💾 Archived View for arcanesciences.com › itanium › myths.gmi captured on 2024-08-18 at 18:27:11. Gemini links have been rewritten to link to archived content
⬅️ Previous capture (2022-04-28)
-=-=-=-=-=-=-
REALITY: Itanium generally compared well on industry standard benchmarks to contemporary x86 systems, and often exceeded contemporary SPARC performance. For technical computing loads, Itanium smelled like roses for a long time.
REALITY: Itanium held its own on technical computing loads, but was often very, very brittle, especially for branch-heavy code streams with a lot of dynamic memory access. After Montecito, no Itanium part was performance-competitive with contemporary Power (Power7 for Tukwila, Power7+ for Poulson) or even really contemporary Xeon (though the Itanium parts had more exposed QPI.)
REALITY: While the physical design of Itanium parts was often impressive, the design philosophy has aged poorly. It was built on two basic assumptions:
Considering we now live in a year where Apple is shipping a deeply OoO core with a wider frontend than any Itanium, ever, the first part can be effectively considered dead. In 2021 general-purpose high-end in-order is nonexistent, and there's a good reason we don't build machines like that anymore.
REALITY: Itanium sales peaked at close to US$4bn a year, #2 in RISC/UNIX and a significant chunk of the server market as a whole, and the product family lasted over ten years - which, honestly, is more than you can say about some Intel product families. Performance on some workloads was excellent during periods of Itanium's life.
REALITY: Itanium never achieved widespread popularity beyond HP, though some vendors achived niche success. By the time Poulson shipped, HP and Inspur were the only two vendors standing, beyond those offering rebadges of either HP Integrity or Intel's own SDVs. Itanium is an end-of-life product line twenty years after it first shipped.
REALITY: In-order machines are brittle, and Itanium more than most. This is true even when hand-writing assembly. Itanium code density was poor, and combined with an undersized (but very fast) L1I cache, this created issues, especially in pre-Montecito cores. Itanium's ability to initiate memory accesses early was always subpar compared to OoO machines, even with Advanced Loads and Speculative Loads.
REALITY: While Intel was the public face of Itanium, it originated from HP in the late 1980s as a long-term incompatible evolution of the PA line. Intel joined the project in 1994, when it was already fairly late in definition, resulting in the cancellation of Intel's own 64-bit RISC program, IAX. The HP design group in Fort Collins, responsible for Itanium design work, was absorbed into Intel's Fort Collins Design Center (FCDC.)
REALITY: Power and SPARC survived - Power via extremely high performance, SPARC via a large installed base. MIPS performance was not in a favorable position at the time of SGI's transition to Itanium. PA-RISC wasn't "killed by" Itanium in any meaningful sense, since IPF was always intended to be the long-term successor to PA. MIPS was non-performance-competitive and SGI was diversifying regardless.
GRAIN OF TRUTH: Alpha was legitimately very good and had a compelling roadmap, and was a victim of a small customer base and of interesting decisions from Compaq management.
REALITY: While it didn't exactly help with making Itanium a general-purpose merchant processor, most IPF customers literally could not care less. HP-UX always made up the bulk of IPF sales and it never ran on x86 to begin with. Nor did VMS, Nonstop, GCOS 8...