Intel Details Eight-Core Poulson Itanium Processor

MojoKid writes "Intel has unveiled details of their new Itanium 9500 family, codenamed Poulson, and the new CPU appears to be the most significant refresh Intel has ever done to the Itanium architecture. Moving from 65nm to 32nm technology substantially reduces power consumption and increases clock speeds, but Intel has also overhauled virtually every aspect of the CPU. Poulson can issue 11 instructions per cycle compared to the previous generation Itanium's six. It adds execution units and re-balances those units to favor server workloads over HPC and workstation capabilities. Its multi-threading capabilities have been overhauled and it uses faster QPI links between CPU cores. The L3 cache design has also changed. Previous Itanium 9300 processors had a dedicated L3 cache for each core. Poulson, in contrast, has a unified L3 that's attached to all its cores by a common ring bus. All told, the new architecture is claimed to offer more than twice the performance of the previous generation Itanium."

Why?

[PCK]

I was under the impression that Itanium was all but dead. I'm guessing Intel must be contract bound to bring out new versions.

Thank you HP?

[jandrese]

I guess all of that money that HP has been dumping into Itanium development is finally paying off. Everybody else assumed Intel was just going to discontinue the product for obvious reasons, but here they are releasing a major upgrade to the core architecture. It still makes me wonder what HP sees in Itanium that makes them so gung ho about it though. Is it the vendor lock in? Is this upgrade enough to finally push Itanium past x86 based processors in some performance metric?

Re:What's twice a small number?

[Anonymous Coward]

Perhaps Intel fell into the trap of not communicating with the compiler team, but HP certainly did not.

The development of the EPIC concept at HP already started in 1992 as the to-be-successor for the HP-PA architecture. Look up e.g. the many joint research papers of CPU-architecture and compiler engineers for PlayDoh (or see e.g. http://www.hpl.hp.com/techreports/93/HPL-93-80.html for an intro to PlayDoh).

The compiler technology to do well for EPIC architectures was mostly available by the time IA64 launched. Arguably it hadn't advanced enough to be considered "ready for prime time", but things like data flow analysis of predicated code (http://www.hpl.hp.com/techreports/96/HPL-96-119.html), if-conversion (various influential papers), VLIW scheduling (e.g. selective scheduling a la Moon&Ebciolu), interprocedural analysis and various other interprocedural analyses were available and actually implemented in the HP compiler and probably also in the Intel compiler.

(John C. Dvorak wrote "How the Itanium Killed the Computer Industry". Perhaps one day I'll write "How the Itanium Revived Compiler Scalar Optimization Research" :-)

I think a bigger problem for Itanium was the underwhelming Merced. We used them mostly to heat the SuSE Maxtor building, they weren't good for much else. Slow, power-hungry, inefficient. And by that time, the market for things like database servers and high-end engineering workstations (where HP-PA was big) was imploding due to the advances on commodity architectures like AMD64 running Linux or even just Wintel32. Merced was a disaster, the chip wasn't ready and the compiler technology was not available widely enough, and by the time Madison came along the reputation damage was too big to be undone.

It probably also didn't help that support for Itanium in Linux was never very good. The "typical hacker" didn't have access to IA64 and the major companies supporting IA64 didn't invest in Linux-for-IA64. Compare how IA64 funding for Cygnus/Redhat got cut before binutils was complete (to this day, binutils for ia64 is still far from complete) to how AMD funded and cooperated with SuSE to get a good x86-64 Linux ecosystem even before First Iron. Also, GCC is only since recently beginning to catch up with the proprietary compilers of HP and Intel (and also e.g. SGI's Open64), but neither HP nor Intel ever really understood how GCC is a corner stone for the whole GNU+Linux system. For example, bash compiled by ecc was much faster than bash compiler by gcc -- but no Linux distribution ever shipped an ecc based complete distribution. An official LLVM port for IA64 doesn't even exist, but a port for the long-dead Alpha *does*. What does that tell you?!

Mis-management and lack of vision are as much a cause for Itanium's failure as the technology itself....