Itanium: Mark Architecture as Orphaned | Hacker News Reader

Itanium: Mark Architecture as Orphaned

git.kernel.org

88 pointscglong75 comments

Ahh the memories of dealing with Itanium for our large SGI boxes. It was super awesome to have a single system multi terabyte memory “machine” in 2005.

It was not awesome to keep adjusting out code to be more amenable to the VLIW-like needs of the processor (Intel’s compiler actually generated very good messages as to “why we didn’t do a good job with your code”, you just often couldn’t do anything with that info).

I still remain amused to guess what kinds of programs were included in the architecture simulations that led them to want so many integer units (12?) and so few floating point ones.

> I still remain amused to guess what kinds of programs were included in the architecture simulations that led them to want so many integer units (12?) and so few floating point ones.

If I had to guess, SPEC Int.

I one had a frank conversation with a CPU designer at a chip company. I was looking at a CPU that was fairly wide superscalar, but had super tiny caches. I observed that this chip was probably only efficient at running Dhrystone. His response was "We win and lose 7 figure CPU sales based off of Dhrystone benchmarks. DMIPS/Watt is the only benchmark that matters for the majority of our sales in dollars"

Sure. Specint and the more modern specintrate are obvious choices for “we can’t suck at this”. But the overall architecture was so extremely over “int heavy” that the box was just not great for the supercomputing segment it claimed to want to play in.

Maybe large Oracle databases, too?

The architecture really was the only game in town for a while for having so many physical memory address bits. I just assume someone “resold” it to the HPC/supercomputing groups as “here, go sell this”.

Not as bad as Niagara, that thing had what, 1 FPU shared by eight cores? Great for Java, a bit less for PHP that only had floating point arithmetic...

Poor Niagara :). The Niagara design though was clearly intended for web serving/databases and Sun didn't go try selling it into the LINPACK/supercomputing segment. That's the disconnect for me.

To be clear, if you had a multi-hundred processor system sitting around (as we kind of did, thanks SGI!), you used it. But... you probably shouldn't have purchased it.

And I thought Itanium was designed for Linpack ;-)

ours mostly just ran x86 code..

On the funny Pentium3 equivalent thing bolted on? Or you mean in emulation mode?

TPC OLTP.

I believe that early AMD64 was limited to 40 bit addressing, and until about 2007-2008 anything Intel-based other than Xeon MP was limited to 36 bit.

(I’m not sure why people downvoted / flagged you, so I vouched this).

I definitely don’t remember the numbers and haven’t looked them up on wikichip, but Itaniun certainly counts as “Intel-based” :).

I remember my time at the OSDL when we got our first itanium servers in. I recall that even though we were "partnered" with Intel, it already felt like an orphaned architecture. once the mainstream press started calling it "itanic", it seemed even more doomed.

I'm mostly sad to remember that the itanium was supposed to be a "replacement" for the DEC alpha, and Compaq selling the IP to Intel. very sad day for computing.

It's really a shame IA64 accelerated the deaths of Alpha and PA-RISC. Not sure they could have survived x86-64 or ARM64 but they still had legs when Itanium hit the scene.

I'm not entirely convinced that x86-64 would have come on the scene in its current form had itanium not been invested into so heavily by intel (and been such a failure).

remember, intel seemed hesitant to extend the x86 architecture, and x86-64 came out of AMD.

Linux used to refer to the arch as amd64 and then Intel copied it a few years later.

AMD licensed the EV6 bus from DEC and there was some discussion of having motherboards that could accept either an AMD K7 or DEC Alpha and just flash the BIOS. That might have made the Alpha last longer by having more low cost motherboards

It's still referred to as amd64 in Debian.

I don't think anything could have survived the rise of X86. Most x86 chips were 32 bit at the time and you need Alpha/Sparc/PA-Risc, or some fancy custom chips to work in 64 bit.

I wonder if the AMD/Intel competetion pushed x86 to get as fast as it got surpassing all those custom chips?

HP was involved in Itanium and was discontinuing there own chip lines. We were talking about moving to it, for the "superdome" line of machines. I was working with PA-RISC at the time, and it seemed like this was the next chip. But it wasn't compatible with PA-RISC but had an x86 compatibility mode..

It was a weird chip, and required really good compilers. I thought it was long dead. It was interesting though..

https://news.ycombinator.com/item?id=14518758

My impression was that HP essentially convinced Intel to build next-gen PA-RISC for them; for HP it was mostly a success in that it kept HP Unix and VMS running.

MIPS as well. SGI announced the Itanium switch and dialed back MIPS investment in 1998, but Itanium wasn't really usable until Itanium 2 in ~2003.

ARMs rise seems to be almost miraculous considering that I struggle to find any contemporary discussion of it as a serious ISA from back in the day.

Back in the day to be considered 'serious' Arm would have needed to compete on performance and that meant competing with Intel which they (primarily Robin Saxby) worked out was a losing strategy - so they didn't.

Then getting AArch64 onto the same CPUs as AArch32 meant in one bound they could cast off the legacy and start to compete on performance.

I was more thinking along the lines of academia than market share - if you read a book like the dragon book or Muchnick's text, they clearly thought we'd all be on Alpha or some other RISC machine rather than ARM - even for teaching.

