Convex and Hewlett-Packard Make a Deal HP and Convex Computer launched an extensive array of technology and business agreements in March, 1992, and followed through with additional ones in October 1992. The March announcements saw Convex choosing HP's 100MHz PA RISC for its work-in-process massively parallel computer (MPP) system and HP buying five percent of Convex. This was a good design win for HP, displacing MIPS, which had fallen behind the power curve. Convex and HP are also exchanging compiler technology. With the announcements, Convex began simplifying its product strategy, dropping a multiprocessor under development in favor of the new PA RISC-based MPP. Convex plans to tie together its main product line, the C Series, with the future MPP machines via OSF distributed software. The C Series is Convex's main product line, and uses its own custom chips and architecture. What was missing from the March announcements was a way to tap into client- server computing via powerful desktops. The October announcements are a step in this direction: HP Series 700 workstations with Convex cluster software -- Convex is offering new cluster software products, with integration and consulting support, on the Series 700. (The products are ConvexPVM, ConvexNQS+, and ConvexMLIB.) These transform a network of 700s into a high-performance, loosely coupled parallel computer. There's a lot of bang for the buck here -- CERN in Geneva has 16 700s delivering twice the power of "dead meat " IBM 3090/600J mainframes at less than a tenth the price! Meta Series -- A Convex C Series supercomputer tightly integrated with the series 700 workstations and designed to deliver very high levels of application-throughput performance. Meta integrates the vector and parallel processing of a supercomputer with the scalar processing of workstations, just what customers need in computational chemistry and seismic analysis. Other undisclosed products are supposedly in the works. Get Ready to Push Back the Obvious Sales Objections Be aware that industry scuttlebutt views H-P as a reluctant bridegroom in this relationship -- perceived in some quarters as having wooed Convex just to get them to sign on for the PA Risc chip. You'll have to convince your prospects that the deal is real and not just another Advanced Computing Environment (ACE) du jour. With the H-P name behind Convex, you should also be able to push back on any negatives that Convex is under pressure -- in a failing minisuper marketplace. The relationship can produce very positive results for customers in this confused environment -- if the HP field is viewed as enthusiastic about its potential. Remember that both IBM and Digital will be positioning aggressively with customers here in 1993 -- the HP field MUST NOT fall off the radar screen as customers plan for 1994 and beyond purchase orders and transition strategies. If you want to participate in this market -- this year is the time to put a stake in the ground with customers. Who's On First In order to position this relationship effectively, you need to know the basics about the turbulent supercomputer market, long dominated by Cray Research (Minneapolis), as well as how this market is changing. So before moving to some details about major competitors, we first present a quick run down of the market terms and conditions. (Note: do not confuse Cray Research with the troubled Cray Computer, the company Cray Research formally spun out; it is located in Colorado Springs and headed by living legend Seymour Cray.) Supercomputer Background Highlights Convex is the only remaining so-called minisupercomputer supplier. Minisuper-computer suppliers had a straightforward business model: offer a quarter-of-a-Cray for a-tenth-the-cost -- or less -- and let the technical community simply recompile their Fortran decks. Now-defunct Convex rivals included Alliant, Celerity, Cydrome, Elxsi, Floating Point Systems, Multiflow, Scientific Computer Systems, and Stardent, to name the principal players. What went wrong? This was yet another example of "too many firms chasing too little business." Cray Research had enormous profit margins, and started using them to make deals. The wee ones simply lacked the staying power against this uniquely entrenched competitor. Fighting back further, Cray Research began selling an entry-level system (the CRAY EL) directly, and later exclusively through DEC, limiting Convex's top-end growth. Late in the minisupercomputer game, Silicon Graphics, a "cult company" with a strong following, quietly began marketing its powerful servers as symmetric multiprocessors, cutting into minisuper revenues. Now, SGI is making serious public noises about taking on the entire supercomputer market against all comers. If all that wasn't enough, a whole gaggle of newcomers have replaced the minisupercomputer pretenders to Cray's throne. These are the MPP suppliers, who are competing viciously with each other to take marketshare from Cray Research. They are succeeding -- Cray lost money for the first time in its history, after three years of flat sales and declining profits. And, the recession and the end of the Cold War did the rest! Complicating the competitive scene are various alliances that are described under the individual companies. But first, the buzzwords about the newcomers. The MPP Fandango You'll hear a lot about Massively Parallel Processing (MPP), all of it confusing and most of it pure hype. Here's what you really need to know. Much of the MPP industry sells on advanced theology of how the machines work -- rather