John West has a great commentary at InsideHPC.com. First I recommend, if you haven’t read it, by all means, read it.
He points out that Addison Snell (a former colleague during SGI days) stands by his HPC spending/market analysis of several months ago. Further, he notes something I saw in another press release recently, that indicates that financial services firms are remaining committed to HPC.
John’s summary of Tabor’s results notes that purchases may be deferred or placed on hold.
We are seeing no decrease in interest. Just a lengthening of sales cycles, as high performance computing has to traverse more layers within the management purchasing chain to get approval.
Frequent readers of this blog will note that I have been espousing the view that HPC is a fundamentally game changing technology. Yeah, I know … cliches. And marketing buzzwords. Marketeers have a nasty habit of taking good wording, and over-using it to the point that it becomes an entry in buzzword bingo. So when something really is describable by the adjectives in question, it gets annoying that you have to come up with synonyms just to get away from language someone else abused in the past.
It is unfortunate in this case … HPC allows users to ask the “what if” questions. It lets them ask them at a very reduced cost per answer. And this ability to rapidly cycle through these what-if scenarios, allows users to explore much wider spaces, or optimize answers, or ask better questions.
But HPC really does this.
Which is why end users can quantify the benefits to their management. And management may take longer to sign off, but they recognize that the null decision (doing nothing) could cost them more in the end.
So I agree that HPC purchases may slow. We have seen some of this from the company side of things. More layers are being traversed to get to decisions.
On the other hand, we are not seeing a slowdown in interest. What we are seeing is a desire for companies to optimize price as well as performance.
John continued to talk about pure play companies. Cray and SGI as the major examples. Cray is in IMO a good state. They have constrained their costs, developed valuable technologies, didn’t build out the whole stack, just where they could be different. And they are selling. This is good. They have a few competitors, but they have a fairly unique product offering, that fills an important niche. If they can make the down-market effort (CX-1 anyone) work, this could get very interesting. Cray has according to Yahoo finance, a market cap of $71.28M, and at 800 employees, a value per employee of about $89k
SGI on the other hand largely doesn’t make its own equipment anymore. Most of the SGI value in the days past was tied up in the graphics, or the NUMA design, or … any number of things. Unfortunately for SGI, you can build the NUMA machine that most customers want out of ScaleMP and (insert your favorite vendor’s blade unit here). If Cray wanted to, it is likely they could do this as well with CX-1. Basically SGI is directly competing with Dell, HP, IBM, … . As I have said many times, it isn’t a wise idea to try to out-Dell Dell. Whether or not they can ride out this storm is anyones guess. SGI has, according to Yahoo finance, a market cap of $29.48M. For 1632 employees, this is a value per employee of about $18k.
For comparison, IBM has a market cap of $114B, and for the 387k employees, has a value of about $295k/employee. Dell has a market cap of $19.97B, and with 82k employees, has a value per employee of about $241k. HP has a market cap of $84B, with 321k employees, a value per employee of $262k.
I consider IBM, HP, and Dell to be quite strong companies. Barring some unforseen circumstance, I don’t see any of these being more than bruised by ongoing events in the market. But HPC, as John points out, is a small fraction of their business. All of HP, IBM, and Dell have embraced Linux on their volume HPC stacks, and have other stacks to augment their volume HPC systems.
John points out that Sun is taking a beating. Current market cap at $2.93B. With 35k employees, their value per employee is about $84k. Comparable to Cray. Sun has $2.63B cash on hand. Which means the non-cash portion of this company is being valued by the market at about $300M. Its worse if you look at their forward P/E ratio, which is at an unsustainable 79.4. Should be 10-20 at best for a healthy company in a growing market. Look at IBM’s (9.71), or HP’s (8.24), or Dell’s (8.28). If Sun’s stock drops by 8x just to get the P/E down to where it should likely be, you are looking at a company valuation of $366M, which very negatively discounts the face value of the cash.
Sun has many differing challenges ahead. What to do about Linux, how to compete against Dell, IBM, HP on the high end, and others elsewhere.
John asks about differentiation, but in the context of the lower tiers. Before we go there, let me ask, how is a HP Proliant different from a Dell 1950 different from an IBM x*** box from a Sun x4200? At the end of the day, the processors and memory are all the same. As are the disks. The fascia are different, sure. As are the IPMI cards. And some of the options you can get with it. But are they different enough that you are willing to spend more money due to the marginal additional perceived value?
Dell has differentiated itself more on business process. They have made it easy for people to order from them. This is not a trivial thing. This is a barrier lowering thing. This is good.
I am not bashing these companies. I am noting that the market has become effectively a completely commodity market, differentiated only on price. If you have some special sauce to deploy, now is the time to deploy it.