Understood. I suspect that Arm was a bit too off the mainstream for teaching (why talk about Thumb / Jazelle etc?).

Plus probably better to teach for an architecture that had the performance cachet and where you could get some desktop machines.

Whenever I hear Itanium comes up the old pang of anger that it killed Alpha. Sorry can't help it.

> While intel is still officially shipping chips until July 29, 2021, it's unlikely that any such orders actually exist.

We got two delivered a month or two ago. These ones are never going to run Linux, so it makes no difference if they drop support.

What do you use them for, if you don't mind me asking?

Most likely VMS

I'd say HP-UX is at least as likely.

They’ll be running OpenVMS. Funny - HPE sold that last batch of hardware, but they refused to sell software to go along with it - you have to go over to VSI to get that.

It’s just an Ingres database and batch processing data into and out of it. Can you believe Ingres is still around?!?

The replacement is a DataBricks/Spark system, but it’s years behind schedule - that’s why they had to buy new servers before production ended. It makes me sad that to match the performance of a 2nd-gen dual core Itanium (the entry-level processor back then - 10 years ago), we need a 10-node cluster and all the guys are hand tuning their partition parameters just to get to that performance level.

Do you actually though? Or is the replacement just badly engineered? Lacking context, but that sounds unlikely to me, unless you were doing something that was extremely highly suited to itanium.

Oh don’t get me wrong - the “architect” on that project is in over his head.

The thing is, starting a DataBricks cluster can take 2-3 minutes. If you can’t keep it running, you have to bear that cost every time you want to use it. You know, like when you are batch processing... But when you’ve got an on-premise server with a nice fat fiber channel connection to an nvme Pure Storage SAN and a processor that is really good at I/O, it’s going to be pretty fast.

I still have nostalgia for OpenVMS. One of the first multi-user systems I worked on was a VAX! I bought an AlphaStation off of ebay a few years back and boot up VMS now and then.

You know an architecture is dead when even Linux gives up on it. Mixed feelings on this architecture. Oh well...

> It's dead, Jim.

I wonder how many folks would miss this reference. If I'm not mistaken, it's only in the original series. On the other hand, I guess Itanium is even more obscure than Star Trek at this point.

I tend to assume that most people know at least as much as I do, so I was taken aback when, after explaining line endings to a co-worker, she asked, "What's DOS?"

I think "he's dead, Jim" is/was common enough that many people would recognize it as a reference. Some would probably even know it as a Star Trek thing, but I suspect there are a bunch of people who know it references something but don't care enough to find out what.

The line stands well enough on its own without context, so it can be used even without knowing what its referring to... but it clearly is a reference.

> I suspect there are a bunch of people who know it references something but don't care enough to find out what.

Yeah I’m in this boat. They say this on Reddit all the time so I know it’s a reference to something, but like you said I never bothered to check what it’s actually a reference to.

Star Trek TOS is all spiffed up on Netflix, give it a try. Some of the episodes are still great.

YouTube's accursed algorithm bubbled a video into my recommendations with the provocative title: "What happened to the A and B drives in Windows?"

It suddenly hit me that there are adults alive today who have only ever known PCs without floppy drives.

Right, they might as well open those up for flash drives. Not like floppies are making a comeback.

You can assign any drive to them. Windows just won’t do it on its own.

A friend of mine is named Jim. I was working with him one day and received a "He's dead, Jim" error from Google Chrome. He became very confused.

"Why is it calling you Jim? How are you signed-in to Chrome?"

"I'm not. I don't use that stuff."

"Why is it calling you Jim then?"

"It calls everybody Jim. That's a quote from Star Trek."

He assumed the error message was personalized based on how you were signed-in to Chrome.

Turns out that, though he's a big Star Wars and Dr. Who fan, he managed to completely avoid Star Trek TOS. I found the phrase to be used ubiquitously too but, apparently, not so much that he ever noticed it.

>after explaining line endings to a co-worker, she asked, "What's DOS?"

How old was she? I bet there are plenty of people born in the late 80s / early 90s who've never used DOS. Windows 95 changed everything.

The averagely computer literate people that I know in my age range (mid-late 20s) refer to the windows command line as "DOS". I can't be bothered correcting them.

whyever? technically correct is the /best/ sort of correct.

What is DOS? Is it COMMAND.COM (of which the Windows command prompt is indeed a descendant) or the API underneath (INT 21H mostly)?

cmd is rewritten. Most of DOS was in assembly language. Most of NT is not. At one point cmd maybe contained some straightforward translation to C of some pieces of assembly, but I'm not sure (I'm sure it was a mess though; it may have been cleaned up a little now)

But cmd is surely quite compatible with command.com. With more capabilities, but still very compatible. And a deceivingly close user interface.

Yeah I didn't mean a descendant source code-wise. But I would say that DOS is more the 16-bit API than COMMAND.COM.

