7 January 2005

Itanic

It’s something I’ve wanted to see happen for a couple of decades now, and finally it’s happening: Intel are getting their asses handed to them by a better CPU maker, in this case AMD.

If you weren’t in the computer industry in the 80s, you might not be aware that IBM picked the x86 processor series for the IBM PC specifically because it sucked. IBM sold lots of high end workstations and word processing systems, and made good money doing so; the last thing it wanted was for the PC to eat into that market. So, they made a PC which was good enough to satisfy hobbyists, but hopefully crap enough not to compete with IBM’s serious business hardware.

What IBM didn’t foresee was companies like Compaq taking the basic PC design and improving it. Soon the PC’s memory was no longer limited to 640K, it had a real FPU, a 16 bit processor, larger and larger hard drives supported via vendor-specific extended versions of MS-DOS, and so on.

Still, by the late 80s it was clear that the x86 architecture was a mess. Even Intel engineers realized it. So, they started to design a completely new CPU with a clean, modern RISC-based design. The result was the Intel i860.

Unfortunately, the i860 was designed with an 80s PC mindset, and its performance was horrible as soon as you started to context-switch. It would have worked fine for single-tasking MS-DOS systems, but it was a dog as soon as you wanted pre-emptive multi-tasking. And by the early 90s, when Intel started pushing migration to i860, PC users had begun looking at the Apple Macintosh and Commodore Amiga and asking “Why can’t I do that?” Worse still, other RISC chips like the PowerPC were quickly outpacing the i860.

So the i860 bombed. Intel went back to trying to solve the x86s fundamental problems by throwing more and more transistors at it. Which is why we now have Pentium chips that dissipate 45W of power to run at the same speed as 15W PowerPCs.

People outside the computer industry often don’t realize that Intel aren’t really a technology company; they’re a marketing company. They’re very like Microsoft, in fact—they achieved success by being lucky enough to get picked by IBM, developed a near monopoly, and became fantastically rich. The difference is that whereas Microsoft has continued to dominate by acting illegally to preserve its monopoly, Intel has dominated by spending huge amounts of money on marketing.

The current egregious Intel marketing is “Centrino”. You’ve heard of it, and you probably think you want it, but chances are you don’t know what it really is. You might even think it’s a kind of low-power CPU. Well, it’s not–in fact, it’s just a code word. All “Centrino” really means is that the computer has a Pentium M CPU, an Intel chipset, and a motherboard with Intel wireless networking. It’s pure marketing, designed to force laptop manufacturers to buy everything from Intel instead of just the CPU. However, if the ads said “Buy a computer made entirely by Intel”, they wouldn’t be quite so compelling, would they?

Similarly, average people go into computer stores demanding a Pentium 4. They have no idea what a Pentium 4 is, let alone how it compares to the alternatives, but the saturation advertising has convinced them that they need one. When I tell people (such as my parents) that they can get an AMD system for hundreds of dollars less that’ll run all the same software faster, they find it hard to believe. If they’ve heard of AMD at all, they have some vague memory about Athlon chips being hot. (They were, when they were first released. Now they use less power than Intel Pentium 4 chips.)

In short, Intel CPUs are overpriced, oversized and tend to overheat. They are badly designed, but Intel compensates by using more silicon and more transistors to work around the bad design. Intel CPUs don’t sell on technical merit or bang-for-the-buck, and never have.

But back in the 90s, Intel decided to try again. This time they got help from HP, who had at least managed to achieve modest success with a RISC-based design in their PA-RISC “Bobcat” workstations in the 1990s, and had a next-generation CPU on the drawing board that would try to take the next step beyond RISC. Intel’s engineers got to work, and the new CPU was named Itanium.

HP also owned the DEC Alpha design. The Alpha was the fastest processor around, and was in use in every university science department where there was a need for number-crunching on the desktop. So HP killed it, perhaps out of hubris, or perhaps because of “Not Invented Here” syndrome.

Initially, forecasts for Itanium were rosy. Forecasts, remember, are made by marketing people and business people, not by engineers who know about Intel’s design history. Major vendors signed on to sell Itanium workstations and servers; SGI killed their own MIPS RISC CPU, predicting that Itanium would take over, and IDC expected $30 billion in Itanium server sales by 2001.

The big problem with introducing any new CPU is that businesses generally want to run Windows, which is built for x86. Initially, Intel’s answer to this problem of legacy back-compatibility was that software companies (i.e. Microsoft) would just have to recompile all their code for the Itanium. Of course, it’s now 2005 and Microsoft has just announced that it’s cancelling all plans for an Itanium version of Windows XP, even though OS X and Linux are both running on 64 bit hardware.

Meanwhile, AMD had used some of the design ideas from the DEC Alpha and put them into their x86 chip designs, and called the result Athlon. When the Athlon began beating the Pentium CPUs in performance, Intel quickly started shoveling an x86 compatibility mode into the Itanium design. It was a mess, and the chip’s power consumption ballooned to 130W. Itanium’s performance running x86 code was worse than the competing x86-only hardware; it was so bad that when they threw in the towel and implemented emulation instead, the emulator ran x86 code twice as fast.

Unfortunately, in the mean time AMD had added 64 bit features to their x86 design. The result was a chip with 64 bit power and full x86 compatibility with no slowdown, called Opteron.

So we reach today. Rather than shipping $30 billion of Itanium systems by 2001, Intel hoped to ship 100,000 Itanium systems in 2004. In the mean time, AMD predicted that they would have shipped 2,000,000 Opterons by the end of the year.

Update: Tony Finch tells me that the i860 wasn’t actually a total flop, and that it found a niche in embedded systems. There, it didn’t matter if it handled multi-tasking poorly, and the code could reasonably be hand-optimized in the appropriate places for the i860’s pipelined arithmetic unit—something compilers weren’t good at at the time.

Quite likely true; embedded software is one of those weird parts of the industry that I know very little about, almost its own separate world, where people run strange OSs like QNX and write programs in odd languages like FORTH. Still, a niche is a long way from what Intel wanted to achieve with the i860. I’m sure they’ll find a niche for Itanic too, but it’s still a flop when judged by what they planned it to be.

© mathew 2017