Blade servers are basically the same story. Everyone has something similar. Blade servers do represent a lock-in to a vendor at the sub-rack level. Better management tools? 6 of one, half a dozen of the other. Thats in the eye of the beholder.
But they need special sauce as well. Which is why the Cisco non-announcement, and my speculation about someone grabbing ScaleMP is so interesting. ScaleMP is differentiation. In the era of virtualization-gone-wild, ScaleMP is the counter to this … SMP built on demand. Another possible suitor would be Dell IMO.
The point is that large portions of the compute side are completely commoditized, and the value in the special sauce is what truly differentiates.
With lots of people running Rocks, or switching to Redhat HPC, or xcat, or Perceus, the cluster stack isn’t special sauce (and that aspect as a market also appears to be over).
The point of all of this is that the survivors in this commoditized market will be those who can move differentiated solutions at the lowest possible cost and lowest barriers to the customer.
John then asks abou the next tiers.
What about the Penguins, Rackables, Veraris, and Bulls of the world? They???ll have a harder time mixing it up over cost in the middle market with a less differentiated product or, worse, driving down to the bottom to compete solely over price. Some of these may cease to be, unless the recovery is very swift indeed.
Unfortunately on the volume compute side (cluster nodes), it is already non-differentiated, and there is little capability to reverse this trend.
What differentiates the Penguin from the Rackable from the … from the IBM, from the Sun, from the Dell, …
Thats the issue. There are some customers who will only buy tier 1. Nothing else. They voluntarily spend more on similar gear they can get from the tier 2’s.
That is, I think I want to extend John’s last paragraph to include the tier 1’s. Their compute nodes are all largely interchangable. Which means the cost is the deciding factor.
For us, we differentiate our storage systems by raw performance, bundles (more about this coming soon), and by some innovative features (more about this coming soon, we promise), as well as price performance. We epitomize “better, cheaper, faster” for what we do. This isn’t braggadocio, our customers tell us how much faster our machines are. Sadly, we can’t use their quotes due to their corporate rules, and our respect for their privacy, but some of their reports on how much faster are simply astounding.
We are focusing more of our efforts on the lower cost aspect these days. When it comes down to it, we are not racing to the bottom … we have to be able to pay rent, salaries, etc. But our cost structures are lower, so we can live on a smaller margin.
So we have our special sauce. And customers find this of value. This is a good thing. We are not trying to out-Dell Dell.
But what is going to happen to the market as a whole? Tabor, and others think it will continue, albeit at a slower pace. John thinks (I hope I am interpreting this right) that some companies are going to face existential challenges, with some not making it. I agree. This will happen.
John also noted that HPC could disappear from some of the companies marketing efforts as an independent focus. We have seen this before from the big companies. It is hard to get attention for small market segments in large companies. Prime targets for cuts in a down market. You hear this all the time … focus upon core markets.
But note that I said we are focusing upon the price aspect of our solutions. Because customers are focusing upon this.
So how can customers, whom are suddenly more price sensitive, get the computing power they need?
Just like in the late 90’s early 2000’s, this is an inflection point. Customers clamored for more processing cycles, and clusters offered better cost per cycle (and therefore more cycles per dollar) than large SMPs. Took a few years, but clusters now dominate HPC.
We are seeing the same sort of transition now.
Users want more processing power. So companies are providing this. Intel/AMD via more cores/die, nVidia via GPUs, and ATI via GPUs. Yeah, I know FPGAs are there as well, in addition to specialized co-processors.
I won’t talk about that aspect. Right now, under $2k gets you 1/2 TF of single precision power, and in short order, even more. So if you can adapt your code to this, why buy a cluster?
This was a very similar question to what was asked in the late 90s. If your code was running MPI on a t3e, couldn’t it do the same calculation on a less expensive cluster? Maybe not as fast, but much cheaper.
So now, what if you can cast your code onto the GPU systems in a few weeks. And then get 10-100x performance delta. Is this of value? Will people flock to it?
Only if this 10-100x is easy to get. Cuda helps solve that problem.
Basically what I am saying is that I think we are seeing is a market transition to lower cost computing platforms from smaller clusters. Many users I have spoken to are interested in accelerators, and some are actively porting.
So while I am not in complete agreement with John on who is going to hurt and who isn’t, I think it will be the nimble adaptable companies whom will survive and thrive. I count Cray in this mix. And us. And folks like SiCortex. ScaleMP. Cisco. IBM, HP, Dell. A number of others. Pure box sellers are going to get beat down on price. Value sellers will have some ways to resist this. Real value costs real money. But the user has to agree that there is value. And that’s the catch.
I wonder when the first Tesla cluster will show up on the Top500.
[update] November 2008 is the answer to the above query.