HP shows smallest pod in the world, talks HPC
Top500 wins - or losses?
Now that every vendor has a shipping container (pod) computing solution, how do you differentiate your offering? You can’t go bigger and stick to the form factor. But wait a minute: you can go smaller… hmmm...
Stalking the floor at SC09, we believed we found Hewlett-Packard pursuing this strategy with its new pod. The diminutive pod measured around 4’ long, 10” wide and about 18” tall – perhaps enough to hold 10 blades with storage, or more without storage. For non-US readers using unconventional measurement systems, I’ll translate: 1.22 meters x 25.4cm x 45.72 cm… or, using the Roman cubit/palm measure: 2 cubits 4 palms long x 3.43 palms wide x 1.03 cubits tall. (Click here for additional conversions.)
Turns out that this was just a scale model for HP's real, regular-sized, pod… my bad… but the video of the perplexed HP booth attendants explaining that this isn’t its actual pod is kind of amusing.
On a serious note, we spent some time with HP HPC management and discussed their approach to the market. Briefly, HP wants every bit of profitable HPC business it can capture. The emphasis on the word ‘profitable’ is key. Many of the much-touted and glorious Top500 wins are anything but profitable to the winning vendor.
At times, vendors have to write a bid that guarantees both the performance and cost of a projected system two or three years before it is to be delivered – with a performance goal that is in excess of what is possible with current technology. So vendors have to accurately predict what will be possible with the next generation of gear, guess the cost of it, and then guess how much labor it will take to make the system perform adequately. They then have to put a dollar figure on it, keeping in mind that other competitors want the business too.
After the deal is awarded, the vendor team has about 10 minutes of “Yay! We rule!” gloating, followed by months of “What the hell did we get ourselves into?” wailing and gnashing of teeth, as they spend increasing amounts of money designing and tuning the system so it comes in on time.
Not our game
HP doesn’t want to play that game. This isn’t to say that it won’t engage on the big, speculative Top500-type deals, but they won’t do it at a loss. They are committed to using industry standard technology and partnerships in HPC while shying away from anything that smells of one-off customization.
You probably won’t see HP doing anything like the hybrid Roadrunner system that IBM used to deliver a petaflop to Los Alamos. However, it might well partner with NVIDIA or Intel to deliver a hybrid solution that’s based on existing (or upcoming) HP blades or racks – but only if it makes business sense from a profit and loss standpoint.
The obvious advantage to this strategy is that it ensures that HP’s HPC business pays its own way – and then some. This is a well-accepted business model, and much superior to the “Sure, we lose a little bit on each unit, but we make it up with volume” strategy employed by others.
But there are some advantages to taking on the bleeding-edge deals. In a lot of ways, supercomputing is the Formula 1 racing of the computing industry. Many of the advances we see in enterprise technology were pioneered in supercomputers. Vendors who are forced to climb seemingly impossible performance mountains are typically the first to discover and exploit these advances in both HPC and, later, enterprise systems.
There is also the cachet and PR advantage that surrounds the vendor who has built the fastest computer on the planet. These benefits are hard to quantify in terms of hard dollars and cents, while the expenses associated with a serious HPC challenge on the Top500 list are easier to see (and be frightened by as well).
Of course, when we talk about the high end, we’re talking about a fraction of the overall HPC and cluster market. The bulk of sales are in the midrange – which is where most of the growth is too. HP’s strong position in x86 systems helps them in this market, giving them a wide range of systems and configurations that can be made into solid HPC systems.
However, we think HP needs to do a better job of addressing HPC market specifics – what HP brings to the table besides their line of x86 boxes. HPC isn’t commercial; the workloads are different and the buyers are different – they’re swayed by different vendor attributes.
From our conversations with HP, it’s clear that HP knows this; but we’re not so sure that HP has proven to the market that they know it. In other words, we believe HP needs to do a better job of talking about its HPC heritage and its unique value, and a better job of proving that it can deliver what the HPC customer wants.