Arrived at SC06

Getting here was fun. Most everything went without a hitch. The TSA did not appreciate the larger sized toothpaste and shaving cream in my carry-on. I’ll refrain from commenting on this.

I managed to leave my camera at home. Go figure. I will work out a way to get photos posted to http://photos.scalability.org . Hopefully not cell-phone quality, but real ones. Tomorrow night is the opening gala. Then tuesday night is Beo-bash, and the folks at Microsoft have invited a few people to have dinner. Of course these are at the same time. The big news at SC06 is in the rise of the accelerators. Last year there were a few brave FPGA people, and some custom chip people. Now we have lots of accelerated projects: GPGPU, ClearSpeed, Cell, PeakStream, Celoxica, …. At the end of the day, the case for accelerators is in the business model, not in the technology. If you can get 5-10x out of your accelerator, what is a) the cost for you to create this acceleration (e.g. how much does it cost you to port it) b) the cost for you to deploy this acceleration (e.g. what is the per seat cost), and c) the cost to maintain it. That is, you can’t start out with “we will charge you 100x because our product is 100x faster”. Which, sadly, appears to still be a few groups model. Others have realized that they can address a much larger market by focusing their efforts on somewhat slower acceleration platforms, and delivering their value to a much larger market. Remember, accelerators as point solutions have been in the market for decades, and with OpenGL, SCSI, and other specific standards, have been turned into highly efficient and cost effective solutions to specific problems. The markets are there, the consumer demand is there. What is missing (largely) is VC money flowing into new accelerator companies … as there aren’t many small highly innovative groups out there with a good idea, a good business model, and a solution. As Deepak at the ever enjoyable http://mndoci.com has pointed out in the past, the market is and has to be pragmatic with respect to any “new” technology. There has to be a defined benefit, and there has to be a reasonable cost-benefit analysis (CBA) that drives adoption. I believe that the defined benefit is there, and the CBA makes sense in some cases. We have been making this case for years to potential investors. Someone must have burned the HPC market out of their collective minds … there is very little in the way of new capital to drive the change that looks like it is coming. 1-2 deals per year isn’t what I would call “active”. What is interesting to me is how quickly the market has switched. I had expected it to take years to grow an interest. The interest here is a strong ground-swell, literally a grass-roots effort. Significant effort needs to be made to come up with a common API for building accelerators and applications. Very much akin to OpenGL. I have been proposing this for a while, and calling it OpenHPC. Maybe a re-naming is in order. Regardless, the most interesting portions of this conference will not likely be in the large booths, but the people wandering the floor, or in the academic booths. From our perspective, this is where we (semi-officially) launch our L. Flavigularis. More about that in the next post. Others should announce interesting things as well. The groundbreaking stuff has been seen over the last few years with the accelerators, it is now time to see how to turn this into a business. Obviously we have a bias (and some good ideas on it). The rest of the conference will be a smattering of new ways to combine Opteron and Woodcrest into scalable clusters, and a few other chips.