Athlon 64: How AMD turned the tables on Intel

This, if intel's compilers and architecture had been stellar and provided a x5 or x10 improvement it would have caught on. However no one in IT was fool enough to switch architectures over a 30-50% performance improvement that require switching hardware, compilers, and software and try to sell it to their bosses.

> if intel's compilers and architecture had been stellar and provided a x5 or x10 improvement it would have caught on.

That sounds like DEC Alpha to me, yet Alpha didn't take over the world. "Proprietary architecture" is a bad word, not something you want to base your future on. Without the Intel/AMD competition, x86 wouldn't have dominated for all these years.

I dunno if you meant it this way, but I've heard waaaay too many people say things like this meaning "if Intel compiler guys didn't suck...". They didn't, and don't (Intel C and Fortran compilers are to this day excellent). The simple fact is noone has proven yet that anyone can write compilers good enough to give VLIW overwhelmingly compelling performance outside of niche uses (DSPs, for example). I remember the Multiflow and Cydrome guys giving the same "it's the compiler, stupid" spiel in the mid-80s, and the story hasn't changed much except the details. We bought a Multiflow Trace...it was really nice, for certain problems, but not order-of-magnatude-faster, change-the-world nice, which was how it was sold.

Now, to be clear, a lot of these folks and their ideas moved the state-of-the-art in compilers massively ahead, and are a big reason compilers are so good now. Really, really smart people worked this problem.

My understanding is that Itanium was also stupidly strict on VLIW, making compilers even harder. Microsoft blogs had examples of some really funky bugs caused by it, too.

In comparison, Multiflow was not so bad.

The DEC Alpha actually did provide very good performance, and you could even run Windows NT on it. As far as I can tell, the biggest problem was just that Alpha systems were very expensive, and so they had a limited customer base. There were some quirks, but the main thing was just that you'd be paying $2000 for a PC and $10,000 for an Alpha based system, and most people didn't need the performance that badly.

I think Itanium would have had a better chance if Intel had made their compilers and optimized libraries free, but your larger point is really important: Intel’s performance numbers for the Itanium seem to have been broadly extrapolated from a few very FPU-intensive benchmarks and I don’t think it was ever realistic to expect anything like that level of performance for branchy business logic or to decisively change the price-performance ratio to become compelling. I worked with some scientists who did have a ton of floating point-heavy code, so they were definitely interested but their code also had a lot of non-linear memory access and the Itanium never managed a performance lead at all, much less one big enough that it wouldn’t have been cheaper and faster to buy 2-4 other servers for each Itanium box. In contrast, when AMD released the Opteron it decisively took the lead in absolute performance as well as price/performance and so we bought them by the rack.

> Alpha systems were very expensive, and so they had a limited customer base.

That's the usual chicken & egg problem... If they sold more units, the prices would have come down. But people weren't buying many, because the prices were high.

Itanium, like Alpha, or any other alternative architecture, would also have trouble and get stuck in that circle. x86-64, being a very inexpensive add-on to x86, managed to avoid that.