Intel Itanium IA-64 Support To Be Deprecated By GCC 10, Planned Removal In GCC 11

[jacob]

ONE of the worst? It was several of the worst ideas put together, with a few good ideas thrown in as garnish.

A few good ideas? I can't even think of one. Except maybe some stuff like SMT, which other CPUs have too and which is basically contrary to the very concept of the Itanium. They implemented it as a band-aid trying to make up for a fundamentally terrible design.

[L_A_G]

The whole story of IA64 is a kinda sad mess when you find out about what it was supposed to be and how things actually ended up going.

Intel originally intended it to replace 32 bit x86 and have AMD waste their time creating their own 64 bit x86, which they gave the rather unimaginative name "AMD64". The Itanimum hardware was supposed to offer much better performance while offering full backwards compatibility, but the hardware that actually came out wasn't much faster than the competition and ran 32 bit x86 code pretty slow due to having separate dedicated hardware for doing so. Some estimates said that the performance was about on par with a 100 MHz Pentium 1 when the original Pentium 4s were already on the market. To add insult to injury AMD soon released the original Opteron which not only traded blows with the Itanium on native code, but also ran 32 bit x86 applications much faster thanks to not needing separate dedicated hardware to do so.

It obviously didn't take long for Itanium to pick up the nickname "Itanic" as it failed to get support from end users and as a result hardware vendors like Dell and IBM along with software vendors like Microsoft and Oracle started dropping support. Silicon Graphics, then going by the abbreviation SGI, even went belly-up partially because of their failed bet on Itanium systems. Eventually it got to the point where HP, who co-developed the architecture together with Intel, started suing software vendors like Oracle and Microsoft to keep them from dropping support. Worse yet, with those court records it came out that HP was literally paying for software vendors to support Itanium and in the case of Oracle this amount was 690 million dollars annually.

Obviously Itanic never reached the lofty heights it was intended for either in terms of sales or performance. The only reason why it stayed around for this long was because of HP's continued investment and how credibly hard it's marketing department pushed their Itanic-based systems up until at least last year. Because of this I'm pretty sure that many companies who fell for the HP sales pitch won't be buying any HP products for a long time to come.

[jacob]

OMG.. WTF? I never even heard of it, and it is absolutely gloriously insane:

Instruction length: 6 to 321 bits.. yes, bit aligned
Object oriented instructions... Yes really
Hardware defined and enforced software data types....
Hardware garbage collection........

The Atari Jaguar was another beauty. The purportedly "64 bit" architecture had in fact a M68000 combined with two proprietary 32 bit cores which were kinda-sorta-but-not-really compatible with each other. It wasn't a SMP architecture but rather a weird brand of master/slave architecture in which any of the cores could be used as master. Now of course some parts were big endian and some little endian, which was especially "fun" to code for. As an extra garnish, in the great Atari tradition there was zero documentation, no compiler to speak of, no development environment and when the Atari technical support team was asked questions at developer conferences, their honest answer typically was that they had no idea.

[jacob]

The whole story of IA64 is a kinda sad mess when you find out about what it was supposed to be and how things actually ended up going.

Intel originally intended it to replace 32 bit x86 and have AMD waste their time creating their own 64 bit x86, which they gave the rather unimaginative name "AMD64". The Itanimum hardware was supposed to offer much better performance while offering full backwards compatibility, but the hardware that actually came out wasn't much faster than the competition and ran 32 bit x86 code pretty slow due to having separate dedicated hardware for doing so. Some estimates said that the performance was about on par with a 100 MHz Pentium 1 when the original Pentium 4s were already on the market. To add insult to injury AMD soon released the original Opteron which not only traded blows with the Itanium on native code, but also ran 32 bit x86 applications much faster thanks to not needing separate dedicated hardware to do so.

