Beam me up, Scotty. Star: Solaris, Planet: Opteron

IN THE FIRST PART OF THIS SERIES we reviewed last year's massive changes at the performance end of the 64-bit battle, or what happens when a vicious bald beancounter runs a global technology business by killing off its crown jewels.

We saw that IBM, with the current technology edge but often ridiculous POWER4 "list" pricing, and Intel, with still more promises than deliveries on the Itanium front, but a truly aggressive attitude and commitment, still have a lot of battles to fight for the 64-bit performance and market supremacy. Even though one of them says how the two things can "co-exist". One Alpha guy told me that too a few years ago, and we know how that "coexistence" attempt ended...

Now let's look at the two other major players in the 64-bit arena - Sun's UltraSPARC and AMD Opteron/Athlon -- let's just call it Hammer for the easy identification purpose.

We'll look at their current situation, future roadmap potential, and whether they may have a chance to fight it out with the POWER and Itanium for the 64-bit championship trophy.

SPARCs fly: Beam me up, Scotty, to see the Solar(is) Eclipse!
Sun was among the RISC CPU pioneers when, in late eighties, the first SPARCstation 1 came. However, Texas Instruments' process problems with the first BiCMOS SuperSPARC caused delays and clock frequency target misses which until now blocked Sun's chances to reach the CPU performance throne. Looks like, as long as Scotty is cosy with TI, SPARCs will not fly - in Sun's another foundry relationship, and in SPARC performance...

Yet, with clever marketing, very aggressive sales, clear focus on one and just one platform (Solaris) and creation of fancy campaigns -- after all, who was behind the dot.com -- Sun managed to keep the market share leadership in the UNIX space.

Oh my, another case of worst performer as the market leader. As if we hadn't had enough of currently the worst, technically most backward CPU platform being the PC hardware market leader, and technically worst-ever OS platform having the stranglehold on that same PC for ages.

Right now, the fastest shipping UltraSPARC is the 1050 MHz UltraSPARC III, if you're lucky enough to get it -- it's 900 MHz still for mere mortals. It runs off a 150 MHz 128-bit front side bus with 2.4 GB/s bandwidth per CPU - just above the Athlon XP and below the P4-based Celeron.

USIII has in fact some similarities (large register file with windowing, no out-of-order execution etc) with EPIC, which might explain their shared (non)performance positions during the Itanium 1 (Merced) age.

Itanium 2, though, is way ahead of USIII in SPEC and seemingly other benchmarks, and consequently I saw the first deals being taken from Sun in favour of Itanic 2 based on - performance! And maybe the fact that Itanic 2 uses normal vanilla DDR-RAM instead of proprietary (Ed: you mean rip-off?) 232-pin Sun SDRAM DIMMs and still gets 2.6 times memory bandwidth, and of course the Intel marketing machine, one among rare ones that can fight Sun head-on. IBM doesn't stand still with POWER either, and, besides that platform's superb performance, the Big Blue's corporate stability and commitment to the platform are the most obvious carrots dangling in front of current Sun accounts.

Sun's roadmaps are colourful and exciting. The problem is, StarOffice presentation slides are not the same thing as shipping systems. According to some slides from independent sources, I should have played with dual-core UltraSPARC IV at the highest-ever clock for any 64-bit CPU up to date (i.e. above 1.3 GHz) , by now. But I can't play with it right now coz it's not there yet...

Let's put the thing in perspective: Sun still has a very strong Solaris/SPARC market share, and has a range of systems that do manage to squeeze the last drip of lacklustre performance from its CPUs. SPARC makes greedy guys like Oracle very happy, because now, due to the smart guy mentioned at the very top of this story, instead of a 1.26 GHz four-CPU Alpha ES45 you may buy say an eight-CPU 900 MHz Sun V880 for the same performance, and Oracle will earn twice as much due to their per-CPU licensing scheme. It doesn't matter whether the CPU is 300 MHz SPARC of 1.5 GHz POWER4+.

So, who said Larry is better-natured than Bill? Well, Larry at least got a good excuse for more money: fast cars, a string of girlfriends, more fun, tough immigrant childhood (sob, sob) and so on.

