Comments on: Itanium upgrade delayed

By: Toon Knapen

Toon Knapen — Tue, 10 Feb 2009 12:25:43 +0000

I agree that there is no added value for ISV’s to support itanium (actually when working at an ISV some time ago I pushed the port to itanium out myself).

Now I only have an itanium at my disposal and I’m trying to understand why people prefer itaniums over others. I conclude it is not a technical issue.

By: John Leidel

John Leidel — Tue, 10 Feb 2009 00:49:53 +0000

Toon, I’ve done quite a bit of development on large IA-64 machines throughout the progression of the EPIC architecture. Many point to the wide instruction pipeline or the delta in cache/memory layout on IA-64. Personally, the majority of time I spend ‘head-scratching’ during performance optimization on IA-64 centers around precision [especially floating point]. Where IA-64 wins with a wide instruction pipeline and comparison registers, x86_64 wins with instructions specifically designed to save cycles in dealing with heavy precision [ie, denormals, etc: see `-daz and `-ftz`].
John [West] is most certainly correct in his statement about ISVs. How much time do you want to spend porting/optimizing for IA-64 with such a limited market?

By: John West

John West — Mon, 09 Feb 2009 22:33:28 +0000

I’m not sure what non-negligible is. If you slice out the server market in terms of chips that are “server grade” (which is one way to look at it) the share could be reasonable, but I don’t have a figure. But a lot of customers run “non-server” grade cpus in their enterprise servers, which implies that those chips are meeting their needs, so it’s also fair to look at the total population of servers and grab the Itanium percentage of those. In those terms, I would guess (I emphasize guess, don’t have numbers) that the share is quite small indeed.

In terms of what’s so hard about it, outside of HPC I think you have to think in terms of an ISV evaluating the cost of porting and then maintaining another build of their product vice potential revenue from customers who won’t buy their product unless there is an Itanium version of it. While many have said its worth it, many (many) more have not. Within HPC this isn’t really such an issue in my mind, since much of what we do is user’s own code.

By: Toon Knapen

Toon Knapen — Mon, 09 Feb 2009 20:34:14 +0000

What’s so hard at porting from x86-64 to IA64?

When working on the ia64 and looking for more info, I was and still am astonished that so few are interested in it and at the same time they seem to have a non-negligible market-share?

By: John Leidel

John Leidel — Mon, 09 Feb 2009 17:21:38 +0000

Once again, this is a story of “application specific.” Many institutions don’t find the need/funs/patience to port their codes to an EPIC platform. The use the additional money to simply buy more x86_64. As John mentioned, there are pockets of supporters for IA-64. If Intel is continuing support and production, they’re probably collecting enough margin on sales to warrant doing so.

…just my two cents post-caffeine

By: John West

John West — Mon, 09 Feb 2009 14:54:55 +0000

Sadly, I’m not being paid by IBM for anything. Just to be clear, I wouldn’t construe that as an anti-Intel comment. I’m not for or against any chip maker in the same way that I’m not for or against any car engine maker. As long as it works, I’m happy. The Itanium might work, but it’s far from successful.

Tukwilla was originally expected in 2006 (maybe 2007, the reports are conflicting), and still hasn’t appeared. The architecture originated at HP to better support HP-UX, and production has been plagued by delays that have made it difficult to build products around. That is, if the marketplace actually wanted products built out of them. Built at least in part for HPC, they haven’t had the price/performance that HPC customers typically want, and outside those few with specialized requirements the Itanium has been a big dud.

In terms of stats, the Wikipedia identifies the Itanium as the 4-th most deployed chip, behind x86-64, POWER, and Sparc. The chip distracts the company, and they should call exit on a failed experiment and move on to other things.

By: joe schmoe

joe schmoe — Mon, 09 Feb 2009 12:51:29 +0000

maybe he was paid by ibm to make the anti-intel comment

i’ve never used itanium, but i’m glad intel is trying to advance their technology

By: Toon Knapen

Toon Knapen — Mon, 09 Feb 2009 09:27:51 +0000

On what basis do you recommend to put a bullet in the Itanium.

I’m not a fan either but I had to work for the past year on itanium machines and I must say that I found the hw and software stack (.e.g profiler) on top of it quite nice to work with.

I guess the memory-bw of the itaniums is way inferior to those of the Xeon’s but once they are both equiped with CSI, would’nt the mem-bw not be identical (and thus would’nt the itanium have catched up) ?