Itanium: Intel's Great Successor

TL;DR

Intel's $5 billion Itanium CPU failed as x86's successor because of architectural overreach, compiler complexity, and AMD's simpler x86-64 extension winning the market.

Key Points

• 1.Intel pursued a clean-sheet 64-bit architecture to escape x86's limitations and AMD competition. A 32-bit CPU could only address ~4GB of RAM, shutting Intel out of the lucrative UNIX workstation/server market dominated by RISC chips like SPARC, PA-RISC, and DEC Alpha.
• 2.HP's internal VLIW-derived PA-Wide Word design became the technical foundation of the project. Bill Worley, Josh Fisher, and Bob Rau — veterans of IBM's RISC revolution and VLIW startups Multiflow and Cydrome — developed PA-WW, which Intel chose over its own 64-bit effort after an internal bake-off.
• 3.EPIC (Explicitly Parallel Instruction Computing) was the architecture philosophy underpinning IA-64. A descendant of VLIW, EPIC shifted parallelism work to the software compiler rather than hardware, aiming to outscale superscalar designs, but required enormously complex compilers that were notoriously difficult to build.
• 4.Cultural clashes between Intel and HP severely complicated the Merced chip's development. HP's consensus-based management collided with Intel's 'constructive confrontation' style while teams ramped up with inexperienced staff, causing HP to take sole lead on the second-generation McKinley chip.
• 5.Merced suffered two major delays, slipping from 1998 to a mid-2000 release at a cost of $5 billion total. Intel underestimated circuit complexity, faced clock-speed bottlenecks at its ambitious 800MHz target, and had to move from the 250nm to 180nm node, pushing servers to customers only in Q4 2000.
• 6.The first Itanium launched in 2001 was outperformed by newer RISC chips and ran 32-bit x86 software poorly. Sun's UltraSPARC III and IBM's Power4 quickly overshadowed it; IA-64's fundamental incompatibility with x86 meant emulation sacrificed the chip's own advantages, and projected 60% server market share never materialized.
• 7.AMD designed x86-64 (AMD64) as a direct counter, authored largely by Jim Keller, and announced it in October 1999. Excluded from Itanium's architecture by Intel, AMD bet that developers would refuse to recompile software for a radically new ISA, offering a 'simple' 64-bit extension fully compatible with existing 32-bit x86 code.
• 8.The rise of commodity x86 compute clusters undermined Itanium's entire market premise. Companies like Google abandoned expensive proprietary mainframes for cheap Pentium/Xeon-based Linux clusters networked together, meaning Itanium's biggest competitor ultimately became Intel's own Xeon product line.
• 9.Intel was forced to adopt AMD's x86-64 extension in 2004, and Itanium limped on until its last chip shipped in 2021. HP bought 85% of all Itanium production and paid Intel $690 million across two deals to keep the line alive; projected server sales of $14 billion by mid-2004 reached only ~$600 million in reality.