The GPGPU phenomenon is continuing to attract lots of attention in the high performance computing community and is starting to bring some new players into the market. The introduction of commodity GPU processors offering teraflop-level performance suggests supercomputing can now be had for near-PC prices. The current challenge is to package those GPUs so that their power can be tapped by the average HPC practitioner.

The latest attempt at this comes from Velocity Micro, which until this week was known for its bleeding-edge PCs and desktop systems for power users, especially gaming enthusiasts. On Monday, the company jumped into the HPC market by launching a new line of NVIDIA Tesla GPU-accelerated HPC workstations. The products consist of customizable desktop systems based on Intel CPUs and NVIDIA's newest C1060 card. The C1060 is based on NVIDIA's 10-series GPU, which offers almost a teraflop of peak single precision performance (and around 100 gigaflops of double precision). With 4 GB of local memory, the C1060 has more than twice the capacity of NVIDIA's first generation C870.

The Velocity workstations, which range in price from $3,995 to $16,995, come preloaded with the CUDA SDK (NVIDIA's C programming framework for GPU computing), along with either Window XP or Fedora Core 8. The hardware is available in three basic configurations: an entry level system containing a dual- or quad-core Intel Core 2 processor and an optional C1060 Tesla card; a mid-level system with almost the same Intel CPU options, but up to two C1060s; and a high-end box with single- or dual-socket Xeon quad-core CPUs and up to three C1060s. The company rates the three configurations at 1, 2 and 3 teraflops, respectively, with the GPU card or cards providing most of the horsepower. NVIDIA Quadro GPUs are also available to drive video, enabling GPU computing and visualization to take place simultaneously.

The three-teraflop configuration with dual qual-core Xeons, three Tesla cards and a Quadro GPU, consumes plenty of juice. In an attempt to max out the system, the Velocity engineers have been able to drive power consumption up to 950 watts, and that's probably the most a real-world application would consume. The systems are all air-cooled, presumably very effectively, since hot chips are standard gear on most Velocity systems. In fact, for the company's high-end consumer boxes, overclocking is fairly common, although not for the HPC product line.

The GPU-equipped machines are designed for typical HPC end-users: scientists, engineers and other technical analysts. Since HPC is new territory for Velocity, the company has partnered with James River Technical (JRT), a reseller that specializes in the HPC market. JRT facilitates deals for vendors like SGI, especially for the higher education and government markets. The Velocity-JRT partnership is an especially nice fit here since the lowest hanging fruit for these new workstations is likely to be researchers at universities and government labs.

These types of users have already shown a lot of interest in GPU-accelerated computing and are on the lookout for production-ready systems. According to JRT president Tom Mountcastle, many of their customers are constrained more by budget, than imagination. "This appeals to the research community because they like being out there on the edge," he said.

On the other hand, since the machines will be on people's desktops, the big government labs and the universities aren't interested in inexpensive systems that lack vendor support, which chews up a lot of system administration time. In this area, Velocity has a good track record. Over the years, the company has collected numerous award for craftsmanship, service and reliability from the likes of PC Magazine, CNET, and PC World.

The company is also among the first, if not the first, to take advantage of the latest hardware technology for its consumer products. In that sense, it sees the new Tesla hardware and CUDA as a game-changer for HPC. From Velocity's perspective, NVIDIA's introduction of the more powerful 10-series GPUs and the maturity of the CUDA software stack indicate that the technology pieces are now in place for a commercially-viable high performance PC. "We've determined there is a hole in the market for entry-level high performance computing and that's where our product will be focused," said Randy Copeland, Velocity Micro's CEO and president.

CUDA, in particular, seems to have reached a critical mass. A quick tour of NVIDIA's CUDA site reveals dozens of academic codes and a smattering of commercial applications and libraries that have been accelerated. Application areas include the usual HPC verticals: finance, life sciences, oil & gas, EDA, digital content creation and basic science research. A number of bindings and libraries are also now available so that Python, MATLAB, and other environments can tap into GPGPU.

Now with the 10-series Tesla products due to be released this month, OEMs and integrators can construct GPU-equipped servers and desktop boxes with double-precision floating point support. Presumably, workstation vendors like Dell and HP could build accelerated HPC desktop systems, but since the demand for these machines is still largely unknown, these firms will probably be content to watch more specialized companies like Velocity from the sidelines. Likewise, IBM could develop an equivalent Cell-BE based workstation, but the market for such a system is likely to be much more constrained than ones based on the more ubiquitous GPU.

It was less than a month ago that Cray introduced its own entry-level supercomputer, the CX1. Whereas the Velocity offering is essentially an SMP machine with GPU accleration, the CX1 is the more traditional cluster architecture, but scaled down for personal use. JRT, which sells both systems, seems to be covering its bases here. It's quite possible both machines can find their own niches -- the CX1 for more traditional MPI-based applications and the Velocity boxes for more global address apps that lend themselves to acceleration. The CX1 is also in a higher price band, with the least expensive configuration starting at $25,000 -- about $10,000 more than the top-of-the-line Velocity machine.

Even though the first Velocity systems just hit the streets this week, the company already has a second generation in the works. They intend to quickly move to four-socket CPU configurations, and will incorporate the Nehalem processor when it becomes available later this year. Further down the road, it may be possible to hook the workstations together for applications requiring greater scale.

If Velocity Micro can make a go of this, the "Attack of the Killer Micros" saga will have added a new chapter. Instead of just commodity microprocessor hardware invading HPC's turf, PC vendors themselves could start eating into the market from the bottom up. Meanwhile, it will be interesting to see if any other desktop vendors are tempted to jump into the HPC arena.