Supercomputing in the financial crisis

SUPERCOMPUTERS have always been the cream of the crop, the technology achievement showcase for the vendors and, for some like nearly-defunct Cray and finally-defunct SGI, their bread and butter for decades. That likely won't change, long term, but the global supercomputing market may slow down considerably, at least for a while.

With high performance computing (HPC) CPUs having become commoditised, as PC clusters have mostly supplanted esoteric vector and RISC processor designs, it has become harder than ever for the HPC vendors to make decent margins on these number crunching monsters.

Then there's the current financial crisis created by the greedy banksters. As if it weren't hard enough for vendors to compete with each other in academic HPC deals before, with the users squeezing the last drop of blood from them, now HPC users' budgets have really been cut across the board and across the world.

Let's look at an example. Last November there was a mid-sized 1,000-core cluster deal here in Singapore. It was supposed to be Nehalem EP but, due to changed schedule constraints, it became a Shanghai Opteron set up. Two vendors competed in the finals, IBM and HP, both making bids at steep discounts yielding what were rumoured to be somewhere around -20 per cent margins (yes, minus 20 per cent, but then vendors' costs are often inflated anyway).

At the end, HP won the deal with a minor price difference, barely enough for a Michelin three-star restaurant dinner with your honey and a pair of Manolo Blahniks for her thrown in. The customer must have been even happier than that imaginary 'demanding darling', as it ended up with 32 large eight-way, 32-core 128GB ProLiant DL785 SMP Opteron systems in a tightly coupled Infiniband interconnect setup with all the extra bells and whistles plus support for only about $800 grand.

Now, that's not much to pay for a system with approximately 10 Teraflops peak performance. The Singapore government should thank and generously reward the institute's humble yet very knowledgeable civil servant in charge who wrote the tender and squeezed the vendors. Do your sums based on last year's Opteron system pricing and you'll see what I mean. Unless the vendor treats this as an advertising effort - which IBM's Blue Gene was, in a way - there won't be much bonus money for the corporate sales droids out of this deal. And, yes, it's much more brain-stretching work for them compared with the usual run of corporate sales. "You want Fortran? What's Fortran?"

So, as you might understand, High Performance Computing has lately become less attractive for profit oriented IT vendors: way high system demands in every aspect for barely-there profits, tough user acceptance tests with high risks, and little customer loyalty, as they always want the latest, fastest kit available. In the HPC market, every sale is a fresh date like in that movie about many first dates with an amnesiac girl, there's vicious competition as though each deal is like a billion dollar outsourcing contract, and every sale is a mental challenge for the account sales force.

No wonder that, after SGI's death and dismemberment - sorry, 'Rackable' will never sound as sexy as 'Silicon Graphics' - and the near-death of Cray, as well as the financial crisis forcing further customer budget cuts leading to more price arm-twisting, some large brands are said to be re-evaluating their options relative to the value of their general HPC market presence. My intuition - I don't dare call it even a rumour yet - is that IBM could be among the first of the big vendors to, let's say, 'de-emphasise' its HPC market activity.

Not that IBM is about to throw in the towel completely, but soon you might see Big Blue keeping HPC to more of an advertising-like effort of human-versus-computer competitions and such promotions, along with a narrow scope for government-related deals where it has a firm footing for whatever reason. After all, IBM's upcoming Power7 ubermachines will still sell to financial institutions trying to calculate how the hell, if ever, to make back those trillions of dollars the investment bankers gambled away, and those will be more attractive sales efforts for it than wasting time on loss-making showcase university deals.

Exceptions? Of course there will be some. The major exception, the defence and intelligence market, is still alive and well of course. After all, every major financial crisis in recent centuries was soon followed by a massive war, so this one might not be different. The armies, navies and air forces - plus the foreign espionage and internal surveillance establishments - of the US, China, Russia, India, Japan and that hodge-podge called the EU will therefore acquire more and more HPC systems, and will probably do so at decent margins for the bidding vendors due to the incestuous relationships within the global military industrial complex.

Another exception might be 'visualisation clusters', my favourite flavour. Not only the system nodes, which are essentially souped-up workstations, have to be the very top of the line, but also everything else - from memory size, interconnects, storage, GPUs used for parallel OpenGL visualisation, to the contrast, colour, sharpness and edge blending on the 3-D projector-based CAVEs or walls - has to be tuned to the hilt, for the right price of course. Not many vendors do it well yet, and yeah, if you know your stuff, you can make money. Over a year ago, HP was said to have made a bigger profit on a 200-node visualisation cluster at Korea's KISTI institute than Sun realised on a five times larger main compute cluster in the same facility during the same timeframe. µ