than what they do! Some are prophets of SIMD (Single Instruction Multiple Data), the others apostles of MIMD (Multiple Instruction Multiple Data). This is comparable to selling watches based on analog or digital movements; they both tell time, and only technoids care how it is done. In scientific circles, however, some customers (notoriously so in the national laboratories) will consume hours of field time asking for finer and finer details of these trivia. Do not get ensnared in these religious tarbaby wars; it's a sure-fire lose-lose! SIMD uses one set of instructions to work on a whole bunch of data in lock step, while MIMD has a whole bunch of processors working on different data, more or less independently, depending upon which brand it is. Both are difficult to program, but in different ways, and that's what the argument eventually boils down to. MPPs are hideously complex systems, and require extensive interprocessor communications and memory management. These are unobtainable in Intel x86 chips, where the design center is the low-end desktop, so all the MPP suppliers use either custom chips, Sparc chips (Thinking Machines and Cray's acquired FPS unit), Alpha chips (DEC, Convex, Cray Research) or IBM's RS/6000 Power series (IBM). Now for a brief rundown of the major competitors: Thinking Machines Corporation (Cambridge, MA) TMC has changed its tune a few times, first espousing SIMD with the Model CM-1, which was followed by the CM-2. In an abrupt about face in late 1991, TMC said that MIMD was the only way to go, introducing the CM-5 (there were no CM-3s or 4s). Not only did the CM-5 leave many earlier customers in the lurch, the floating point unit just didn't work for most of 1992, which seriously impacted 1992 revenues, which were as high as $90 million in 1991. According to TMC critics, the company is being kept alive through massive DARPA subsidies, which has prompted more than one congressional enquiry as to fairness -- DARPA is supposed to support advanced research (the AR in DARPA), not production machines. 1993 is a make/break year, particularly as a much-rumored on-again- off-again acquisition by DEC is again in the off position. The CM-5 uses the Sparc chip, a bizarre choice, since Sparc is old and designed for cheap commodity desktops and lacks the glue and other nitty gritty that makes or breaks MPP. Systems begin at $750K. Theoretical top-end "bragging rights" machines take a power plant to operate, a Niagara to cool, and a bond flotation to finance. You should not see TMC in competition at the low end, although they are desperate for any and all business. Intel Supercomputing (Beaverton, OR) Intel has had a series of MPPs, the most recent ones based on what everyone else says is the wrong chip: the Intel i860. Alliant went bankrupt with it, Stratus tried it for a while and backed off, and even Intel corporate says the i860 is no longer mainstream. With the PC chip business booming, Intel is fat and happy, but not dumb, and rumors of the Directors pulling the plug on this red-ink operation just don't fade away. Intel's current line of systems is called the Paragon, with process starting at under $1 million. If you se them, it will mainly be through Digital. Cray Research Trying to regain market mindshare after its poor corporate performance, Cray Research bought certain of the Floating Point assets from the bankruptcy trustee, and is now marketing the Sparc-based FPS box that FPS literally bet and lost the company on. Prices start at $500K. You may see this nearly abandoned stepchild in some sales situations, but even Cray will not promise a long-term future for it. In MPP, Cray found that it couldn't beat the MPP crowd, so it joined them, annou ncing an Alpha-based MPP for a vague 1994 initial delivery. While the unit will consume gobs of power, it eventually will have the Cray software library. If you see it, demand benchmarks; everyone else has new models on the drawing boards, too. MasPar Computer (Sunnyvale, CA) MasPar avoids the high end, selling MPP products in the $75,000-to-$1.6 million range and choosing only those application areas where its flexible "autonomous" SIMD (called ASIMD!) can deliver dramatic gains in performance and price-performance on computationally massive problems. In October 1992, MasPar upgraded its line with the MP-2 with five models that are one-and-a-half to five times the speed of the earlier MP-1s. These are based on MasPar's new custom 32x32 Risc CPU technology, which implements 32, 32-bit processors on a single chip. Multiple chips are used to make up 1,024-processor boards which in turn make up the processing portion of the systems. A DECstation 5000 is used as a frontend processor. Digital Equipment Digital is playing the role of British merchants Fortnum and Mason -- "Purveyors of the Finest Goods to Her Majesty." While it is developing its own Alpha-based MPP system, it is repackaging and reselling the MasPar (SIMD) systems, a selection of low-end Intel's (MIMD), and entry-level Crays (vector). Also known as the Burger King play -- have it your way. Digital is also heavily investing in software to make MPP easier to use, regardless of architecture. It leads the High Performance Fortran Initiative, which has moved surprisingly quickly in agreeing on common standards. Note that Digital takes sheer delight in retaking a Convex account. "After all, Convex built their base from ours!" Digital MPP product line consists of two distinct families. The DECmpp 12000/Sx is a SIMD system based on MasPar's MPP products and the DECmpp 12000/Mx