64-bit Future Scenarios

WE'RE JUST A WEEK ahead of the "major milestone" - for the first time, the next step in X86 PC CPU platform was supposed to be defined not by Intel, but by "perpetual underdog" AMD - with the Opteron, the high-end portion of X86-64 Hammer CPU family. And this might be a serious one: the whole extension of instruction set is done by AMD, including 64-bit operations and enhanced version of SSE2 FP, too. So, you supposedly get both top 32-bit performance and top 64-bit performance.

Plus that nice and very scalable "poor man's EV7 Alpha" SMP architecture with each CPU having its own direct-attach memory buses and ultra fast HyperTransport links connecting all of them to avoid shared-bus contentions and scalability problems, and you got a clear 64-bit mainstream champion! Oh, provided AMD can execute the whole thing properly and produce enough CPUs that will work without melting... what could be a major "undertaking".

And yes, the 32-bit performance of Opteron is way better than 32-bit X86 hardware-assisted emulation of Itanium2, however the initial 1.6 GHz Opteron version does not seem to match the 3+ GHz Pentium 4 / Xeon, according to most initial benchmarks seen on the Web.

Let's assume for the moment, even though we know it is a VERY tall order, that AMD executes the Opteron launch and shipments flawlessly, and many thousands of Opteron 240, 242, 244 etc chips start heating up the server rooms around the world soon. Let's also assume that the same happens some six months from now with the desktop version, the Athlon64 - an even TALLER order knowing AMD's "track record". What's the effect on the rest of the 64-bit gang? What decisions may some of the other vendors make? And, how will those decision affect their other partner vendors? Finally, what's in for all of us?

Intel Itanium...
Of course, everyone from Intel might publicly tell you this - the epic Itanium effort is the be-all and end-all of Intel 64-bit strategy. After all, three of the best 64-bit architectures - MIPS, HP-PA, and the speed king of all, Alpha - were sacrificed to the EPIC deity. At least that is the case in the general workstation and server market (MIPS of course has a great second life in the embedded arena). Well, sometimes it's just the usual case of "heroes die, crooks survive." Is it maybe utter injustice that SPARC still survives, for instance?

However, even after the Itanium 2 arrival, the initial take up seems to be somewhat slower than expected, and Intel is a volume-driven business, isn't it? The already ancient E8870 chipset for Itanium2 is around to stay for Madison and Madison2 (9 MB cache version early next year) too, according to feedback at IDF this February. The official line is "enabling full compatibility with the new CPUs for the existing Itanium2 systems". Well, how many of those are out there to justify "compatibility with huge installed base"? Not too many for sure.

It does cost money to develop a new generation chipset for a complex platform like Itanium 2, and, if the volumes and $$$ are not happening, what then? Of course, stick to the old one as long as possible. Unfortunately, unless the common sense prevails and Madison2 9M gets a new 667 or preferably 800 MHz FSB and chipset, that also means that we might be stuck with the old 6.4 GB/s Itanium2 FSB until deep into next year. That bandwidth is fine for Canterwood Pentium4, but is at the bottom of the field among new 64-bit CPUs this and next year! 64-bit applications usually happen to have very large data sets, hard to efficiently cache in, so the strain on the memory bus will be much greater than on a PC desktop, especially when you have several hungry CPUs competing on the same bus.

Look at it this way, a 4-way Itanium2 (McKinley or Madison, probably Madison2 too) will have its 4 CPUs contending to share 6.4 GBytes/s of common FSB bandwidth, while a 4-way EV7 Alpha or 4-way IBM Power4+ HPC in p655 system will have their 4 CPU comfortably basking in humongous 51.2 GB/s total memory bandwidth, spread nicely a quarter for each CPU, and every part of memory still accessible to all CPUs - in parallel! Do you want to guess which platform could then be more suitable for high-performance computing, whether commercial or technical?

All this doesn't mean the underlying Itanium EPIC architecture is a cripple - Itanium2 is among the speed champions right now. The 1 GHz McKinley is right up there with both Power4 and Alpha EV7 on many apps, except when memory-intensive stuff kicked in. With 6 MB cache in a 1.5 GHz Madison, and 9 MB cache in a 1.8+ GHz Madison2, the CPU core itself is a top-performer - as long as the software vendor spends a humongous amount of time to optimise it. It takes "a bit more" effort to fine-tune software performance on Itanium compared to other platforms.

So, right now, not the core, but the FSB change is the priority, at least converting the 200 MHz double-pumped (400 MHz throughput) 128-bit FSB into a quad-pumped (800-MHz throughput) FSB. A minor change on the CPU and chipset side, but a major boost for streaming-sensitive apps - hopefully Intel will make that step for Madison2 at least.

Itanium2 the core CPU logic is huge and damn complicated compared to elegant Alpha or even Power (after all, that epic effort ended up in something like 100 different instruction FORMATS - not instructions, mind you), but at the end of the day, the benefits are there too. For instance, the humongous 128-entry integer and FP register files help a lot in matrix-multiply and other computationally intensive code. Itanium2 can achieve up to 90% LINPACK FP throughput efficiency (Rmax) while Power4 manages only 60% at present - but Power5 will improve that dramatically towards the end of this year. Of course, there are many real jobs where Power beats Itanium, and also jobs where Itanium has an advantage.

In summary, at present, FSB or the overall bus architecture is the main limitation of Itanium - but in the long run, massive core upgrade is due too: after all IBM CPU cores will change substantially in the meantime with Power5 and 5+. But what if 64-bit X86 cores in the meantime gain on performance? Will it then be Intel Itanium...

...or Intel Yamhill?
The rumours are surfacing that, if the Itanium volumes end up disappointing and X86-64 does make serious inroads (as it may now, with Microsoft support), Intel's Yamhill "Plan B" may get into action. And no, this doesn't mean Itanium gets killed off (yet) - in the worst case, it may be passed on to HP to continue the platform, since HP anyway bet their high-end futures on the EPIC.

There was an analysis on the Web of Prescott chip die, with conclusion that the ALUs and data paths for integer portion really look 64-bit wide now. Could it be that Intel is already building 64-bit capabilities into its X86 chips, and just waiting to see how AMD chips fare in the market? And, are they using X86-64 according to existing AMD platform spec, or their own 64-bit X86 extensions?

Now that Microsoft is officially behind AMD X86-64 spec, it would be hard for Intel to create another incompatible 64-bit X86 platform. If there is any new 64-bit X86, it would from now on have to be according to X86-64 definition - but again, as you will see, that may not be a problem for Intel at all.

So, if X86-64 comes out strong and Intel then brings out Yamhill, what happens? If Yamhill sees the light of the day, it will initially be "enabled" on Xeon CPUs for workstations and servers, and only afterwards on whatever desktop chips Intel has at that time. Well, Itanium also aims at workstations and servers. If you extrapolate the Prescott/Nocona expected performance and add extra zing if it or its successor has X86-64 extensions a la AMD, you could end up with a processor that equals or exceeds Madison2 in 64-bit performance (at least SPEC figures), but costs far less and runs all the legacy 32-bit stuff at top speed.

In fact, Intel would probably be able to create a faster and cooler X86-64 processor than AMD itself. Pentium64 or Xeon64 sounds fine, isn't it? Combined with Chipzilla's monstrous technical, financial and marketing might, they can still seriously threaten AMD survival in the long run. So, whichever way it goes - IA-64 or X86-64, or both - they could be the winner. However, IA-64 had another big-size development partner from Day One. And that's where our Part 2 starts... µ