It was a primitive program loader.

msdos.sys, io.sys, and command.com

Source is interesting, I took a look:

https://github.com/microsoft/MS-DOS/tree/master/v2.0/source

Yeah, but I think you're supposed to follow it with the obligatory "You get his phaser, I'll get his wallet"

I wonder how would people feel about itanium failure had it been developed by anyone other than Intel; I mean, would there be that lingering feeling it could have been something had only the right amount of resources put into it etc etc?

Itanium was actually developed by HP starting in 1989 and then Intel joined in 1994.

Not only that but in the early 2000s HP effectively gave Intel all their hardware design team to help with the effort.

Maybe? The architecture is weird -- it required user programs to be aware of and utilize superscalar pipelining. Superscalar pipelining was a great bet, the assumption that programmers and application writers could explicitly take advantage of it was not. It's a really COOL architecture, but I'm not sure it could ever have been a commercially successful one.

For comparison, how do you feel about Transmeta?

Probably not, it felt a lot like an experiment that was picked to win and then still didn't. Almost like a reverse i960, which ended up being really good at a few niche applications that certain problem spaces kept it in at least some demand long after it was effectively abandoned.

no. in order VLIW dependent on handwaving compiler magic never works.

I still think there is a medium between the two that could work - if you look at uops per port measurements there's still a lot of room left for improvements i.e. the mantra is "The compiler will never know more than the chip" but sometimes the CPU is pretty dumb too

Same thing that happened to Transmeta ???. Too much dependence on software magic.

Transmeta was reborn as Nvidia Denver... which is not that successful either.

Can we un-orphan this by recreating this architecture using open source hardware? (A tall technological order, I know :) )

Would there be patent problems, or potentially copyright problems?

Would there be a single person in the world who actually likes Itanium enough? Is there really no point at all?

The kind of people who want to work on vintage hardware usually want to work on vintage hardware that is either cheap or beloved or both. I think Itanium is neither.

I think it also used 200 watts at idle. Put a bullet in its head.

How about we not and devote resources to crowdsourcing open source Alpha-architecture chips. You know, the 64-bit powerhouse architecture Itanium killed with its weirdness and lameness.

I know Jay Maynard (of Tron cosplay fame) runs an Itanic server in his house. Mainly because the ISA is so rare in the wild that it's unlikely to be affected by shotgun malware attacks.

Yes. Die, trash, die!

The king is dead, long live the king!

https://en.wikipedia.org/wiki/TeraScale_(microarchitecture)

This architecture accelerated Intel's downward spiral. It should have been shut down years ago, and I find it surprising in the changelog that Intel "is still officially shipping chips until July 29, 2021" no wonder Intel is behind on innovation.

They should have closed out the team working on this long long ago and changed everyone working on this to making their newer chips work better.

> This architecture accelerated Intel's downward spiral. It should have been shut down years ago, and I find it surprising in the changelog that Intel "is still officially shipping chips until July 29, 2021" no wonder Intel is behind on innovation.

I don't think so:

https://en.wikipedia.org/wiki/Pentium_4:

> Launched November 20, 2000

https://en.wikipedia.org/wiki/Itanium:

> Launched June 2001

https://en.wikipedia.org/wiki/Intel_Core:

> Launched January 2006

You seem to be basically claiming their downward spiral started before their biggest recent upswing, which doesn't make sense. Itanium (and the Pentium 4) were mistakes that they managed to recover from 15 years ago. However, it's only just being discontinued because Intel's not a company that's willing and eager to leave their existing customers in a lurch.

Edit: it sounds like HP was basically paying Intel to keep Itanium alive:

https://en.wikipedia.org/wiki/Itanium#HP_vs._Oracle:

> ...court documents unsealed by a Santa Clara County Court judge revealed that in 2008, Hewlett-Packard had paid Intel around $440 million to keep producing and updating Itanium microprocessors from 2009 to 2014. In 2010, the two companies signed another $250 million deal, which obliged Intel to continue making Itanium CPUs for HP's machines until 2017.

According the the article, the latest Itatium processor was released in, surprise!, 2017.

When they say "still shipping" I imagine a box in a warehouse somewhere full of new old stock for enterprise customers with really long maintenance contracts. I'm sure nobody is actually working on the architecture anymore - I wonder when the last chip came off an assembly line.

more itaniums shipped as keyrings than socketed processors.

If you were to ask most Intel employees about this announcement, the reaction you'd get would be "wait, we still make Itanium?" Whatever Itanium team is working on it still is clearly pretty small.

there's the small matter of these things called contractual obligations.

Intel is behind at the moment, but it's hardly in a downward spiral.

True before the M1. False now.

Don’t get me wrong, the M1 is clearly good, but remember AMD has had many years of being behind in both power and efficiency before getting to where it is now.

I don’t see intel just rolling over, I suspect it will just take some years for them to get back to a good arch, just as it did when they had their long P4 period leading up to the Core series.

AMD wasn't going against Apple in the performance silicon space when they were behind.

In the coming years, any fab process innovation will see Apple buying up most or all of the fab capacity for the new process in the supply chain, locking other vendors out, meaning that in order to keep pace with Apple Silicon, Intel and AMD will have to build bigger, hotter, more power-hungry chips -- and then they're not really competing with Apple, just retreating into the enthusiast and high-end server hardware markets.

Reply on news.ycombinator.com