Its a very big HPC world. According to IDC and others, its a $7B+/year world, growing at a nice healthy clip, 10-20% CAGR, again, depending upon which market research report you read.
Seems like a hot market… right?
Well … sort of.
The largest growth (>15% CAGR) is in the 10-25k$US region (small computing engines), with the higher end stuff coming in at a somewhat anemic (~5% CAGR) rate. Still, a 5% growth rate in a market this large is nothing to sneeze at. After 3 years, your market has grown at the high end by 15.8 % assuming that growth rate. In the high growth rate arena, a 15% CAGR would yield about 50% more market size after 3 years.
Again, these are nothing to sneeze at. One might think that this market would be getting some serious attention from investors.
After all, there are all sorts of new and interesting technologies coming out, some of which have real potential impact, and in some specific cases, possibilities for much larger markets. Right now, the largest consumer group of supercomputing hardware are the millions purchasing game consoles with significant processing power. This is not represented in the various studies we have read.
The new and interesting technologies coming out all seem to cluster (pardon the pun) about special purpose computing systems. The general purpose computing systems have in essence hit a wall in terms of designs, where the cost to re-design and re-architect is far greater than the cost to do die shrinks, add multiple cores, and so forth. More importantly, each new general purpose processor absolutely must be compatible with the previous general purpose generations if you want to have any hope of market acceptance. The market has decided that momentum is good. Whether or not that is technically the right thing is debatable, it is however, the reality of the general purpose computing market. This effectively relegates a number of technologies rapidly to the dustbin of computing for this market.
The special purpose computing technologies don’t necessarily have this particular baggage to carry forward. However, they have their own set of problems.
* They are hard to program
* There are no "standard" parts/technologies
* There are no "standard" reference designs
* The current board vendors pricing is simply astronomical
The part about being hard to program is somewhat an indictment against the current state of affairs of programming education, where everything is done on an IDE, and frankly the IDE users either lose, or worse, never acquire the skills required to do the sort of programming that technologies such as reconfigurable computing requires.
No standard parts refers to the wide range of choices in reconfigurable computing technologies, as well as DSP and similar computing technologies. This isn’t such a terrible thing, until you realize that the software industry (which drives hardware purchases) didn’t start getting moving until it had a single platform about which to work. Multiple platforms increase costs. Too many platforms are a distraction.
No standard reference designs also refers to the wide range of choices among cards. This also leads to the final point. When you spend 10-20k$ on the hardware alone, and it is hard to (re) program (the field part of FPGA is still somewhat of a bad joke for application programmers in C, though it is getting better with the advent of C->Verilog/VHDL compilers), it is often very hard to articulate the value of the hardware, when the software license will cost 10-25k$US, and the customers budget is 10-25k$US.
On the other hand, if you have a system which can be built and sold as a hardware card for a 2-4k$US, with requisite software in the 10-20k$US region, is adaptable, programmable in C/Fortran, and thus targeted at the largest and fastest growing section of the 7B$US pie, specifically at the 3 segments that comprise $4BUS of this market, one might expect that VCs would be falling all over themselves to find and grow a few good startups here. The one that wins is going to make a killing, in large part because supercomputing technology goes downstream and gets commoditized over time, so if you can facilitate this and own key sections of that technology …
Of course, you would be wrong about VCs being very interested in this market. There is something decidedly “un-sexy” about high performance computing from their perspective.
This is a shame. Supercomputing is being pushed to the desktop. Someone with a viable desktop supercomputing product (no, not “the grid”(tm)) ought to do well if they can get the funding they need.
And that is the problem, isn’t it.
We have been asked about this a bit. In short our experience with very interesting and high potential accelerator designs, benchmarks, and business plans have met with a “ho-hum” by the private equity community. We have been talking about this for years, and find it mildly ironic that heterogenous computing is now all the rage in HPC, using the same arguments we have been making from the outset. As noted later in this blog, we have sought capital sources, yet few have expressed anything more than mild interest in high performance or accelerated computing.
Meanwhile, since this was written, IDC has indicated an HPC market of $9.2B last year, up from $7B. Clusters, the area we play in now, has risen to about 50% of the market, and is growing at 60+% CAGR. Everyone we have spoken to indicates that the story is great, and the numbers are good.
But we are out of style. Who wants to invest in something that could catalyze tremendous change and increased value in an existing market, when there are so many RoR startups to go after, or open source startups to form, or non-revenue producing video clip sites to fund?
I think the problem is how to makeover HPC to encourage interest. Its not the business case and model, its not the numbers, or the presenters, or the team. HPC is just not sexy. RoR is sexy. So are image/video sharing sites. Like VOIP was.