It obviously didn't take long for Itanium to pick up the nickname "Itanic" as it failed to get support from end users and as a result hardware vendors like Dell and IBM along with software vendors like Microsoft and Oracle started dropping support. Silicon Graphics, then going by the abbreviation SGI, even went belly-up partially because of their failed bet on Itanium systems. Eventually it got to the point where HP, who co-developed the architecture together with Intel, started suing software vendors like Oracle and Microsoft to keep them from dropping support. Worse yet, with those court records it came out that HP was literally paying for software vendors to support Itanium and in the case of Oracle this amount was 690 million dollars annually.

Obviously Itanic never reached the lofty heights it was intended for either in terms of sales or performance. The only reason why it stayed around for this long was because of HP's continued investment and how credibly hard it's marketing department pushed their Itanic-based systems up until at least last year. Because of this I'm pretty sure that many companies who fell for the HP sales pitch won't be buying any HP products for a long time to come.

I don't remember all the details but it was actually pretty simple. The WLIW/EPIC concept originated at HP. Someone must have smoked something particularly heavy and thought that it would ever be a good idea, but HP needed a parter to implement it. Then there was Intel with two problems. On one hand, they really, really wanted to get AMD out of the market, but that required changing the instruction set, because they couldn't stop AMD from making x86-compatible CPUs (they tried, and failed). That was one of the primary motivation for Intel's many failed attempts at launching a new architecture (iAPX432, i860, i960 etc.) On the other hand, the Apple-IBM-Motorola alliance released its PowerPC processors which, Intel feared, could offer a better price/performance ratio than its then-flagship, the Pentium (and for some time, they indeed did). The result of all this was the Intel-HP alliance where Intel would implement and manufacture HP's design.

The thing is, nobody outside of HP (not even sure about Intel) seriously thought this would ever work. IIRC the nickname Itanic appeared on the NEXT DAY after Intel announced the Itanium brand. Then the first core (Merced) was released, extremely expensive, with lacklustre performance on native code and a x86-compatible mode that ran at the speed of a 100MHz Pentium and virtually everyone made their mind at that stage.

Meanwhile AMD feared that if the Itanic caught up, it would indeed lose its market. They created AMD64 as a desperate attempt to stay relevant by offering the first affordable 64-bit CPU for the masses. So Intel had its Itanic and AMD had a design that offered excellent performance at a fraction of Intel's cost and, at the same time, had basically no downsides as in the absolutely worst case, it would just run 32-bit software like any ordinary x86 processor with no penalty. The rest, as they say, is history.

[bison]

A few good ideas? I can't even think of one.

In-order execution, which is not susceptible to side channel attacks.

[L_A_G]

...

I think the later failure of the architecture has colored your perception pretty badly. In case your memory fails you or you never heard of this, Silicon Graphics and Compaq/DEC abandoned MIPS and Alpha respectively specifically to focus on making Itanium-based systems instead. Both IBM and Dell made servers using the chips completely out of their own will as shown by how HP couldn't sue them for breaking any contractual obligations like they did when Oracle decided it was time to abandon ship. IBM and Dell's exit from the Itanium server market didn't even happen until 2005, 4 years after the first Itanium chips came out and 2 years after the second series came out.

AFAIK Itanium server sales topped out at about two billion a year and the real phase-out of properly invested parties didn't really start until around 2010 when Intel dropped IA64 support from their then latest version of ICC and Microsoft announced their phase-out for the architecture.

As for AMD64, now more commonly known as x86_64 after Microsoft told Intel in no uncertain terms that they weren't going to support their alternative 64 bit x86 architecture, I highly doubt it was an act of desperation. All CPU makers outside of ones dedicated to embedded systems knew by the mid 90s that a move to 64 bit was inevitable and needed to get underway as soon as possible so AMD was obviously just doing what everyone knew when it started work on what would be known as AMD64.

