Two technology themes were prominent this year, looking at the smaller companies on the exhibition floor; solid state storage, and reconfigurable computing implemented on FPGA hybrid accelerator platforms. One example of each :

Solid State Storage

Fusion IO were using a video wall of 16 40" panels to demonstrate video streaming from their IODrive Duo; each panel being tiled to show 64 HD streams all coming from the one drive which was feeding the 1024 streams. This is a current product.
For two government installations they've built a system which can sustain one terabyte per second (1TB/s) of aggregate bandwidth with access latencies under 50 microseconds. This uses 220 IODrive Octal cards, in 6 racks. Octal cards use 8 IO memory modules, and fit a PCI Express x16 Gen2 double-wide slot. Steve Wozniak, Chief Scientist, said “We look forward to productizing the ioDrive Octal in the future, and bringing the power of this solid-state storage technology from the world of HPC to the enterprise.”
The fundamental improvement from using solid state memory compared to disk drives is in input/output time - no rotating media means no seek time. It is also possible to keep the same amount of data in a much smaller space, using less power, than discs require. It's early in the life cycle for large scale solid state storage. Analysis of application requirements and behavior will be required to determine if the higher cost of state storage compared to disc drives is justified.

Reconfigurable Computing

Convey Computer presented themselves well compared to the other FPGA accelerator vendors I met. The device combines one or more general purpose CPUs (Intel) with Xylinx FPGAs. There are application specific instruction sets (Convey calls them personalities), since the instructions which most benefit from acceleration for geophysical analysis are not the same as those for financial analytics. It is possible to use the device as a 'closed box' application platform; to recompile existing code to get quick perfomance gains, or to develop customized instruction sets beyond the supported 'personalities'. For some applications, the hybrid device takes less power and space to deliver the same performance as a commodity server.

Many of the early employees of Convey worked for Convex prior to its being acquired by HP. It raised a $25m B round in June 2009. Public customers include Imperial College, Stanford CEES, LBNL, and Oakridge NL.

Other reconfigurable computing solutions include using Cell processors, and Graphical processing units.

Supercomputing technology applies to more general problems

The annual Supercomputing events are a conference and exhibition, with the target audience being researchers who have government or industrial funding to work on big science projects. Big science projects (computational biology, climate modeling) need big tools. At one time, supercomputers were highly specialized machines made by Cray, IBM, Sun, SGI and others. Now, the Top500 supercomputer list is dominated by machines using thousands of Intel x86 processors. Half of the interconnection on the list is Gigabit Ethernet. The path from innovations in hardware and software tools developed initially for tackling large science problems to techniques which can be applied to more general commercial problems is much shorter than it used to be.

Further reading

Updated to add reviews 28 Nov 2009
SC09
Fusion-IO video wall (video)
Reconfigurable Computing research HPCwire article
HPCwire SC09 review
Networking products review The Register

Cunning Systems - Technical Advisors

Cunning Systems evaluates product and service ideas in computing and communications. If you would like to discuss an idea, contact us at info@cunningsystems.com