Furthermore, Solaris has advanced well by now. Version 9 is a pretty slick, competitive UNIX port, getting close to the ultimate excellence that the assassinated, still barely surviving, Tru64 Alpha UNIX was, and the applications abound - this is the best supported UNIX port, after all. Just what to do about the CPUs? Sun did have some changes in its CPU team over the past year, and we might finally see the light at the end of the tunnel:

Firstly, the new CPUs such as USIIIi, with integrated high-bandwidth DDR memory controllers a la Hammer, and very simple, inexpensive high-performance chipsets, might bring reasonably fast, commercial 64-bit UNIX down to high-end Linux PC prices. This trend will continue with USIVi etc. I won't mention years because again slide dates and actual dates never match here.

On the high end, there are many surprises to come, if they can come on time: After the dual-core USIV, which will have two USIII cores, large on-chip cache plus much faster system bus, two years down the line we should see US V, which is supposed to combine 2++ GHz highly parallel, out-of-order, multithreading, even dual-core in one of the revisions, could have a shot at the performance pinnacle. Another future CPU, off the main evolutionary tree, is expected to have four USIII cores with large common cache on one chip - good for say Blast bioinformatics throughput-based task loads. Again, if executed timely, and if the SPARC architecture can be a bit simplified and cleaned up for better performance and scalability [Ed: Alpha-betised?], Sun may rise again with a vengeance.

But, what if things go wrong? Well, in the worst case, Sun can count on...

AMD Hammer: The Dark Horse?
There we go - everything you wish is in: perfect X86 compatibility (it is X86, after all) at top speeds, a reasonably good, fast 64-bit bolt-on which has gone about as far as you could go on top of X86 (more registers, SSE2 FP etc), balanced integer&FP performance; integrated high-speed, cheap DDR, memory controller and standardised inter-CPU cum I/O interconnect - in summary, an Alpha EV-7 (or a Transputer grandson) reborn as a PC CPU. And yes, it's affordable!!!

Oh, just two minor problems. One: it's not shipping yet - expect Q4 (no, it's not precise) for the desktop "Athlon(64)" and Q1 for the server "Opteron" version. The other is, it comes from a company named AMD. Michael Dell might say: "Don't bother, these aren't the guys who bankrolled me all the way". OK, forget about him. But some irritated users and partners might say: "Geez, these guys' mindset and track record in handling AthlonMP doesn't really match a server CPU supplier - we got problems with overheating, or no-one addressed poor memory performance, or what about extra reliability features". AMD got to further elevate its thinking and way how partners and users are handled if it is to succeed in the workstation and server space - even if there is no Tier One vendor to back them right away (one would come their way sooner or later if AMD takes correct steps first). Intel's Xeons, for instance, are designed from the ground up to perform well and reliable in a workstation or server system working non-stop: I never had problems with any (P4) Xeon, but did have some stability and RMA issues with AthlonMP and its early Tyan boards. Things have improved since then somewhat.

If Hammer processors, both desktop and servers, come out on time, they may be at least the integer performance leaders, and close to the top on FP, for a while. The point is that, for the first time since affordable, but not widespread, Alpha clones three years ago, we'll have inexpensive yet very fast 64-bit systems at PC price levels. You don't need to wait for 64-bit apps ports to use your Hammer system to the fullest - while the port is done, you can use the existing 32-bit X86 apps at top speed, and then gradually move on to 64-bit stuff as it arrives.

AMD needs to put maximum effort in tuning their upcoming systems: for instance, the initial Clawhammer (Athlon) boards we saw at Computex support up to DDR333 (PC2700) SDRAM - on a single channel config, that gives you 2.7 GB/s memory bandwidth - still much lower than P4/533 FSB 4.2 GB/s. Yes, the on-chip memory controller will reduce latency drastically, but the bandwidth difference will still stick up like a sore thumb.

