple was that blunt about it.

Guy English

The long pipelines on the G5 made it slower at changing what it was doing, while giving it an advantage with inflight instructions. Branches were bad and could require a full pipeline flush (some of which had already started running speculatively). This setup is terririfc for SIMD operations like bitmap operations. Same way GPUs are today. But if you had to do a lot of branches (like it you're compiling) you'd waste an awful lot of what made the G5 fast. One theory was that smart compiler optimizations for generated code would avoid these problems. To some extent they did but never enough before the other issues with the PowerPC line made it pointless to keep pursing.

@Guy Thanks for filling in those details. How was the G5’s deep pipeline an advantage when it only resulted in a clock rate that was about the same or lower than Intel’s (hence Jobs talking about the Megahertz Myth)? Did it enable it to be wider?

The thought was that deeper pipelines enabled higher clock speeds, and the higher clock speeds would more than make up for the drawbacks of deep pipelines (i.e., branch misprediction penalties). The problem was that no one (neither IBM nor Intel) were able to achieve the higher clock speeds at a reasonable power or temperature envelope.

If I recall, the Megahertz Myth was more of a G4-era thing to deflect from its lower-than-intel clock speeds, whereas the original G5 marketing was just the Photoshop/Logic demos and Apple claiming it was the world's fastest personal computer (and then having to retract that marketing copy in the UK).

I had that original 2006 Core Duo iMac too. Best deal on a Mac I ever got, at least as far as performance per $.

@remmah Yes, Apple definitely started talking about the Megahertz Myth before the G5 era, and the G4 had a much larger megahertz gap than the G5 did. That iMac was great, but mine completely stopped working shortly after the Apple Care expired—the only Apple computer that’s ever done that for me. Still, it served its purpose well.

"If I recall, the Megahertz Myth was more of a G4-era thing to deflect from its lower-than-intel clock speeds"

IIRC, the whole Megahertz Myth was not just to "deflect". The PowerPC really did perform far in excess of what a clock speed comparison with Intel chips would indicate due to the RISC/CISC difference, at least for most tasks where the software had been engineered to take advantage of the difference. And given the adoption of RISC in smartphone CPU's, it seems it was the correct architecture going forward. Again, I think Apple's move to Intel was sheerly due to the lack of capital investment in PowerPC development, which Apple foresaw had no viable solution. Though I suppose the ability to promote Macs capable of running Windows natively played some pretty minor part in the shift.

(Also, regarding Michael's reliability issues with his first Intel Mac, I also had severe reliability issues with two separate first-gen Intel Mac laptops, and IIRC, think reliability issues with early-gen Intel Macs was pretty widespread, which I'd speculate was due to Apple's learning curve with heat dissipation issues it hadn't had to face before.)

@Chucky IIRC, PowerPC did outperform on graphics tasks, but that wasn’t because of CISC vs. RISC. It was because the PowerPC had more floating point units, plus AltiVec. The chips were optimized for different tasks. Plus, as Guy notes, some of the RISC elements like the deeper pipeline made the PowerPC slower for more general tasks.

In my view, the RISC choice wasn’t really vindicated in the obvious way. One of the big promises of RISC was increased clock rates. You do less work per cycle but more than make up for it through faster cycles. Yet despite being RISC, the PowerPC was still behind in clock rates. Intel had higher clock rates and complex instructions that did more per cycle. And RISC was supposed to be better for power, but Intel was ahead there, too.

Newer Intel chips are essentially RISC on the inside, plus a translation layer so they can handle the old instruction set. I’m not sure whether this shows that RISC is better or that the instruction set doesn’t really matter that much.

Also, the PowerPC was not really a true RISC design, and x86 is not really as CISC as what RISC was a reaction against. ARM is an entirely different beast, so I don’t think its success says much about what the PowerPC could have been.

Thanks for all that clarification, Michael. Obviously far more complex than I understood.

> some of the RISC elements like the deeper pipeline

The G4 - which is, I think, the chip Apple was using when they were talking about the "Megahertz Myth" - had a pretty short pipeline, though. I think early G4s had a four-stage minimum pipeline. The pipelines only really got substantially longer with the G5.

Funny that we're still discussing this stuff :-)

Leave a Comment