The future of the HPC market ... is it growing or shrinking?
By joe
- 8 minutes read - 1573 wordsHPC as a market, is under stress, and will continue to be for a while. The Inquirer has an interesting article saying very similar things to what I have been saying for a while about the market. It is a very good read. I’ve been saying for years … no … decades now … about 15 years to be frank, that HPC has been moving relentlessly downmarket. Each wave of its motion has a destructive impact upon the old order, and opens up the market wider to more people. All the while growing the market. Identifying the next big thing early, and riding the wave … thats why HPC is such a fun market. Change is constant. Performance envelopes pushed. But … this is a market. This is a business. With customers, competitors, and so forth. This is where things get interesting .
Specifically, there is huge margin pressure in this business. The relentless pressure to drive prices into the ground has a number of interesting effects. First, all the oxygen is sucked from this segment of the market. Or at least in a metaphorical sense. Really it is the margin thats gone. More in a minute. Second, this loss of margin provides effectively an incentive to decrease investment and continued efforts in this market. Who benefits from this? In the short term, you spend less on your HPC. In the long term … what long term? This is meant in an ominous sense for this market. If the market is incapable of supporting vendors, the market effectively evaporates as vendors leave. Which decreases customer choice and options. Loss of the market means less interest in investment in this market. Which means fewer innovations. Remember this the next time someone brags about how low they were able to drive prices. Sales people missioned on revenue aren’t focused upon sustainable businesses objectives. They just care about “meeting their number”. Which is the wrong approach for a business seeking a profit. The Inquirer talks about a deal in singapore:
That last sentence is I think a fairly universal belief. It is also wrong. Bills of materials cost, curiously, what bills of materials cost. It is very hard to buy a $1000 part for significantly less than $1000. Some of the vendors provide marketing support, but they can’t really pay you to use their products (though if you speak with some … well … less smart … vendors, they may utter words to that effect). You can work hard at negotiating harder, but you aren’t going to get 50% discounts on parts. Look at the gross margins of the parts vendors. They aren’t at 100-200%. This means they don’t have lots of room to negotiate. When you assemble rather than design and build your own, you are bounded by the fundamental costs you have to pay for the materials. The bill of materials (BOM). Your margin is the percentage more than the BOM that you get for the sale. You have to run your business off this margin. A 1% margin on a $100k sale is, $1000. This won’t pay the rent on a tiny space, never mind salaries and benefits for the people doing the selling. See the problem? Your net margins are your gross margins minus your (amortized) costs. Margins are usually expressed as a percentage. A positive net margin could mean a positive profit, which you can then re-invest into new HPC goodness. A negative margin means you are paying your customer to take your gear. This is not a long term survival strategy for a business. This is a way that either forces businesses out, or the business dies not understanding what is killing them. SGI was in that latter category. Notice the rumored margins. The article goes on to point out
True across some swaths, but remember HPC as a market is changing. It is going down market … touching more people, being larger as a market in general. This is more about the dying of a segment of the market. They go on …
The acceptance testing is largely what killed LNXI. Forced them into some really bad decisions. High risks are standard. In a normal market, high risks come with high rewards. The risk absorbed by the vendor for dealing with onerous T&C;, and acceptance tests, should translate into higher prices. This is how insurance works, this is how risk is priced into a deal. But its not. Which means that the price of this risk is borne only by the vendor. And somethings this doesn’t quite work out. This scenario is simply not stable. Well run companies are going to take a long hard look at where they are making money and where they aren’t. And it is a simple business decision to close or reallocate resources involved in the loss making operations. The article goes on with … well … formenting their own rumor. More in a moment:
Well, I know a fair number of very recent ex-Big-Blue-ers. I also note which groups they came from. This rumor the author was formenting, is sadly, not so farfetched. But they aren’t the only ones. Microsoft’s HPC effort isn’t likely to hit break even for years if ever. I haven’t heard rumors, though I have seen them cut back, and lay off here in the US, and in India. Unprofitable divisions will be shuttered. Previous analysis of their HPC offering suggests that they would need huge (unachievable) volumes to hit break even, never mind a profit. AMD has cut hard and deep. John McCalpin, the father of the infamous Stream benchmark, was cut, and was taken up by Jay Boisseau at TACC. Sun has sliced very deep. Dell has trimmed some. It seems that most everyone is hunkering down for the storm to pass. Some in the market will probably focus elsewhere … de-emphasize HPC. Not that HPC is bad. It is just a small market relative to the rest of their work.
Way back at SGI, a professor at my alma mater asked me to give him free machines. Something like $100k of free machines. I asked why would we do this. He indicated that the next generation of students would learn on them and buy them. Um … that is a really weak argument. No, he didn’t get them from us. He did get some from Sun. Sun didn’t sell much at a profit there. Pretty much nothing at a profit there. Look at this as quid pro quo. What can you do as a customer to provide us as a vendor, equivalent or better value to what we give up? If the answer is “nothing, don’t be silly”, then you don’t understand the problem. Basically, the days of these sorts of freebies and all the other things that drive profit out of this market, are waning. Its an unsustainable scenario. One that will drive choice out, and decrease competition. Which will enable the remaining competitors to have more control over the pricing and their margins. Imagine an RFP with T&Cs; such that there are no responders. We got a call from a university about a year ago, asking us to bid on a system. The system wasn’t terribly hard, but the T&C; were impossible. I asked them if they would accept bids without their T&C.; They indicated no. My understanding is that they didn’t have bidders for this. I wonder why. Rhetorical. I know why. Create a bad enough market, and market forces will correct the market for you. It does this. Not always to the benefit of the consumer. Here in the Detroit area, house prices are in free fall. But they are starting to sell. House prices were inflated, and until they started hitting what was perceived to be reasonable levels, people weren’t buying. The massive oversupply in the market creates a buyers market. Conversely, massive undersupply in the market creates a sellers market. Drive the prices down hard and long enough, and this is what you create as you force vendors from the market. This is where we are in HPC. But we are doing this against the backdrop of a technological wave. Sort of like what happened starting in 1998 to 2000. Those pesky Linux clusters effectively destroyed the market for large SMP systems. Which previously had destroyed the market for large vector systems. And in the process grown the market more than an order of magnitude in size. The same thing is likely to happen again. We can throw huge numbers of cheap and easy to program cores at problems. Using inexpensive ScaleMP software we can tie together resources and build SMPs on the fly. Using inexpensive JackRabbit and DeltaV storage, we can provide massive I/O firepower. Using inexpensive many-core deskside systems, we can combine all of these technologies apart from ScaleMP’s. I think it is correct to say that the market is changing. The whole economy is shrinking, and there is nothing wrong in noting HPC is shrinking as well. However, it is such a useful tool, such a powerful tool, that I don’t expect to see it disappear as a market. I do expect to see some the bigger players, for whom this market is but rounding error, make critical tactical decisions, and alter investment. Businesses will focus where there is money to be made. Drive enough vendors out, and the remaining ones have more control over the market.