So, AMD, please enable DDR400 (PC3200) support on both Clawhammer and Sledgehammer (Opteron). If, for instance in workstations, one could use quad-CPU Opteron board, each CPU having dual-channel DDR400 SDRAM, you'd have up to 8 GB RAM capacity at 25.6 GB/s total bandwidth even if you use just one 1 GB non-registered DDR400 DIMM per channel! If you use three registered 2 GB DDR400 DIMMs per channel for a server, there you go - 48 GB RAM on a single board, with bandwidth that in theory lets you stream through all of it in two seconds...

AMD does have a good Hammer roadmap, supposedly moving to the 0.09 micron SOI process in late 2003 (maybe let's say early 2004 to be safe), and covering the one CPU to eight CPU server, workstation, desktop and mobile space across all price ranges with a unified, stable, 64-bit architecture that also runs 32-bit stuff like a champion. We still got to see how well these plans actually get executed.

The rumourmongers say the first Opterons should have a much faster SPECint than the McKinley (but maybe not than the Madison or POWER4+), and about the same SPECfp as the McKinley (but of course lower than the Madison or POWER4+). Why compare against the forthcoming competitors? Simply, we're talking about Q2 2003 here for Opteron, and, even if using the lowest-bin say 1.33 GHz Itanium3 (sorry, Madison) with 6 MB cache and 533 FSB, we're expecting at least 1,000 SPECint and 1,700 SPECfp 2000. And IBM POWER systems should reach 1.8 GHz by then, maybe reaching something like 1,200 SPECint and 1,800 SPECfp 2000 on their side, too - both Intel and IBM are expected to also make compiler optimisation improvements on their platforms in the meantime.

Keep in mind, Intel can easily get a X86-64 based Pentium out of the door anytime if they want to, and it may not be entirely impossible to marry IA-32 and IA-64 on a single chip - look at Ace's Hardware recent story on more issues related to this.

Why not than marry Alpha EV-8 and IA-64 on the same chip, than gradually let Alpha take over and run IA-64 stuff through a translator [Ed: stop daydreaming]. You may also see the Itanium platform get some unexpected performance boosts in the McKinley/Madison timeframe due to the compiler improvements to finally make better use of their immense resources.

Also, if Prescott and Nocona still manage to keep a substantial 32-bit performance advantage over the Hammer family in 2003, which is quite possible if Intel manages to break 4 GHz with these babies next year, and continues the price pressure, Hammer might appeal to a much lesser user base, those who need 64-bits early on their X86 system.

Sun + AMD ?
Let's assume that SPARC continues in the Solaris space for a while if the new chips come out within reasonable time. Is there still a space for AMD within Sun? I believe yes. Sun needs to have a strong Linux presence to avoid some of its key users having to deploy Linux portions of their installations using HP, IBM or, God forbid, hardware from an overgrown garage assembler called Dell [Ed: you mean "Mr. -1% R&D"?]. At the same time, they're not really making love with Intel (you can see who is THE ONLY target of most Itanium2 benchmark comparisons...). So, why not use their Cobalt brand to have a parallel AMD Linux-based line in parallel (officially no conflict) with their SPARC Solaris systems? They could even do quad-Opteron workstations and server cluster nodes then, while keeping the two worlds separate from the sales point of view. And, ultimately, they could "test-port" Solaris to Hammer - just in case...

At the same time, while there will surely be a huge following for AMD among Tier Two and lower assemblers and techie end-users, AMD still needs at least one major Tier One player to really "validate" the new platform in the eyes of the enterprise market. I'm not sure that IBM, HP, Dell or any other major vendor will jump first - it may have to be a fresh face beaming with Sun-shine.

Well, the conclusion is simple: Sun still holds on dearly to their UNIX market leadership, but the performance issues are eating that away. There is a program in place to get some sparks fly out of stagnant SPARC, but, whether that bears fruit or not, next year we may see a completely new Tier One high-end PC vendor: whether the systems are called Sun or Cobalt, doesn't really matter - these two sorely need each other, and it better be a passionate romance for a long term if they're going to make successful, attractive offsprings... µ

Watch out for the third instalment in few days' time...

See Also
The 64-bit saga: POWER4 versus Itanium