|
AH, AH, aren't these Intel chaps so unpredictable - who would guess it would add the iAMD64 (IA32e) into those mainstream Xeon and soon Pentium CPUs so easily? Xeon made compatible with Opteron? Ouch.
The ultimate insult to add to the injury would be similarly "64-bit extending" the Dothan and its future Pentium M successors, enabling 64-bit mobile Intel computing way before Itanic could even dream of fitting into anything sexy and slim. Oh sorry, theire is 64-bit mobile X64-64 computing already, with a few Athlon64 notebooks which somehow aren't advertised too loudly (can't seem to figure why, huh).
Well, what happens than to the expensive, giant luxury-free on-board, cruise liner, Itanic? Is the deadly iceberg that seals the doomed ship's fate coming straight from its home harbour's waters around Satan Clara?
Captain Carly should have considered these questions, asking both herself and her then-deputy Curly, when doing the unthinkable: abruptly dismantling her two speedy flagship star liners, Alpha and PA, when they were still in tip-top competitive condition - probably better than the Itanic is in right now.
Performance, dollar for dollar
The current top-of-the range Itaniums out there right now (1.4GHz 1.5MB cache for DP systems, and 1.5GHz 6MB cache in quad++ MP systems) are not bad overall performance-wise: the expensive Madison 6M is among the leaders in both SPECint and SPECfp - it is basically its slow FSB chokepoint that diminishes its performance potential. Of course, SPEC is not the only performance metric. The user's application performance in real life seems to depend greatly on how efficiently the code is handled by both the EPIC compiler and the in-order CPU execution. Some perform greatly, some others horribly. Look at newsgroups and forums all over for more on that topic.
Right now, against the current 3.2GHz 2MB cache XeonDP "Extreme Edition", or 2.2 GHz Opteron, Itanium's performance mileage varies - sometimes it is faster, sometimes it is slower. The problem is, even in the DP variety, Itanium is quite a bit more expensive than the other two - at least 40% more for a similar system.
This year, besides the interim FSB-improved Madisons (1.6 GHz/533 MHz - what about 1.67GHz at 667MHz FSB?), you'll probably see the 9MB cache Madison as well, at similar clocks, still on the 130nm process.
On the Xeon front, well Nocona and Jayhawk will pretty much be in lockstep with Prescott and Tejas: I expect to see the 4GHz parts at least announced, if not shipping, this year, all of course on the 90 nm process.
How would these stack up against, say, a 1.6GHz Madison? My guess is that, with optimised Intel or PathScale compilers, and low-latency dual-channel DDR2-400 RAM, the 4GHz Nocona or Jayhawk will beat even the Madison 9M flavour, in both SPECint and SPECfp. It might only be a slight win, but at quarter the system price (one-sixth CPU price), and still fully 64-bit with all the associated memory addressing goodies! Oh, I forgot the full-speed, 100%, 32-bit legacy software compatibility, something the Itanic kinda lacks.
What about Opteron? Oh my, that may turn to be an even nastier situation for the Itanic. Why? Well, not only it will be able to reach 3 GHz by the end of the year (probably beating the 4 GHz Nocona in most apps then), unless AMD screws up the 90 nm transition - but also it scales, very smoothly, up to 8 ways, adding the memory and I/O bandwidth instead of 4 CPUs quarrelling for a single narrow FSB pipe of XeonMP or Madison systems. Dollar for dollar, it beats both Itanium and Xeon there.
And, with a common software base for both Intel and AMD 64-bit X86 and yes, I do believe those few minor differences will soon be worked out - AMD is already putting SSE3 in their next core revision platforms, guess where software firms will put their porting effort to? No prize for that one.
The Alpha Experience: rekindling the past
Now, some may say, Itanic still shines in certain specialist apps, and people will be willing to pay a premium for those apps and use Itanic there, in niche markets, therefore justifying keeping its existence and development.
But hold on, some of us have some experiences, fairly recent ones, for such cases. There was a CPU called Alpha, which was as purely 64-bit (if not more) than Itanic is, many years ago - far more elegant, too. And, whether in Windows, UNIX or Linux, it was an undisputed performance champion for quite a few years, whether you look at integer or FP.
In its heyday, in fact all the way till its untimely murder, Alpha was quite a bit better supported in scientific and technical apps than the Itanium is right now. And there was a range of systems, including some OEM units going for as low as US$ 2,000. Yet, it had trouble increasing its market share.
The users will say: "It (Alpha) must be at least twice as fast as the top Intel, in whatever application I run natively, but not more than 20% more expensive". Even when that demand was met (remember the EV56-500 vs P6-200 days?), then they'd say "well, not enough software was native", and FX!32, while providing very fast X86 translation on Alpha, wasn't the answer to all headaches.
Now, does Itanium meet these requirements? Is Itanium at least twice as fast across the board than Opteron or Xeon, while only 20% more expensive for a similar configuration? Well, neither seems to be the case - by far, either now or in near future.
Of course, Intel's backing was supposed to be the key drive behind Itanium's expected success. Why would otherwise Compaq, HP, SGI and umpteen others throw in the towel and kill perfectly fine platforms, just for the promise of Itanium? But now, as Intel refocuses on the 64-bit X86 in its front-line battle against ever-stronger AMD, and maybe kindly asks HP to adopt the IA64 baby, where does that leave HP then? And SGI? And few others, too...
What can HP do, especially if their Opteron sales soon outdo their own Itanium sales? As the Xeons also become 64-bit this June, the major Itanium differentiator goes away... If Itanic is to be judged just on price, performance and software support, it would lose most battles - in fact, more than than Alpha lost in its days. What about often-mentioned RAS feature set? Well, that one, also very well done in IBM and Fujitsu systems, for instance, can't alone justify for the platform's continued existence. After all, Alpha EV7 had excellent RAS too, with CPU mirror lockstepping and real-time memory RAID5 on top of its leading performance - where is it now?
It is ironic that the same "legacy compatibility", which Intel benefitted from so much when Alpha, MIPS and HP-PA were canned, now comes back to haunt the Itanic ship passengers, in particularly the main tenant, HP. And it was AMD, led by ex-Alpha designers, who forced Intel to adopt X86-64 fully and therefore push this fatal iceberg towards its Itanic ship.
Others' headaches
Well, in any case, have Sun folks seem to have found their temporary panacea in the Opteron? I really, really doubt that SPARC will dig themselves out of the deep (non)performance pit they fell into years ago, no matter how many parallel threads (pun intended) they spin to climb out of it. Opteron Solaris could be an interesting "commercial UNIX" solution for the existing Sun base, giving them an easy Linux migration path on the same CPU platform, if desired.
IBM seems to be headache-free these days - its POWER5, due this summer, is expected to have the performance lead for 8-way and larger systems, while the PPC 970FX G5 will extend that to both low-end AIX, Linux and MacOS systems - with its low power, it may be the first platform to allow for a dual-CPU 64-bit notebook machine! Its Opteron sales are doing well too, however Sun and HP Opterons (who'd ever think HP Opteron would be a reality so soon?) are encroaching there now.
For HP, well, it is still not too late for it not to have to re-learn the Alpha lesson: while it is unlikely it would return to Alpha, even though if it did, the EV8 on 90 nm would still beat the crap out of everything else in the market, it can at least either cut the losses early and focus on "services and added value" - justification was when the Alpha was killed, or simply go all-out after Mikey Dell. µ
Source: INQ!
|
Threaded Mode