is a MIMD system based on Intel Supercomputing's products. Digital first introduced its MasPar-based 12000/Sx family in late 1991 with the Model 100, and expanded the family in October 1992 to include more powerful systems (model 200) based upon MasPar's newest products introduced at the same time. List prices for systems range from $175,000 to $1.7 million, including base disk capacity and frontend processor. Digital introduced the first model of its Intel-based DECmpp 12000/Mx in September 1992, and currently markets four models that range in price from $630K to $1.5 million. Kendall Square Research Kendall Square Research is the latest MPP power player, being run by former industry wunderkind Henry Burkhardt, cofounder of Data General and Encore Computer. KSR is the company to beat in 1993. It went public in 1992, has no cash problems, and will break even in 1993. KSR shipped its first systems in late 1991 and is ramping up fast. Be careful, it takes no prisoners! The KSR series is far easier to program than other MPPs, using a novel architecture that retains the "conventional shared-memory programming model." This masks the underlying complexity of MPP and allows programmers to quickly port old applications or develop new ones onto a single processor, and tune them later. The product family consists of 36 systems in configurations ranging from 8 to 1, 088 processing units. The models scale linearly in all major dimensions of performance and capacity. List prices for KSR1 computer systems, including base disk capacity and peripherals, range from $500K for the KSR1-8 to over $30 million for the KSR1-1088 The KSR1's operating system, KSR OS, is an enhanced version of OSF/1 UNIX. It runs symmetrically on all processors and provides flexibility in allocating computational work among multiple processors. It simultaneously supports multiple computing modes, including batch, interactive, OLTP, and database management and inquiry running concurrently. KSR OS supports the most widely accepted Unix interface standards, AT&T SVr3, BSD 4.3, Posix 1003.1, XPG, and FIPS 151.1. MOTIF and X-Windows user interfaces are in the works. Kendall Square Research is aggressively porting application-software packages on to the KSR1, notably computational chemistry, computational fluid dynamics, petroleum reservoir modeling, signal processing, and supporting mathematical libraries. We expect an Oracle port this year. IBM IBM just pulled the plug on the "machoflop" machine it was backing, the one designed by Supercomputing Systems Inc (SSI), headed by Steve Chen, former Seymour Cray protege'. Years late in delivery and based on vectors which had since gone out of fashion, it is not a competitive factor. IBM has made so much hay of its workstation farm at the Livermore National Lab displacing a Cray, that Cray issued a news release saying it wasn't so. Nonetheless, it is a real competitive factor for H-P/Convex cluster systems, since they operate on the same principle. We expect IBM to increase its supercomputer marketing intensity during the first half of 1993. The company announced an internal joint relationship between its Enterprise Systems Supercomputer Group and the Austin-based RS/6000 AWD Group in February, 1993. It will begin to beta test its PowerParallel 8-64 processor RISC-based systems in Q2 and Q3 of this year. As a result, IBM is clearly attempting to hold off competitive threats to its vector installed base until the market steadies with deliverable parallel hardware and applications software technology in the 1994 timeframe. Note that IBM has for some time been cutting back on technical-market field support, shifting them into their consulting division, elsewhere, or out. Under yet-to-be-named new management, expect more cutbacks. Aberdeen believes that it will be relatively easy to create suspicion about whether or not the PowerParallel machines are for real -- or when they will actually do useful work -- against the backdrop of a company in serious financial trouble. However, HP will have to really SELL its own commitment to supercomputers if the Convex deal is to succeed. nCUBE (Foster City, CA; near Silicon Valley) nCUBE is a smaller, quieter MPP suppler, focusing on a oil, autos, and more recently chemistry in the technical markets, but now betting heavily on the Oracle market. They are impressive here, with audited 1,000 TPC-Bs on a mere 64-processor system and at less than $2.5 K per TPS. Aberdeen believes there are bigger Oracle plans in the works. You should not see them very often, but with Larry Ellison backing them (the largest shareholder), they will be here for a while. Prices begin at $175K for the entry-level system and $400 for standard systems. Conclusions The big two producers and archrivals Thinking Machines and Intel Supercomputing are totally dedicated to putting the other out of business using a high-tech version of MAD (from the Cold War: mutually assured destruction). Last year they came close to killing each other off; this year they might. KSR is coming on strong; be careful. Digital is just getting started in MPP, but is getting better at it every day. They have a full line of other people's wares. Cray is thrashing, and concentrating on the high end, while IBM is concerned about corporate survival. With growing interest in MPP for both technical and commercial applications, plus a vigorous 700 Series, you should have a headstart in the Convex-based technical accounts, if you take time to learn a little more about the business and take a positive account management approach.