[carewolf]

I don't remember all the details but it was actually pretty simple. The WLIW/EPIC concept originated at HP. Someone must have smoked something particularly heavy and thought that it would ever be a good idea

VLIW seemed like a good idea in the late 1980s if you never imagined out-of-order superscalar processors, but by the time Itanimum was designed, we already had superscalar processors(*), and when comparing VLIW's statically scheduled "super-scaling" to dynamic super-scaling, static scheduling is only better in power consumption, but failed terribly in performace, and Itanium wasn't meant to be low power / low performance processors. Transmeta made a better gamble there, but also failed, again mostly due to performance.

(*) Not only did we have it, with it X86 had started killing off every other otherwise superior architecture that didn't have it.

[torsionbar28]

"Considering the GCC compiler is used to compile the Linux kernel and IA-64 doesn't enjoy coverage from other compilers like Clang able to build the Linux kernel, it will effectively mean the end of the road for new Linux support moving forward."

Pretty sure when RHEL and SLES stopped supporting IA-64 a while back, that marked the end of the road for Linux on IA-64. This is enterprise hardware, nobody is running Gentoo on these things lol.

[torsionbar28]

Meanwhile AMD feared that if the Itanic caught up, it would indeed lose its market. They created AMD64 as a desperate attempt to stay relevant by offering the first affordable 64-bit CPU for the masses. So Intel had its Itanic and AMD had a design that offered excellent performance at a fraction of Intel's cost and, at the same time, had basically no downsides as in the absolutely worst case, it would just run 32-bit software like any ordinary x86 processor with no penalty. The rest, as they say, is history.

You must not have been around back then. AMD64 was anything but a 'desperate attempt', it was AMD's brilliant response to intel's failed market strategy. AMD played the chess pieces like a grand master while intel (and HP) threw good money after bad, on a failed architecture. Itanium was so bad, it cost $thousands (for even the cheapest part), yet underperformed a desktop Pentium on launch day, and never was able to catch up. With each iteration, intel claimed the next one would be better. And year after year, it never was. HP went whole hog into Itanium as a PA-RISC replacement, so their sizable HP-UX market share depended on Itanium's success... which never came. Believe me, I know, I worked for HP specifically with this stuff, for over a decade during the transition from PA-RISC to IA-64.

If anything, Itanium was intel's desperate attempt to displace AMD, as AMD was producing faster and cheaper 32 bit x86 chips than intel at the time, and AMD was winning the Ghz race with their excellent Athlon chips. Remember AMD's "Super socket 7" chips? While intel topped out at 233 Mhz with the Pentium MMX, AMD was up to 550 Mhz with the K6-3. Of course the Athlon64 and Opteron with AMD64 instruction set only furthered AMD's lead. It wasn't until ~2006 when intel finally dumped their garbage Netburst P4 architecture and replaced with the Core Duo, and then the Core 2 Duo which cloned the AMD64 instructions. If you recall, Netburst P4's were so terrible, they regularly overheated, couldn't compete with Athlon on performance, and only scaled to 3.8 Ghz max. intel claimed that Netburst would scale to 10 Ghz. Yes, intel actually said that, TEN gigahertz! So not only was intel failing hard on IA-64, but they also failed hard in their IA-32 department too, from the late 1990's through ~2006.

[HyperDrive]

VLIW seemed like a good idea in the late 1980s if you never imagined out-of-order superscalar processors, but by the time Itanimum was designed, we already had superscalar processors(*), and when comparing VLIW's statically scheduled "super-scaling" to dynamic super-scaling, static scheduling is only better in power consumption, but failed terribly in performace, and Itanium wasn't meant to be low power / low performance processors. Transmeta made a better gamble there, but also failed, again mostly due to performance.

(*) Not only did we have it, with it X86 had started killing off every other otherwise superior architecture that didn't have it.

I'm not entirely convinced VLIW architectures are dead. I'm yet to find any show-stopping flaws in the Mill, for example (though I still have some doubts about the spiller section, I'd love to see more details about it).