Sun rising again with Opteron in sweet spot

SUN IS AGAIN in the headlines - this time for some positive stuff, believe it or not. If things like Java and, soon after, Solaris, go open source, Linux fans are bound to suddenly like some Sun-shine. Then, its own SPARCle has diminished substantially with the marching orders for the UltraSPARC V team, but the combination of Opteron from the bottom, and Fujitsu's truly SPARCling processors on top, might just be the right thing for the beleaguered Silicon Valley player.

Opting for Opteron
Sun is making some nice steps forward - its Opteron move, which we highlighted almost two years ago, will take a part of the SPARC user base ultimately, but, for Sun, it is still better that they be taken by another Sun platform then, well, by a non-Sun platform, isn't it? With its wz1100 dual-CPU Opteron workstation supporting both Linux and Solaris (Windows would probably run anyway as well), you'll now have a beautiful looking branded 64-bit X86, non-Intel, workstation for the first time. And the work with Serverworks on the large-scale Opterons beyond 8 ways can only help - Opteron is very scalable, and runs all the software, without the application support headaches that partly killed Alpha and that could sink the Itanic barge as well.

Why is Opteron good for Sun? Well, Sun's own SPARC efforts have created dismal results CPU performance-wise over the past, what, ten years? The last time Sun was any kind of workstation performance leader was maybe SPARCstation 1, in the 486 era? The complexity of SPARC architecture (in some aspects, like register windowing, not dissimilar to Itanic) and disastrous underperformance by Texas Instruments semicon process for the subsequent SuperSPARC and UltraSPARC designs, has pushed SPARC out of the RISC performance race and even below all typical PC processors. Ever since Fujitsu got its SPARC64 V, then VI, ready, my point was that, if SPARC is going to live, it will have to be the Japanese cook making the dish this time.

On the other hand, Opteron was a surprise of its own, a good mix of excellent Alpha EV7-like SMP interconnect and memory design, but through a cheaper and fully open platform like HyperTransport, and proven Athlon K7 core. So, K7+EV7=K8 ? Well, about that, plus a sprinkle of extra enhancements, of course. The resulting performance, where 2.4 GHz Opteron / Athlon64 consistently holds its own against ALL platforms, X86 and RISC alike (not to mention the EPIC one), on many different apps, in both single-CPU and SMP tests, is simply astounding. Hey, this is a bloody X86 after all! Yes, it does have bigger register sets and a bit cleaner architecture, but... How could it beat a beast with double 128-entry register sets and six instructions per cycle issue? Well, it could - good compilers like PathScale C and Fortran, with their Cray-like optimisations, do help a lot.

HP case helps
More importantly, AMD, some parts of which still seem to lack even rudimentary strategic marketing and competitive skills in the high end arena, didn't really have any Tier 1 vendor truly committed. The sad IBM story, with sudden disappearance of its e350 quad-Opteron model just about as it was to be launched two months ago, and less sudden re-focus on Xeon64 at the same time (don't ask me why this happened), is one example. The surprise there comes from HP, the bastion of Itanic, who slaughtered two of the finest-ever heroes of the processor wars, the Alpha and PA-RISC, for the "promised thing" (something akin to, in LOTR Tolkin-speak, killing off Aragorn and Legolas, to replace them with one of those disgusting Uruk-hai, since the latter one "promises better fighting performance" based on the spec sheet, plus utter ugliness as a bonus to scare off enemies - but it is small, Hobbit-like platforms like ARM that save the day at the end).

What's the surprise? Well, the best quad-Opteron today comes not from Tyan or Iwill or Sun or IBM, but... from HP! Their ProLiant 585, looking a bit like long-gone, sleek quad-CPU AlphaServer ES40LP, is a 4U system with highest-ever memory capacity in such format (32 DIMMs - 8 per CPU - for up to 128 GB RAM if using 4 GB DDR DIMMs), but also multiple PCI-X buses with two exposed PCI-X 133 buses for, say, dual-rail Quadrics QsNet II, good storage capacity etc. Most importantly, it is their own design - not a Newisys or other package with a label stuck on its front panel. I wouldn't be entirely surprised if HP launched a similarly configured quad-Opteron workstation, too, it could sell well.

I guess, if Alpha was around, now they could have a quad-EV8 21464 2.5 GHz racehorse in that package, and sweep the competition, from Armonk, NY to Tokyo, Japan, like a turbocharged vacuum cleaner. Too bad, well - a nice thing got murdered, but a few got fat gilded handshakes afterwards... that's life.

Where to for Sun's Opteron?
These are important lessons for Sun's hardware business - since the move to embrace Opteron is made, then do it fully, not half-heartedly. That means, everything from an entry-level dual-CPU workstation, to quad-CPU workstations, 8-CPU 4U servers, and even large-scale machines up to 128 CPUs! After all, there is still no large-scale 64-bit X86 system around. Not to forget 64-bit workstation notebooks with OpenGL graphics, too - something like a 64-bit version of Dell M60 with even better screen and nicer design. A full product line will not only enable greater market share, but also give the prospective buyers more confidence that Sun is truly serious this time - enhancing sales across the board.

And, do them properly - push the performance to the limit within each segment. Any crippling, say not to outdo SPARC too much, would back-fire. For instance, a quad-Opteron workstation with dual PCI-E x16 graphics buses for multi-card OpenGL 3-D - one each per PCI-E bridge (you can attach one of each per CPU via HT), and 16 DIMMs at least, optional 32 DIMMs, with speed or capacity choice, would sound very nice. Couple it with tryly nice and exotic enclosures, quality build, and good pricing, and there you go! Having a 64-bit native Solaris 10 together with Linux and 64-bit Windows can only help.

If would, of course, help even more if AMD's move to 90 nm goes well, its memory bandwidth and FP SSE2/SSE3 unit gets further improved, and, of course, more HT channels get added to the high-end versions. A top-end Opteron chip with, say, four or five HT 2.0 channels plus larger cache, would make do for a very scalable 16-way system without extra chipset logic! And that is not counting the dual core option...

SPARCs flew between us...
Well, with Fujitsu and Sun working together now, it does make sense to bring Fujitsu SPARC64 VI into the high-end workstation segment, too. One thing Sun always had was its depth of workstation application support, unmatched by any other non-X86 platform. What it lacked was a good high-end platform that is decisively ahead of PC in performance and scalability. That is where Fujitsu SPARC fits perfectly.

Having a good, say dual and quad, band of SPARC64 VI stations, will nicely round off a viable workstation/server line with US IIIi at the low-end, and huge multi-SPARC64 systems with up to 128 CPUs at the high-end. Again, no performance cuts - it must be tuned up to the hilt in everything, in particular memory and graphics systems, if expected to beat anything PC market (including Opterons) throw at it.

That's about it for now. As a "hardware boy" I can only comment on the hardware side of business, but, I guess, there's plenty of opportunities there for a fix and,, why not, a business rebirth in a way - after all, Sun did start as a unique, quality high-end hardware vendor. µ