Itanium

» see intel.com as it lived, on the wayback machine

~ the obit ~

It started at HP in 1993 as PA-WideWord, a plan to get past what RISC could do, and HP brought Intel in that November. The joint project went public in June 1994. The idea was called EPIC, and the argument behind it was clean enough to be taught: processors were spending enormous transistor budgets working out at runtime which instructions could run in parallel, so stop doing that in hardware and put the parallelism in the instruction format where a compiler could settle it ahead of time. Intel's Albert Yu said that if he were the competition he would be really worried.

The first chip, Merced, was floorplanned in mid-1996 and came back far too large, so features were cut to make it buildable. It taped out on 4 July 1999. Intel settled on the name Itanium on 4 October 1999 and shipped the thing on 29 May 2001, with OEM systems from HP and Dell following in June. Performance against the RISC and x86 parts it was meant to bury was disappointing. Sales were in the thousands.

The compiler never arrived. That is the whole failure in one line. EPIC only pays off if a compiler can statically schedule around cache misses and branches it cannot see at compile time, and after a decade of very serious people working on it, the compilers were still not good enough to beat a big out-of-order machine that just worked it out as it went. Then AMD shipped Opteron in 2003 with AMD64, which added sixty-four bits to x86 without asking anybody to recompile anything, and by 2004 Intel was shipping the same instruction set in Xeon.

I have never had my hands on one, which puts me with almost everybody. What keeps me interested in it is the bet rather than the failure. Moving work from the runtime to the compiler is exactly the argument you get in any conversation about scheduling for accelerators today, and it is still not settled. Itanium is the version of that bet that got twenty years and a large fraction of two companies' credibility to prove itself, and did not.

~ the rap sheet ~

~ the funny part ~

The nickname landed on Usenet within hours of the name being announced, before a single chip had been sold, and it stuck for the entire twenty-year run. Itanic. There is no version of the story where a project survives being correctly summarized by a pun on its launch day.

The other joke is what the numbers looked like once both things were in the market. In 2007, roughly 55,000 Itanium servers were sold. In the same year x86 servers sold about 8.4 million. HP accounted for around ninety-five percent of what Itanium sales there were, which meant the architecture Intel had built to own the enterprise had precisely one customer that mattered.

~ what got buried with it ~

The software walked away first, and quietly. Red Hat dropped Itanium from Enterprise Linux 6 in 2009. Windows Server 2008 R2 was the last Windows release to support it at all. In 2021 Linus Torvalds marked the Itanium code in the kernel as orphaned, noting plainly that HPE was no longer taking orders for the hardware and Intel had stopped a year before that. Linux 6.7 removed the port entirely.

By the end it was being kept upright by contract. Documents that surfaced in the 2012 HP versus Oracle case showed Intel taking $440 million in 2008 and another $250 million in 2010 from HP to keep developing Itanium through 2017. The last generation, Kittson, launched in May 2017 and was a clock bump. Intel closed new orders on 30 January 2020 and made the final shipment on 29 July 2021, both dates exactly as announced.

~ the inheritance ~

Itanium's real legacy is x86-64, which exists in the shape it does because Intel spent the late 1990s insisting that the sixty-four bit future had to be a clean break. AMD disagreed, shipped the compatible thing, and won so completely that Intel had to license its rival's instruction set to stay in its own market. The architecture built to end x86 is the reason x86 is still here.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »